Memory mechanisms in agentic systems

Type: types/note.md

Each row uses one retained analysis artifact and its stated memory boundary. Values carry their evidence basis. Revision 2 shows each unit's scope, findings and coverage; a strong witness does not upgrade weaker alternatives. Revision 1 retains its original aggregate meaning. Admission-control semantics differ between revisions; do not pool them. Absence, inapplicability, uninspected mechanisms, and indeterminate classifications remain distinct. This is the selected population, not the historical memory-review corpus.

code-grounded (176)

System	Profile revision	Compared boundary	Storage	Write agency	Read-back	Push selection	Trace learning	Authority	Evidence
A-MEM	2	Accumulated MemoryNote content, semantic fields, links, usage/history fields and Chroma text/metadata/vector snapshots; admission, direct edits, automatic evolution, index maintenance and later reads within an initialized instance. Provider-side retention is unresolved. Static prompts, inference/embedding model weights, test assertions, external task agents and hypothetical cross-session restoration are not accumulated memory in this boundary.	partial coverage (The local object and collection inventory is explicit; dependency internals and provider retention prevent complete substrate coverage.); Live note content, metadata, links and usage/history attributes: known — in-memory [wired] (All live note field groups are inventoried in the cited objects.); Chroma enhanced documents, metadata and indexed vectors: partial — vector [wired] (Dependency-internal physical storage and durability are uninspected; no sqlite, files or cross-session substrate is inferred.); Possible provider-side retention of supplied memory payloads: uninspected (Adapter requests do not expose provider storage or establish that accumulated provider memory exists.)	partial coverage (Automatic transformations coexist with unresolved caller control; README review advice does not establish manual admission.); Missing/default semantic-field admission and model-directed relationship/neighbor edits: known — automatic [wired] (Starting ingestion is not approval of each subsequent generated edit; exposed process_memory shares this automatic disposition.); Threshold-triggered bulk snapshot republication: known — automatic [wired] (Agency is resolved despite unresolved successful replacement of existing IDs.); Caller content/metadata admission, direct CRUD/live-reference edits, explicit maintenance and direct retriever writes: not-determinable (Generic Python caller identity does not establish a human content decision or automatic caller policy; physical software I/O does not resolve admission agency.); Possible provider-retained state admission: uninspected (Adapters establish inference requests only, not provider memory admission controls.)	partial coverage (Both local directions are wired. Provider-internal uncertainty is not resolved by those witnesses; no cross-session or task-answer route is inferred.); Caller-requested ID, similarity and linked retrieval/formatting: known — pull [wired] (Optional link expansion remains part of the requested return, not a second unsolicited delivery.); Automatic neighbor selection and delivery to evolution during later add or direct processing: known — push [wired] (A caller's ingestion request can trigger internal push without becoming an autonomous scheduled loop.); Possible provider-internal retained context delivery: uninspected (Stateless local requests do not reveal provider retention or later-consumer routes.)	partial coverage (Embedding-based internal push is positive; caller pull does not make unresolved provider alternatives inapplicable.); Automatic evolution neighbor push: known — inferred-embedding [wired] (IDs label retrieved neighbors and bound subsequent writes; they are not the push selector. Five is a budget, not a separate coarse selection route.); Explicit caller ID/similarity/link retrieval: inapplicable (These operations fulfill pull requests; identifier and embedding selectors on them are not additional push signals.); Possible provider-internal push selection: uninspected (No provider-retained payload or selector is inspectable from the adapters.)	not-determinable coverage (No controlled system finding is established: bounded absent feedback does not resolve the origin/control of other included writes. Improvement evidence is not the missing requirement; trace-fed provenance is.); Automatic usage/history/rationale feedback through note usage fields: absent (The searched shipped paths do not increment access/count fields, append evolution reasons or feed them back. This bounded absence is not a system-wide no.); Caller input to automatic semantic fields/neighbor edits and later embedding/evolution consumption: not-determinable (Automatic retained transformations and later consumers exist, but original inputs are not established as activity traces. No qualifying trace-fed yes or comprehensive no follows.); Direct caller replacement/history admission and possible provider learning: not-determinable (Caller agency and payload origin are unknown; provider internal retained updates are uninspected. Local absence of parameter training does not settle artifact-based trace learning.)	partial coverage (Local context, ranking and routing paths are positive; call-local evolution controls and static schemas are not retained-memory enforcement/validation criteria. Opaque consumers prevent a complete union.); Retained content and semantic snapshots delivered to later evolution model and Python readers: known — knowledge [wired] (Delivery wiring is established; actual model reliance or a downstream answering agent is not.); Content/semantic enhanced documents and dense vectors at similarity selection: known — ranking [wired] (This is ranking influence, not evidence of superior retrieval.); Retained link IDs at optional linked-result expansion: known — routing [wired] (Symbolic selection does not grant retained links executable instruction force.); Unconstrained caller history payload and opaque provider retention consumers: not-determinable (No automatic usage/history selector exists locally; payload meaning and provider consumers are unresolved. No learning or instruction authority is inferred.)	AAS-2026-10-16-a-mem-sys-eff95cd80440-01
academic-research-skills	2	Accumulated task memory: Material Passport and referenced working artifacts, run/reset and inquiry histories, imported/exported corpus and screening decisions, resolver cache, reading declarations, criteria bindings, revision/audit markers and evidence, evaluation reports, process narrative, activity store, Pi activation state and the opaque host continuation interface. Includes incompletely inspected adapters, enrichment/audit/criteria/secondary-retention alternatives rather than assuming coverage. Excludes shipped static skill/rule files, ordinary transient returns without later retention, provider parameter implementations and external scholarly source truth; these are not classified as accumulated ARS task memory.	partial coverage (The inventory covers named task stores and the opaque host interface. It does not classify all runtime storage or equate versioned static instructions with accumulated memory; host storage and incompletely inspected optional histories prevent complete coverage.); Local retained task files: passport/artifact references, run and inquiry ledgers, human read log, criteria manifest, activity store, screening decisions, working markers and revision/evaluation artifacts: known — files [wired], files [afforded] (The named local objects are files; JSON/YAML/Markdown encoding does not imply a database or graph.); Imported/exported corpus files and prescribed process narrative files: known — files [wired], files [claimed] (Corpus files have executable producers; process-file production is a model protocol.); Resolver-decision cache: known — sqlite [wired] (The local persistent cache is explicitly SQLite, independently of its JSON payload encoding.); Host-owned Pi session entries and host continuation summaries: uninspected (Branch-local activation is persisted by the host session manager, but its physical substrate and compaction storage are not inspected. No repo, in-memory, service-object or other substrate is inferred.)	partial coverage (Admission is classified per mechanism, not by authorship or filesystem write. Known manual declarations and automatic transformations coexist with unresolved initialization and external control.); Automatic admission by inspected folder/export writers, resolver cache computation, screening merge, criteria receipt checks and activity append: known — automatic [wired] (Software admits each result without a separate human content decision; starting the import/run or selecting a store is not per-content approval.); Run-ledger event capture under orchestrator protocol and model process-summary generation: known — automatic [claimed] (The model selects/captures event content or generates a narrative, rather than receiving per-entry human approval.); User reading marks/rescissions, explicit author inquiry adoption and item-level manuscript authority: known — manual [afforded], manual [claimed] (The operator explicitly supplies or approves this retained content; subsequent mechanical I/O does not alter that agency.); Pi activation entry admission: known — automatic [wired] (A recognized start/stop/input event causes software to append/restore activation state, without per-entry content approval.); Initial passport/criteria/profile creation, generic caller-injected cache dictionaries, other adapter and host summary admission: not-determinable (Generic caller identity does not settle operator control. Criteria initialization and opaque summary control are incompletely inspected; automatic known cache computation does not upgrade these routes.)	partial coverage (The inventory preserves pull and push within different chains and keeps opaque alternatives. A known request-driven route cannot make uninspected host/optional delivery inapplicable.); Requested corpus supply, cache lookup, reading-coverage query and activity render: known — pull [afforded], pull [wired] (Consumer requests are fulfilled from retained material. Code delivery of a requested result is not an extra push.); Explicit fresh-session resume and named artifact reload: known — pull [afforded] (The user supplies the passport and the resumed orchestrator loads named references; requested reconstruction is pull.); Automatic checkpoint inquiry display: known — push [afforded] (Named checkpoint moments or a recorded signal trigger a selector without a separate request for each displayed branch.); Branch/session activation restoration and automatic recovery checks: known — push [wired], push [claimed] (Activation restoration and required post-return/compaction ledger checks are distinguished from the requested resume artifact read.); Requested later comparison of saved evaluation reports: known — pull [wired] (A later gate invocation explicitly consumes baseline/compare reports.); Host summary delivery, all optional cross-model retention, detailed criteria consumers and later process-lesson adoption: not-determinable (No unified external summary/secondary-store consumer route or automatic lesson reload is established; prescribed output alone does not classify an unknown later delivery.)	partial coverage (Only independently selected push operations receive signals. Inquiry display code is wired even though orchestrator delivery is afforded. Pull requests are excluded from signal findings, but unresolved directions prevent axis inapplicability.); Checkpoint-triggered inquiry summary selection: known — coarse [wired] (Trigger: designated checkpoint or recorded signal. Inputs: feature flag, branch count, live statuses, summary moment and optional signal event. Selected parts: live statements/status/provenance and stale-artifact context.); Session-start/tree-change activation restore: known — coarse [wired] (Trigger: session/branch event. Inputs: current branch custom entries and ARS state type. Selected part: latest applicable activation entry. This is coarse availability/latest-state selection, not a requested path identifier.); Protocol-triggered run-ledger recovery comparisons: known — identifier [claimed] (Trigger: compaction, subagent return or stage close. Inputs: claimed decisions/expected steps and recorded artifact bindings. Selected parts: corresponding preserved words and valid receipt outcomes, with conservative counter recovery.); Opaque summary selection and optional inquiry signal/payload, criteria and audit alternatives: not-determinable (No inspected payload/selector account classifies all host and optional paths. A cache composite key and requested corpus filename belong to pull routes, not evidence of identifier push.)	partial coverage (The cache and prescribed ledger route support local positive trace-fed writes. Host and optional routes remain unresolved. This axis is not a claim of improved capacity, reflection, autonomous self-improvement or changes to provider weights.); Resolver-result to retained match decision to later verification: known — yes [wired] (Automatic computation writes a durable behavior-shaping decision that a later exact-key verification invocation consumes. This satisfies the route definition without parameter training or measured gain.); Selected user/tool outcome capture to durable run ledger to later recovery: known — yes [claimed] (The model-operated capture path retains decision words, execution receipts and counters for resumed behavior. Its software append is inspected, but the complete trace-selection/consumer chain is protocol-level.); Activity history and process narrative at their later consumers: not-determinable (Activity extraction is automatic but model/gate use is forbidden; its human display does not establish a behavior-shaping update. Process summary has a human output reader but no established automatic lesson-adoption route. Neither provides a bounded absence record for a local no.); Host continuation summaries and remaining optional write/consumer paths: not-determinable (Host compaction could qualify if automatic retained summaries reach later context, but its actual generation/payload route is opaque. Optional criteria/audit/secondary retention lacks complete trace-fed write coverage; no system-wide no is warranted.)	partial coverage (The inventory distinguishes retained parts, consumers and effects, including update versus downstream authority. Several controls have wired local effects but model refusal and context-use remain protocol/affordance. Unassessed optional effects prevent an exhaustive authority union.); Corpus prose and metadata at later literature roles: known — knowledge [afforded] (The consumer re-screens retained records as untrusted research data and can skip them; it does not adopt them as instructions.); Run-ledger decisions, receipts and counters at recovery and dispatch: known — instruction [afforded], routing [afforded], knowledge [afforded] (Actual consumer roles are the resumed orchestrator and its selected handoffs; the entire ledger is not dispatched.); Pi activation bit at adapter callback: known — routing [wired] (The retained bit selects active compatibility-guidance injection, rather than recalling paper content.); Cached resolver decisions at later verification wrapper: known — knowledge [wired] (Matching/version/type checks permit use of a saved lookup decision, which then supplies finalization evidence.); Reading log scopes and rescissions at the coverage resolver: known — validation [wired], enforcement [wired] (The ledger supplies criteria and withdrawable eligibility; it is not proof of actual reading.); Criteria-binding manifest at initialization/receipt/handoff validation: known — validation [wired], enforcement [wired], routing [afforded] (This is identity/binding validation, not established semantic application of criterion prose.); Inquiry projection at author/foreground checkpoints and process narrative at human reader: known — knowledge [afforded], knowledge [claimed] (Selected alternatives and narrative supply context/advice; the inquiry ledger cannot override integrity gates.); Retained working marker statuses and revision authority at formatter/writer consumers: known — validation [claimed], enforcement [claimed], instruction [claimed] (These forces attach to retained task controls, not the static rules that interpret them.); Stored screening decisions, pilot results and evaluation measurements at reporting/admission consumers: known — routing [wired], validation [wired], enforcement [wired] (Different retained results have different effects; none is a scientific truth oracle.); Resolver execution evidence at automatic cache-update consumer: known — learning [wired] (Learning here is the force of trace input at the durable decision-update path, not demonstrated improvement or downstream instruction authority.); Activity history at human display; host summaries, semantic criteria consumers and remaining audit/review effects: not-determinable (Activity is explicitly excluded from model/gate inputs; its human counts do not establish a controlled behavioral force. Host-summary effects, full criterion application, optional audit paths and retained synthesis weighting are incomplete. No unsupported ranking or blanket learning force is inferred.)	AAS-2026-10-17-academic-research-skills-6fe859d62c16-01
Agent S	2	Included: S1/S2 persistent task/subtask summaries, query/search caches and embedding access metadata; ACI notes across generations; S3 retained coding summaries/action history; offline rollout evidence, derived captions and saved comparison/resume state; unresolved initial archive/local content and provider-side retained-state alternatives. Excluded: static shipped procedural instructions, ordinary turn/reflection accumulation, eviction without a retained replacement, external work products, and provider weights as use-updated memory absent a recorded update route.	partial coverage (Named local content, access metadata and offline evidence have classified substrates. Static source policy and external model weights are not accumulated-memory substrates. Provider-side retained state remains unresolved; the positive local witnesses do not close that alternative.); Historical narrative, episodic, query and search content; embedding access cache: known — files [wired] (All five named operative stores are local filesystem JSON or pickle; dense encoding is not a vector-service substrate.); ACI notes and cached code-agent result: known — in-memory [wired] (These retained objects have instance-lifetime RAM storage, including one-shot coding return delivery.); Offline raw rollout evidence, captions and persisted comparative result: known — files [wired] (Files survive producer invocations and have named later consumers.); Uncaptured initial bundle payload beyond the named loaded content and vector stores: not-determinable (The downloader and named local destinations are known; archive bytes and any additional operative payload inventory are not. Extraction alone cannot classify every retained part.); Included provider-side retained state behind generation, grounding and embedding endpoints: not-determinable (The accepted component records establish external calls, not the existence, contents or mechanisms of use-accumulated server state. No storage_substrate value for that state follows from client configuration or model weights.)	partial coverage (All established local admission mechanisms are automatic, with afforded alternative APIs kept separate from wired callers. Initial human control remains unresolved. Provider-side retained state remains unresolved; the positive local witnesses do not close that alternative.); S1/S2 CLI summary admission: known — automatic [wired] (Human task entry is not a per-summary decision.); Exposed S2 save/finalize admission mechanism: known — automatic [afforded] (Generic caller identity does not establish human content approval; the implemented model/key admission itself is automatic.); Query/search and embedding cache miss admission: known — automatic [wired] (Writing to disk is distinguished from the model/software content admission.); Notes append and coding-result retention: known — automatic [wired] (Coding capability opt-in is not approval of each retained string or summary.); Driver evidence, narrator captions and comparative choice persistence: known — automatic [wired] (Workflow launch does not make each recorded/derived item manual.); Release archive installation: known — automatic [wired] (Asset selection does not establish per-entry review.); Initial content creation and prior local-file/release review: not-determinable (No explicit human supplying/editing/approval decision for retained entries is recorded. This prevents a manual finding without weakening later automatic transformations.); Included provider-side retained state behind generation, grounding and embedding endpoints: not-determinable (The accepted component records establish external calls, not the existence, contents or mechanisms of use-accumulated server state. No write_agency value for that state follows from client configuration or model weights.)	partial coverage (Requested retrieval and independent automatic context supply coexist; a request-driven return is not relabelled push simply because code selects the returned entry. Provider-side retained state remains unresolved; the positive local witnesses do not close that alternative.); Historical experience requests and their fulfilled returns: known — pull [wired] (The first planning/subtask invocation requests retrieval; fulfillment is not counted again as unsolicited delivery.); Exact-task query/search cache reuse and vector access: known — pull [wired] (These reads fulfill requests for retained query/result/access material.); Automatic full notes delivery on worker invocation: known — push [wired] (Independent renderer supply is distinct from the earlier save action.); Pending coding return delivery to next worker invocation: known — push [wired] (The earlier delegation request is not a retrieval request for the subsequently retained cache.); Offline caption/evidence loads and saved-choice resume/scoring reads: known — pull [wired] (Explicit producer/judge/scoring invocations request recorded material; file delivery is not another push.); Initial unseen bundle read alternatives: not-determinable (Named historical loaders are classified above; unseen archive material does not establish the full alternative consumer inventory.); Included provider-side retained state behind generation, grounding and embedding endpoints: not-determinable (The accepted component records establish external calls, not the existence, contents or mechanisms of use-accumulated server state. No read_back_direction value for that state follows from client configuration or model weights.)	partial coverage (Coarse availability is the only supported local push signal. Embedding ranking belongs to requested experience retrieval, so it does not establish inferred-embedding on this axis. Provider-side retained state remains unresolved; the positive local witnesses do not close that alternative.); Notes push: every worker invocation, full-list availability selector: known — coarse [wired] (Trigger is worker message construction; selector input is the current notes list; the selected retained part is the whole list.); Coding-result push: next worker invocation, pending-return availability selector: known — coarse [wired] (Trigger is the next worker call; selector input is nonempty cached return; selected parts are summary, completion fields and formatted actions, not all thoughts/outputs.); Request-driven historical similarity, exact cache/vector access and offline file loads: inapplicable (These are pull operations. Their embeddings, exact keys and filenames are not push-selection signals.); Unseen initial-content delivery alternatives: not-determinable (The named pull routes do not resolve any additional consumer selection hidden in the absent bundle payload.); Included provider-side retained state behind generation, grounding and embedding endpoints: not-determinable (The accepted component records establish external calls, not the existence, contents or mechanisms of use-accumulated server state. No read_back_signal value for that state follows from client configuration or model weights.)	partial coverage (Positive automatic trace-fed retained artifacts are supported without a claim of observed improvement. Unresolved writes prevent a system-wide negative or complete inventory assertion. Provider-side retained state remains unresolved; the positive local witnesses do not close that alternative.); S1/S2 CLI trajectory-to-summary-to-later-guidance chain: known — yes [wired] (Automatic trace-fed durable guidance qualifies architecturally; improved capacity is not established.); Alternative S2 summary lifecycle and shared-store later-consumer capability: known — yes [afforded] (The alternate write is afforded, not a demonstrated shipped lifecycle caller.); Coding history-to-summary-to-next-worker chain: known — yes [wired] (Within-task retention qualifies; neither cross-session storage nor complete controller feedback is required.); Offline rollout-to-caption-to-comparison chain: known — yes [wired] (Later comparison/scoring is the consumer, not a live GUI policy update.); Arbitrary model note admission: not-determinable (The append and later consumer are wired, but actual note input origins are not established for this generic string-admission mechanism; no universal trace-fed classification follows.); Query/search caches and key-embedding transformations: not-determinable (These mechanisms persist responses and access metadata, but the records do not establish every original input as a qualifying activity trace or bound absence of trace-fed alternatives.); Initial bundle and prior local content production: not-determinable (Original production routes and trace mixtures are unavailable; import alone establishes neither trace learning nor a bounded no.); Included provider-side retained state behind generation, grounding and embedding endpoints: not-determinable (The accepted component records establish external calls, not the existence, contents or mechanisms of use-accumulated server state. No trace_learning value for that state follows from client configuration or model weights.)	partial coverage (Authority is classified at actual update, advice, ranking and mechanical consumers. Static procedural policy, format checks and external work-product effects are not attributed to accumulated memory. Provider-side retained state remains unresolved; the positive local witnesses do not close that alternative.); Narrative and episodic summaries at historical manager/worker consumers: known — knowledge [wired] (Optional reference advice in a system prompt is knowledge, not a standing instruction solely because of its channel.); Retained raw textual historical trajectories at summary-update consumers: known — learning [wired] (Experience feeds durable behavior-shaping summaries; this force is separate from downstream advice.); Embedding vectors and associated entry keys at similarity selection: known — ranking [wired] (Fixed scoring code still gives retained access material ranking influence.); Query values at subsequent search/task-identity consumers: known — routing [wired] (A reused query chooses search input and retained task identity rather than granting factual truth.); Cached search content at fusion and planning consumers: known — knowledge [wired] (Model fusion is advisory and may omit content.); Notes and reduced code return at outer-worker consumer: known — knowledge [wired] (Delivery can shape decisions but no enforced compliance or accepted truth follows.); Recorded coding action/thought history at summary producer: known — learning [wired] (The update consumer differs from the outer worker and omits direct controller outputs.); Raw offline rollout evidence at caption production: known — learning [wired] (Trace evidence feeds retained material shaping later comparison, not GUI-worker weights.); Caption answers and available endpoint images at comparative judge: known — knowledge [wired], ranking [wired] (Evidence and relative-priority influence coexist at this consumer.); Saved task keys and selected directory at resume/scoring consumers: known — routing [wired], enforcement [wired] (Only these mechanical fields, not stored rationale, establish these effects.); Unseen initial payload authority beyond named consumer paths: not-determinable (The known downstream consumers classify named loaded parts, not every unseen release entry or alternative payload.); Included provider-side retained state behind generation, grounding and embedding endpoints: not-determinable (The accepted component records establish external calls, not the existence, contents or mechanisms of use-accumulated server state. No behavioral_authority value for that state follows from client configuration or model weights.)	AAS-2026-10-16-agent-s-db6fee212353-01
Agent Skills for Context Engineering	2	Use-accumulated research queues, source captures, evaluations, proposals, registry/ledgers, novelty/router results, digital-brain JSONL, optimizer traces/analyses/prompts/generated skills, CAS/export products, vector/graph helper state, scratch observations, summaries and plan checkpoints, including described mirror/latent/opaque alternatives and unresolved remaining shipped memory/training/governance paths. Includes their access, admission, maintenance and later-consumer routes. Excludes static shipped instruction distribution as accumulated memory, ordinary nonretained current context, external host/provider internals and unsupplied private deployment state; those exclusions do not resolve included opaque payloads or host-consumption gaps.	partial coverage (Positive file, repository, process-memory, vector and graph witnesses do not enumerate the partial-loop inventory.); Research queues, registry/ledgers, results, retained digital-brain JSONL, optimizer artifacts, plans, scratch and CAS/export files: known — files [wired], repo [wired] (These retained parts have inspected file storage; tracked registry files also reside in the repository.); Process-local numerical vector index: known — in-memory [wired], vector [wired] (Synthetic dense index is retained in the helper object, not a persistent database.); Process-local property graph: known — in-memory [wired], graph [wired] (Node/edge dictionaries implement graph storage.); Observation store, structured summary and optimizer trace/analysis objects: known — in-memory [wired] (Named retained intermediate objects live in process memory.); Described latent KV/opaque compression payload storage: not-determinable (No inspected tensor persistence or opaque payload implementation determines a substrate. A KV tensor label is not a key-value database.); Remaining included helpers, examples, governance and lifecycle paths: storage: uninspected (The accepted memory coverage inventory leaves multi-agent/RDF/sandbox helpers, book training, prompt-lab/book workflows, digital-brain producers and governance/migrations uninspected. These records do not enumerate their storage; no corpus-complete conclusion follows.)	partial coverage (Automatic transformed-content admission is supported; no manual classification is inferred from caller authorship or reviewer strings.); Research queue/state/result recording and program-derived registry/export projections: known — automatic [wired] (Software admits these program-generated retained parts; caller approval of initial content is a separate unresolved mechanism.); Optimizer replacement, artifact and generated-skill admission: known — automatic [wired] (Starting optimization is not approval of every revision.); Helper offload, regex summary, masking and graph-property updates: known — automatic [wired] (Software thresholds/extractors/update rules admit transformed retained content.); Initial facts/plans/CAS bytes/proposals and digital-brain content approval: not-determinable (Generic API/CLI caller or reviewer string does not establish a human per-content decision; manual versus automatic initial admission remains unresolved.); Human-reviewed export correction and source-review protocol admission: not-determinable (Correction review is prescribed; actual content admission control is not authenticated. Source-review implementation status is disputed in reconciliation.); Remaining included helpers, examples, governance and lifecycle paths: write/admission mechanisms: uninspected (The accepted memory coverage inventory leaves multi-agent/RDF/sandbox helpers, book training, prompt-lab/book workflows, digital-brain producers and governance/migrations uninspected. These records do not enumerate their write/admission mechanisms; no corpus-complete conclusion follows.)	partial coverage (Concrete requested comparator/scaffold/generator consumption is wired; demonstrated plan reconstruction is afforded. Memory, CAS and observation/lazy-loading APIs do not establish later consumer chains. Automatic successor instruction delivery is independently wired; these gaps and unresolved host alternatives keep direction coverage partial.); Requested registry comparison, digital-brain scaffold and final-prompt artifact reload: known — pull [wired] (Wired requested consumption is limited to these three routes, not adjacent storage APIs.); Requested plan reconstruction in the save/load context-refresh demonstration: known — pull [afforded] (The supported demonstration affords requested reconstruction; deployed model continuation and an automatic refresh hook remain outside this unit and unresolved.); Vector/graph helper API as a later-consumption route: not-determinable (Query methods afford requested access while an object survives, but later read-back is explicitly uninspected. No traced later consumer establishes a memory direction here.); CAS body get/materialize as a later-consumption route: not-determinable (Storage and requested reconstruction are implemented, but the recorded search found no production evaluation/agent caller. These operations alone do not establish later consumption.); Observation retrieval and lazy-file loading as later-consumption routes: not-determinable (Exact-ID observation access and requested cached summary/detail loading are capabilities; successor delivery and a later enclosing consumer remain unresolved. No wired direction is inferred.); Optimizer automatic successor-capture prompt delivery: known — push [wired] (On a continued round, orchestration supplies selected revised system text without a capture request for retained material.); Described mirrored context replacement on later turns: known — push [claimed] (This is declared host behavior, not target implementation.); Generated skill installed host, downstream export, scratch-file reader and untraced helper enclosing consumers: not-determinable (Storage/return capabilities do not close every later request/selection/delivery chain.); Remaining included helpers, examples, governance and lifecycle paths: request and delivery chains: uninspected (The accepted memory coverage inventory leaves multi-agent/RDF/sandbox helpers, book training, prompt-lab/book workflows, digital-brain producers and governance/migrations uninspected. These records do not enumerate their request and delivery chains; no corpus-complete conclusion follows.)	partial coverage (Coarse automatic prompt supply is supported. IDs on requested files, registry entries or observation retrieval are not identifier-based push witnesses; unresolved routes prevent inapplicability.); Continued optimizer round selects current/best prompt for successor capture: known — coarse [wired] (Trigger is continuation; selector inputs are loop scores/growth/current-best state; selected retained part is whole chosen prompt.); Described mirrored edited-message delivery: known — coarse [claimed] (Trigger is a later turn; selector uses accepted edit/current mirrored context; selected part is replacement message content.); Unresolved installed-host, opaque/KV and export/helper delivery selectors: not-determinable (No actual selector-to-delivered-retained-part chain resolves these alternatives. Attention-based doctrine alone does not prove embedding retrieval.); Remaining included helpers, examples, governance and lifecycle paths: push selectors: uninspected (The accepted memory coverage inventory leaves multi-agent/RDF/sandbox helpers, book training, prompt-lab/book workflows, digital-brain producers and governance/migrations uninspected. These records do not enumerate their push selectors; no corpus-complete conclusion follows.)	partial coverage (A wired yes establishes existence, not complete learning inventory. No bounded absence record supports a local or system-wide no; runtime improved-capacity learning remains independently uninspected.); Automatic trace-fed optimized prompt, durable artifact reload and later skill generator: known — yes [wired] (Activity feeds automatic instruction writes; final_prompt.txt survives and a separate generator consumes it. Within-run revised text also reaches capture.); Direct generated skill at later installed host: not-determinable (Direct generation is wired, but installed activation/later behavioral consumption is not established as a separate completed learning route.); Context summaries/offload, mirrored edits and latent/opaque alternatives: not-determinable (Automatic extraction is not sufficient without an established retained-result-to-later-context consumer for these branches.); Remaining included helpers, examples, governance and lifecycle paths: qualifying trace-fed writes and later consumers: uninspected (The accepted memory coverage inventory leaves multi-agent/RDF/sandbox helpers, book training, prompt-lab/book workflows, digital-brain producers and governance/migrations uninspected. These records do not enumerate their qualifying trace-fed writes and later consumers; no corpus-complete conclusion follows.)	partial coverage (Local consumer forces are classified separately; unconsumed metadata is not granted authority.); Registry prose at lexical novelty ranking consumer: known — ranking [wired] (The retained text influences relative overlap priorities; ledger rationale is not consumed here.); Retained research state and router result identity at later code consumers: known — routing [wired] (State reload chooses work; resume results choose skipped operations.); Novelty report/rationale and evaluation fields at active readiness and promotion: known — validation [wired], enforcement [wired] (Default active checking has closed-state and explicit override limits.); CAS access metadata at artifact get/materialize: known — validation [wired], enforcement [wired] (Access checks are local code effects, not complete production privacy assurance.); Retained numerical vectors at helper cosine-ranking consumer: known — ranking [wired] (Local ranking is implemented even though later model consumption and semantic quality are not established.); Retained digital-brain insights at scaffold consumer: known — knowledge [wired] (Content contributes draft reference material without truth checking.); Exposed trace at analyzer/update consumer: known — learning [wired] (Learning force here is input to instruction revision, not demonstrated capacity gain.); Revised retained prompt at successor capture: known — instruction [wired] (Standing system text directs the next capture; actual compliance is unobserved.); Persisted final prompt at separate skill-generation consumer: known — knowledge [wired] (Artifact generation reads final text/summary but not complete iterations.); Trace analysis scores and recommendations at optimizer and continuation consumers: known — ranking [wired], routing [wired], learning [wired] (Scores affect selection and recommendations shape replacement guidance.); Generated skill at installed host; CAS/export payload at later model; helper facts/observations at enclosing agent: not-determinable (File generation or API return alone does not establish the actual later behavioral force; host and production integrations remain unresolved.); Remaining included helpers, examples, governance and lifecycle paths: actual consumers and effects: uninspected (The accepted memory coverage inventory leaves multi-agent/RDF/sandbox helpers, book training, prompt-lab/book workflows, digital-brain producers and governance/migrations uninspected. These records do not enumerate their actual consumers and effects; no corpus-complete conclusion follows.)	AAS-2026-10-19-agent-skills-for-context-engineering-7e663338370b-01
Agent Workflow Memory	2	Included: accumulated/replaced Mind2Web and WebArena workflow files; their raw conversation/trajectory inputs, gate signals, retained reasons and annotation metrics; initial concrete exemplars/specifiers and website workflow contents insofar as they supply initialization or later selection; optional retrieval output files/vector access structures; optional in-process episode notes and their recorded later display route. Excluded: unchanged shipped induction/evaluation instructions as independent memory, ordinary untransformed action/observation history, task configuration and benchmark ranking data as general runtime storage, auxiliary basic agent without workflow memory, provider model weights/internals and BrowserGym internal persistence implementation. Initial content provenance, external production control and optional-note durability remain unresolved rather than excluded.	known coverage (The recorded inventory covers operative files, optional vector access structures and episode notes. External trace files are included at their AWM-visible boundary, not as a claim about BrowserGym storage internals. Provider model weights are computations, not recorded use-updated memory in this boundary.); Mind2Web and WebArena workflow payloads, including shipped initial contents and replacement outputs: known — files [wired] (Both consumers load plaintext; neither workflow syntax nor source-repository shipment makes the operative store a graph or a repository-backed memory mechanism.); Concrete exemplars/specifiers, saved Mind2Web conversations/metrics, WebArena raw evidence and evaluator signals/reasons: known — files [wired] (The records identify JSON/text and screenshot files at the consumer boundary. Unknown content provenance and external file-production details do not change this visible substrate.); Optional retrieval index and selected output file: known — vector [afforded], files [afforded] (The index is an optional access structure, not proof of successful semantic workflow retrieval. The workflow branch saves a FAISS store but returns the wrong object; ablation constructs its store directly and does not persist it.); Optional episode scratchpad: known — in-memory [afforded] (The included note list is in process only. Diagnostic return does not establish external persistence or cross-episode reload.)	partial coverage (Software/model admission and exact human rule approval are classified separately. Starting a pipeline does not make subsequent content admission manual. Initial shipped content and external trace production remain unresolved.); Mind2Web offline/online induction and WebArena neural regeneration: known — automatic [wired] (The human chooses modes, paths and invocation, while software/model stages admit each replacement without per-content approval.); WebArena rule concrete-example admission, manual and auto variants: known — manual [wired], automatic [wired] (Both variants have inspected admission branches and retain distinct control. Manual review reasons are not persisted.); Mind2Web conversation/metric writes and WebArena evaluator result/reason writes: known — automatic [wired] (Models may propose judgments and predictions, but software retains them without per-record human approval.); Optional retrieval-selected file/index admissions and episode-note retention: known — automatic [afforded] (Operator initiation selects configuration, not each retained item or note.); Initial shipped exemplar/specifier and website workflow content: not-determinable (No author or admission history establishes an explicit human content decision or an automatic initial write. Plaintext editability alone is not a manual route.); External BrowserGym raw trace/reward/screenshot production: not-determinable (AWM reads these files automatically, but that read does not determine control of the external production/admission mechanism.)	known coverage (Recorded later-consumer routes request files, traces or configured history. Optional retrieval and scratchpad paths remain afforded, not upgraded to ordinary wiring. The inventory contains no independent unsolicited delivery chain; static unchanged corpus shipment is not itself read-back.); Default workflow reads at later Mind2Web sample and WebArena action invocations: known — pull [wired] (Configured workflow paths are request inputs. Reading and delivering the selected whole file is one pull chain, not a second push; Mind2Web can omit oversized content.); Saved conversation and experiment/result reads at later evaluator or induction calls: known — pull [wired], pull [afforded] (Update/evaluation consumers explicitly read accumulated result artifacts. Human report review is a possible reader, not unsolicited delivery.); Standalone workflow/exemplar retrieval and configured output-file handoff: known — pull [afforded] (The operator requests selection, and a later configured consumer requests its output. Semantic workflow failure limits implementation strength, not the requested direction of the intended route.); Optional scratchpad return through enabled history: known — pull [afforded] (Enabled history/memory flags request the prior notes as context; ordinary CLI disables note display. No independent push selector is recorded.)	inapplicable coverage (This axis classifies push selectors. The complete recorded direction inventory is pull-only, including optional alternatives; file names, corpus tag filters and requested semantic search are not push signals.); Default workflow reads, requested update/evaluation evidence, optional retrieval outputs and enabled episode history: inapplicable (Every included read-back operation is requested file/evidence/history consumption. No independent push operation remains to classify; identifiers on requested files do not establish identifier-based push.)	partial coverage (Automatic trace-fed durable workflow updates qualify architecturally, without requiring improved performance or learning from only the agent's own executions. Manual rule approval is not automatic. Optional episode notes have no established durable later-consumer horizon, so a positive workflow witness does not resolve that alternative.); Mind2Web online saved conversation to durable workflow to later prediction: known — yes [wired] (Recorded model messages feed computational induction/overwrite and a later user-exemplar consumer. Order-based task pairing limits provenance accuracy, not existence of the write chain.); Mind2Web offline annotated training trajectories to durable workflow: known — yes [wired] (Imported multi-step action examples record task activity; traces need not originate in this agent's own execution. Automatic admission and later workflow consumption are inspected.); WebArena neural and automatic-rule trajectory to durable system-guidance workflow: known — yes [wired] (This is wiring for configured eligible traces, not observed default pipeline success. Filename mismatch and empty-example generation prevent asserting each replacement is actually trace-fed.); WebArena manually approved rule examples: inapplicable (This variant has an explicit human per-item admission and therefore is not a qualifying automatic write, even though its content derives from traces.); Optional episode scratchpad production and later history display: not-determinable (A later in-episode note consumer is afforded, but durable behavior-shaping retention is not established: notes are an in-process list, no checkpoint/reload is recorded, and external diagnostic persistence is uninspected. Prevents a yes or bounded no for this route.)	known coverage (The inventory distinguishes acting directions and factual reference context, online experience and imported offline training examples at update consumers, eligibility controls, retrieval ranking inputs and display-only reasons. Authority means force at a recorded consumer, not observed compliance or benefit. Unconsumed fields have no inferred force.); WebArena workflow text at the legacy acting model: known — instruction [wired], knowledge [wired] (System-message directions are interpreted standing guidance, not enforced subroutine dispatch; factual assertions retain their knowledge effect in the same channel.); Mind2Web workflow and concrete demonstration content at the prediction model: known — knowledge [wired] (User-message demonstration delivery supplies contextual advice rather than standing system instructions or an action oracle.); Saved prediction records and WebArena thought/action experience at automatic workflow-update consumers: known — learning [wired] (These inputs feed durable behavior-shaping text updates; their update-consumer force is separate from the output guidance channel.); Imported offline annotated training examples at the Mind2Web workflow-update model: known — learning [wired] (Learning authority at this update consumer does not require examples from this agent's own execution; downstream prediction knowledge authority is a separate path.); Retained reward/rm signals at induction eligibility consumers: known — validation [wired], enforcement [wired] (Truthiness supplies an admissibility criterion and a software veto for source trajectories, not certification of generated workflows.); Exemplar specifiers at default filter and optional similarity selector: known — routing [wired], ranking [afforded] (Specifier strings are removed before model delivery. Default tag filtering and optional semantic ranking have distinct consumers and evidence bases.); Workflow names/docstrings, dense vectors and index metadata at optional retrieval: known — ranking [afforded], routing [afforded] (Ranking is supported only at the optional mechanism level; the semantic workflow object-handoff defect prevents successful retrieval claims.); Evaluator reason/display files at human review; annotation metrics at result analysis: known — knowledge [afforded], knowledge [wired] (Readable review is afforded and metrics serve diagnostics. Neither is consumed as the online Mind2Web success gate; evaluator reasons do not reach the induction model.); Optional scratchpad at later enabled acting-model history consumer: known — knowledge [afforded] (Programmatic display supplies contextual notes, not a standing system instruction or established cross-task learning route.)	AAS-2026-10-12-agent-workflow-memory-cb2da079273e-01
Agent-R	2	Retained search-tree interaction text and score/branch metadata; derived revision conversations, loss flags and feedback/path sidecars; exposed converted conversations; evaluation results and optional progress metadata; their automatic writes, selection, maintenance and recorded or unresolved later consumers. Includes declared external training as an unresolved consumer/update branch, not an established stored learned checkpoint. Excludes unchanged actor/verifier weights, static shipped prompts, ordinary current-rollout context, external environment state and opaque provider retention. Caller-supplied demonstrations/prefixes remain unresolved delivery/update alternatives rather than asserted retained objects.	known coverage (Complete recorded retained-object inventory: tree parts, export parts, evaluation results, optional progress storage and converted output. Static weights, provider retention and hypothetical trained checkpoints are not operative retained objects established in this boundary.); Persisted tree interaction content: known — in-memory [wired], files [wired] (The recorded object is assembled in process memory and retained in a file; no other operative payload is recorded.); Persisted tree score and branch metadata: known — in-memory [wired], files [wired] (The recorded object is assembled in process memory and retained in a file; no other operative payload is recorded.); Revision conversation content: known — in-memory [wired], files [wired] (The recorded object is assembled in process memory and retained in a file; no other operative payload is recorded.); Revision loss flags: known — in-memory [wired], files [wired] (The recorded object is assembled in process memory and retained in a file; no other operative payload is recorded.); Revision feedback and path sidecars: known — in-memory [wired], files [wired] (The recorded object is assembled in process memory and retained in a file; no other operative payload is recorded.); Evaluation result conversation and metadata: known — in-memory [wired], files [wired] (The recorded object is assembled in process memory and retained in a file; no other operative payload is recorded.); Optional progress index: known — in-memory [afforded], files [afforded] (The recorded object is assembled in process memory and retained in a file; no other operative payload is recorded.); Converted conversation output: known — in-memory [afforded], files [afforded] (The recorded object is assembled in process memory and retained in a file; no other operative payload is recorded.)	partial coverage (Automatic internal writes are established independently of caller authorship and filesystem I/O. Caller-supplied helper inputs and external checkpoint admission remain unresolved; no manual finding is licensed.); Tree child admission and completion save: known — automatic [wired] (Batch launch/configuration does not approve each retained child.); Revision conversation, flags and sidecar admission: known — automatic [wired] (Construction and append admission are model/software controlled without per-example human approval; threshold defects remain.); Evaluation result admission: known — automatic [wired] (Result persistence follows rollout completion, not per-result approval.); Dataset conversion transformation: known — automatic [afforded] (Fixed positional extraction automatically determines output contents once invoked; input supply/approval is separately unresolved.); Converter input admission and progress-index supply: not-determinable (Generic caller identity does not establish an explicit human content decision or autonomous admission of those supplied inputs.); External training and changed-checkpoint admission: not-determinable (Documented launch instructions establish neither per-checkpoint human approval nor automated admission of a learned derivative.)	partial coverage (The shipped cross-invocation tree consumer is pull; the converter affords requested consumption. Known pull does not resolve external/auxiliary delivery alternatives or justify an axis-wide no-push assertion.); Persisted tree request, selection and delivery: known — pull [wired] (A separate CLI requests the selected directory; enumeration and reward selection fulfill that batch request.); Supplied retained export rows at rewrite: known — pull [afforded] (Library caller supplies rows to a consumer operation; no deployed driver or trainer is wired.); External trainer and changed actor checkpoint: not-determinable (Consumer intent is named, but ingestion request/selection/delivery and checkpoint adoption direction are not established.); Auxiliary demonstrations, prefix source, progress and evaluation-result reuse: not-determinable (Helpers accept caller data or afford storage, but durable source-to-later-invocation delivery is unspecified; current rollout history is excluded.)	not-determinable coverage (No positive push signal is established. Pull-route inapplicability is local; unresolved included alternatives prevent axis-wide inapplicability. Coarse, identifier and inferred selectors are not guessed from file enumeration or task labels.); Requested tree-file and rewrite consumption: inapplicable (These are pull operations, not automatic unsolicited delivery; filename IDs and reward filters do not establish a push selection signal.); External and auxiliary possible later delivery: not-determinable (No actual push trigger, selector input and delivered retained part are recorded for these alternatives. Whether any push exists and which signal selects it cannot be determined.)	partial coverage (Positive architectural/artifact findings do not claim exercised training, improved capacity or a closed self-improvement loop. Unresolved alternate update chains preclude a complete classification; no local or system-wide no is inferred from unchanged inference weights.); Automatic tree-to-revision-training-artifact write: known — yes [wired] (Qualifying scope is automatic construction of durable behavior-shaping training material from retained experience, not executed parameter learning. The wired construction is conditional on reaching past defective thresholds.); Converted training artifact and its later consumer: partial — yes [afforded] (Automatic conversion from trace-derived conversations is afforded, but actual training input contract and downstream adoption are unresolved.); Advertised trace-fed parameter training and later actor: partial — yes [claimed] (Iterative training is declared; optimizer ingestion, updated checkpoint identity/admission and later behavioral dependence are not implemented in the captured artifact.); Other retained-result and auxiliary input update chains: not-determinable (Result retention, supplied demonstrations/prefixes and optional progress storage do not establish automatic durable behavior-shaping updates or their trace-fed later consumers.)	partial coverage (Tree-to-builder knowledge, learning, ranking and eligibility effects are recorded; external training is claimed only. No standing instruction authority is assigned to a stock correction inside an example, and unconsumed metadata receives no positive authority finding.); Retained tree text at the revision verifier: known — knowledge [wired] (The loaded text supplies task and action/observation context; its later use is not direct actor policy.); Retained tree content and outcomes at revision-data construction: known — learning [wired] (Experience is an input to an automatic durable training-artifact construction, conditional on the export code reaching its writer.); Retained score and terminal metadata at path selection: known — ranking [wired], validation [wired], enforcement [wired] (Separate relative-priority and eligibility consumers read symbolic fields.); Revision text at the exposed converter: known — knowledge [afforded] (Conversion reads stored text but does not evaluate or learn from it.); Revision text and loss flags at intended external trainer: partial — learning [claimed] (Training is declared, but no captured optimizer establishes input schema, mask interpretation, checkpoint adoption or actual effect.); Retained feedback/path sidecars at a later consumer: not-determinable (The judgment has immediate prefix-selection force, but retained reasons and path sidecars have no established later reader; immediate judgment authority is not assigned to their stored copy.); Evaluation result at a later consumer: not-determinable (No shipped later result consumer establishes authority; storage alone does not grant learning, ranking or checkpoint validation.); Progress index at a resume consumer: not-determinable (Read/write helpers have no wired caller; intended access metadata is not evidence of an actual routing consumer.); Converted output at an optimizer: not-determinable (Output text exists as an afforded storage object, but a training consumer and effect are unspecified.)	AAS-2026-10-12-agent-r-216224d4270a-01
AgentFly	2	Local memory accumulated or changed through shipped use: both driver case banks and their access views; result traces and restart keys; retrieved-case training pairs; trained retriever checkpoints and later numerical ranking; workspace payloads/logs and document/media retained side files with their write and reuse surfaces. Include opaque local payloads and unresolved initial/admission/consumer alternatives. Exclude static prompts, ordinary current-query history (including accidental exception carryover without a durable artifact), temporary return-only extraction, external provider retention/training and dependency caches, pretrained assets as use-accumulated memory, and bare helpers as established consumer routes.	known coverage (The local retained-object inventory is the two case banks, rich results, training pairs, checkpoints, workspace files/logs and document/media side files. All have file storage; loaded case and training views and initialized weights additionally reside in process memory. Binary encoding and unavailable exact weights do not obscure these substrates. External provider retention and dependency caches are excluded, not classified as local memory.); Two case banks, result JSONL and training-pair JSONL on disk: known — files [wired] (The distinct local JSONL artifacts cover both driver alternatives and their recovery/training outputs.); Loaded case pools/metadata and converted training samples: known — in-memory [wired] (These are operative derived views, not an additional durable bank.); Persisted retriever checkpoint and later loaded weights: known — files [wired], model-weights [wired], in-memory [wired] (The numerical artifact is separate from the retriever component; missing checkpoint contents do not prevent substrate classification.); Workspace payloads, execution/process logs and document/video side files: known — files [wired] (Opaque binary payloads remain included as files; readable summaries do not replace them.)	partial coverage (Driver admissions and numerical saves are automatic despite operator-started workflows. Caller identity and unauthenticated initial content prevent complete enumeration of control; no explicit per-content human decision warrants manual.); Completed-query result/case/pair admission and checkpoint optimization/save: known — automatic [wired] (Models propose plans/judgments, software appends labels and saves updates, and humans configure runs rather than approve items.); Executor-selected workspace writes/commands and document/video tool writes: known — automatic [wired] (This unit is the inspected model-dispatch path, not every possible external caller of the APIs.); Initial corpora/checkpoint supply, external file edits and generic caller-controlled workspace content: not-determinable (Possible file editing and operator path choice do not identify a content-supplying/approving human decision. Generic callers do not resolve admission control.)	partial coverage (Query-requested example selection, explicit file/CLI reads and operator-started training are pull operations. Automatic startup/reload supplies persisted pools, checkpoint state and restart keys without a consumer request for individual records. Opaque file/log alternatives lack established later routes.); Query-triggered example retrieval and fulfilled delivery to planner/executor: known — pull [wired] (The records explicitly describe driver retrieval requests followed by delivery, not an additional independent unsolicited example push.); Requested workspace reads, later file-path tool reuse, standalone retrieval CLI and separately started trainer: known — pull [wired] (These callers request retained files/pairs/cases; immediate converted-document return is excluded from later read-back.); Automatic case-bank initialization/reload, configured checkpoint initialization and result-key restart scan: known — push [wired] (These loading/supply operations are separate from subsequent requested query selection and from choosing a checkpoint replacement.); Execution/process logs, opaque payload alternatives and unspecified library consumers: not-determinable (No additional later-consumer route is established for logs or the bare helper. Actual opaque payload reuse and alternative delivery are unresolved rather than absent.)	partial coverage (Push loading is coarse availability/configuration based. Embedding/classifier relevance belongs to pull selection, not an inferred push signal. Unestablished alternative consumers prevent axis inapplicability or complete coverage.); Automatic bank/restart-file scan and configured checkpoint loading: known — coarse [wired] (Trigger is driver initialization or completed-query reload; selector input is configured path/existence or persisted result availability; selected material is the loaded pool, numerical checkpoint or restart keys. A path alone is not targeted identity matching of delivered parts.); Query-requested case retrieval, explicit workspace/file/CLI reads and requested training inputs: inapplicable (These units are fully classified as pull; their embeddings, judgment scores or filenames are not push-selection signals.); Unestablished log/opaque-payload and unspecified helper consumer alternatives: not-determinable (No recorded push selector or complete direction assessment permits a signal value or inapplicability for these alternatives.)	partial coverage (Automatic outcome-fed case writes and numerical training produce durable behavior-shaping material with wired later consumers. Other tool-file transformations, initial provenance and raw logs do not establish a qualifying update or bounded system-wide negative.); Automatic latest-plan/outcome case writes followed by later model examples: known — yes [wired] (Reduced experience is retained and delivered to later questions; no improvement or exercised reliance is required for this architectural classification.); Retrieved-case/outcome pair admission, separately launched parameter update and later checkpoint ranking: known — yes [wired] (Operator launch does not make each gradient/save manual; later inference requires initialization, not just text reload.); Workspace/tool payload writes, retained raw results/logs beyond known consumers and initial data/assets: not-determinable (Arbitrary commands may derive durable content, but no complete trace-to-derived-to-behavior consumer chain is recorded for these alternatives. Raw log retention alone is insufficient, and no bounded absence record warrants no.)	partial coverage (Authority is consumer-relative: examples advise models, retained text/weights rank cases, pairs train parameters and result keys govern restart. Unread rationale/logs and opaque tool payload paths do not inherit these forces.); Case descriptions and labels at later planner/executor example consumers: known — knowledge [wired] (User-role examples request imitation/avoidance; standing interpretation rules are static prompts outside this memory boundary.); Retained question/case-plan text at embedding/classifier scorers and checkpoint parameters at classifier: known — ranking [wired] (Ranking includes both fixed scoring of retained text and numerical policy state; observed usefulness is not asserted.); Training pair text and truth labels at gradient-update consumer: known — learning [wired] (Update authority is separate from downstream example guidance and numerical ranking.); Persisted result question/query keys at restart suppression: known — routing [wired], enforcement [wired] (Only keys are read for restart; retained rationale receives no demonstrated reassessment force.); Stored workspace text delivered to executor by requested read: known — knowledge [wired] (Returned text supplies context, with truncation; a script execution is a different consumer.); Executable workspace scripts/configuration and opaque tool-file consumers: not-determinable (Execution is wired but a generic executable payload does not establish one of the seven controlled memory forces without its content/effect. Binary displays and potential later media reuse cannot classify unseen payload authority.); Rich result rationale/tool traces and execution/process log content beyond restart keys: not-determinable (No implemented later rationale/log consumer is established. Their mere retention supplies no controlled force.)	AAS-2026-10-12-agentfly-bbfc9fb2d5c7-01
Agentic Local Brain	2	Includes collected document/metadata/chunk parts, mined entities/document vectors/relations/topics, topic wiki products and unresolved entity-card alternatives, conversation turns, reading events, retained results/confidence, mining status/history, local archives and their later consumers. Includes unresolved collector/maintenance/API alternatives and opaque provider accumulated state as coverage limits. Excludes static shipped prompts/configuration, code/install artifacts, development-agent/Trellis memory and external host state; does not substitute opaque provider model components for inspected accumulated memory.	partial coverage (Files, SQLite, vector storage and process-local retained state have wired witnesses. Remote accumulated state and untraced alternatives prevent complete coverage; source code and shipped prompt strings are not counted as memory substrates.); Collected document files and knowledge metadata: known — files [wired], sqlite [wired] (The cited record identifies this storage part and its retained read/write route.); Chunk text, IDs and dense search vectors: known — sqlite [wired], vector [wired] (The cited record identifies this storage part and its retained read/write route.); Entities, mentions and entity relations: known — sqlite [wired] (The cited record identifies this storage part and its retained read/write route.); Document embedding payloads and keys: known — sqlite [wired] (The cited record identifies this storage part and its retained read/write route.); Cross-document relation rows: known — sqlite [wired] (The cited record identifies this storage part and its retained read/write route.); Topic memberships, labels and centroids: known — sqlite [wired] (The cited record identifies this storage part and its retained read/write route.); Wiki article files and category/article metadata: known — files [wired], sqlite [wired] (The cited record identifies this storage part and its retained read/write route.); RAG sessions, turns and source dictionaries: known — sqlite [wired] (The cited record identifies this storage part and its retained read/write route.); Reading/search/RAG event rows: known — sqlite [wired] (The cited record identifies this storage part and its retained read/write route.); Enhanced result cache including confidence: known — in-memory [wired] (The cited record identifies this storage part and its retained read/write route.); Mining progress and JSONL history: known — in-memory [wired], files [wired] (The cited record identifies this storage part and its retained read/write route.); Locally retained backup archives and listing metadata: known — files [wired] (The cited record identifies this storage part and its retained read/write route.); Opaque generation, expansion, reranking and embedding provider state: not-determinable (Calling records identify remote weights/services but not an inspected use-accumulated provider memory inventory. No model-weights or service-object finding is inferred for opaque accumulated state.); Untraced alternative collectors, entity-card production, tag/export and exposed API retention paths: uninspected (Accepted records leave alternative writers and active callers incompletely inspected; local positives do not enumerate their retained parts.)	partial coverage (Known software admission remains automatic regardless of workflow initiation. Manual backup adoption has its own operator decision. Generic caller identity and alternative/provider admissions remain unresolved.); Missing metadata extraction and chunk/index admission: known — automatic [wired] (The initiating caller may select inputs/flags; this does not make generated admission manual.); Mining entities/vectors/relations/topics and topic-article compilation: known — automatic [wired] (The web full-mode confirmation is not review of each output.); Conversation turn/cache and reading-event admission: known — automatic [wired] (Cache hits, empty retrieval and save failures bound activation, not admission agency.); Topic centroid revision and title-cleanup admission: known — automatic [wired] (Starting maintenance is not manual admission of each resulting change.); Mining history and scheduled/archive packaging writes: known — automatic [wired] (Backup scheduling/selection is separate from subsequent archive admission.); Operator adoption of extracted backup material: known — manual [wired] (Manual adoption is distinct from automatic extraction; successful adoption is unobserved.); Initial file/note content, supplied metadata overrides, HTTP edits and requested withdrawal: not-determinable (Generic CLI/HTTP caller identity does not establish an explicit human per-content decision; manual_input is only a source label.); Untraced collectors, entity-card, tag/export/reset and expiry admission mechanisms: uninspected (Inspected admission witnesses do not resolve all included alternatives.); Provider accumulated-memory admission: not-determinable (No accessible provider-side admission control is established.)	partial coverage (Both requested reads and automatic context supply occur within request chains. The absence of an independent unsolicited scheduler does not eliminate push operations inside those chains. Unknown alternatives prevent complete coverage.); Requested knowledge retrieval: keyword/semantic, legacy/enhanced answering and mining input reads: known — pull [wired] (Requested delivery is not another push solely because selection/delivery is automatic.); Wiki article browsing, recommendation request, history/status and backup selection: known — pull [wired] (Identifier-based GET and requested archive selection are addressing, not targeted push.); Automatic session-history supply to expansion/answer generation: known — push [wired] (The outer question is requested; automatic contextual supply is a distinct operation within that chain, not a second push for the final answer return.); Automatic recent-view supply to enhanced-query enrichment: known — push [wired] (The caller requests an answer, not those particular event records. Recommendation-list return itself remains pull.); Cache reuse fulfilling repeated single-turn query: known — pull [wired] (Cache delivery fulfills the question request; no unsolicited delivery is inferred.); Entity-card, other API/maintenance consumers and provider-side delivery: uninspected (Untraced consumers/alternatives prevent an exhaustive request-versus-supply inventory.)	partial coverage (Session identity and coarse contextual selection support push findings. Semantic, lexical and model judgment used to fulfill requested retrieval are not automatically additional push signals. Unknown alternatives prevent complete coverage.); Session-history push: current query with session identity, recent-turn selector, formatted prior content: known — identifier [wired], coarse [wired] (Identity selection and coarse truncation are separate operations; saved source dictionaries are not delivered history text.); Recent-view push to enrichment: enabled query, latest-view selector, event IDs supplying score inputs: known — coarse [wired] (Matching chunk.source to viewed IDs conditionally boosts scores, but does not independently establish identity-selected delivered chunks; no identifier push finding is added for that boost.); Requested knowledge/wiki/recommendation/cache/status/archive returns: inapplicable (These operations are pull; push-signal classification is inapplicable locally, not for the mixed overall boundary.); Alternative and provider-side push selection: uninspected (Trigger, selector inputs and selected retained parts remain uninspected for included opaque alternatives.)	partial coverage (A wired conversational-input to durable assistant-text to later-context route supports yes. Other raw logging, transient selection and opaque input routes remain unresolved, not system-wide no. This axis is not a claim of improved capacity.); Automatic assistant-turn production/admission from conversational question/history to later model context: known — yes [wired] (This qualifies as trace-fed durable behavior-shaping text, not merely keeping raw logs. No authored continuation summary, parameter training, activation or improved capacity is required or asserted.); Reading events to ephemeral recommendation profile/score boost: not-determinable (The profile/reasons are recomputed, not durably written. Raw event persistence alone does not establish a qualifying derived update; no bounded absence record warrants a local no.); Mining-history status and archive routes: not-determinable (Status logging/archive copies do not establish durable trace-derived behavioral updates; absence of such updates is not a supplied bounded absence.); Collected documents, summary/mining/wiki routes with unresolved original trace content: not-determinable (Original inputs may include traces, but accepted records do not identify qualifying trace origins for these content routes.); Entity-card, remaining maintenance/collector and opaque provider trace-fed writes: uninspected (No complete inspection of these update/consumer chains is recorded.)	partial coverage (Knowledge, ranking, routing and a conversation-to-durable-answer learning path are locally supported. Static prompts/configuration and fixed program checks are excluded; no retained criterion is classified as instruction, validation or enforcement merely because code checks or deletes a row.); Document/chunk/entity/topic text and prior turns at expansion/answer models; articles at wiki readers: known — knowledge [wired] (Knowledge authority names delivery for interpretation, not observed obedience or correctness.); Retained text and dense embeddings at lexical/semantic retrieval; topic/graph associations at selection: known — ranking [wired] (Ranking influence does not imply learned scoring policy or calibrated relevance.); Retained reading-event IDs/times and associated vectors/topics at recommendation/enrichment: known — ranking [wired] (Transient user_profile is a consumer result, not an additional retained object.); Article source-ID/category state at recompilation selection: known — routing [wired] (This selects work, not factual acceptance or a universal freshness check.); Session/order fields at history loading; cache keys/TTL at result reuse: known — routing [wired] (Actual deployed reuse remains unobserved.); Retained topic centroids at subsequent document classification: known — routing [wired] (Classification updates and future use are wired, not demonstrated improved capacity.); Conversation experience at answer generation producing later-retained assistant turns: known — learning [wired] (This is the durable behavior-shaping artifact update path; downstream turns separately have knowledge authority. No parameter update or improvement is asserted.); Cached confidence and mining-history status at API/display consumers: known — knowledge [wired] (Confidence is advisory arithmetic metadata, not answer-truth probability or a veto.); Unconsumed user_notes, saved source dictionaries, archive payloads and remaining graph/topic consumers: not-determinable (Storage or display alone cannot establish every behavioral force or downstream consumer.); Entity-card and opaque provider accumulated-memory consumers: uninspected (Actual alternative consumption and effects are not recorded completely.)	AAS-2026-10-12-agentic-local-brain-f00cbd577178-01
Agno	2	Included: accumulated session/history/summary/state, MemoryManager rows, six built-in learning stores, generic and managed-page knowledge text/access indexes, compaction summary/boundary/transcripts, retained tool-result offload/compressed fields, use-changed component configuration, scored SFT exports and their later-consumer routes. Included unresolved alternatives: custom stores/readers/retrievers/capture hooks, FileSystemKnowledge, opaque media/provider-cache retained state, workflow checkpoints/forks and downstream training/replay. Excluded: static shipped skills/prompts/tool code, ordinary current-run-only state and work-dispatch tickets, inference weights absent an established use-fed update, external terminal/provisioning/hosted UI responsibilities. Known routes do not exclude opaque alternatives.	partial coverage (Supported file, vector, database and optional process-cache parts coexist with unresolved backends and opaque state. Storage of model configuration does not establish use-fed model-weight memory.); AgentSession database rows and optional full-session process cache — rdbms part: Configured session/run database persistence is wired; no uniform backend guarantee is inferred.: partial — rdbms [wired] (Database and optional cache are distinct retained parts. Backend-specific substrate refinements, including SQLite activation, are not recorded.); AgentSession database rows and optional full-session process cache — in-memory part: An optional full-session in-process cache holds retained sessions across invocations.: partial — in-memory [wired] (Database and optional cache are distinct retained parts. Backend-specific substrate refinements, including SQLite activation, are not recorded.); Built-in MemoryManager and five learning-row stores: known — rdbms [wired] (UserProfile, UserMemoryStore, SessionContext, EntityMemory and DecisionLog are database-row objects; this does not assert every adapter is inspected.); LearnedKnowledge and generic Knowledge content/vector retention: partial — vector [wired] (Vector storage is established; the configured content backend and opaque remote alternatives are not fully classified.); Managed-page published text, vector index and catalog — files part: Published page text is retained in the transactional filesystem.: partial — files [wired] (Distinct filesystem, vector and catalog parts are included; full search SQL and snapshot reconstruction remain unresolved.); Managed-page published text, vector index and catalog — vector part: Published page vectors feed later page/Knowledge search.: partial — vector [wired] (Distinct filesystem, vector and catalog parts are included; full search SQL and snapshot reconstruction remain unresolved.); Managed-page published text, vector index and catalog — rdbms part: Publication retains catalog state in the database transaction.: partial — rdbms [wired] (Distinct filesystem, vector and catalog parts are included; full search SQL and snapshot reconstruction remain unresolved.); Compaction summary/boundary and optional transcript archive: known — rdbms [wired] (Unsupported record backends can apply a fold without separate recoverability. In-session summary persistence follows the session path; no archive-on assumption.); ResultStore raw payload and access index — files part: Session-namespaced FileSystem holds the exact output string.: known — files [wired] (Preview is a derived display, not another payload substrate. Complete filesystem contracts are uninspected.); ResultStore raw payload and access index — rdbms part: A tool-result database row indexes payload path, scope and identity.: known — rdbms [wired] (Preview is a derived display, not another payload substrate. Complete filesystem contracts are uninspected.); Compressed-content field when retained with messages: partial — rdbms [wired] (Persistence is conditional and storage scrubbing is unresolved; current-only transformations are not counted as durable memory.); Changed component configuration versions and current pointer: known — rdbms [wired] (Only configurations admitted or changed through use are counted; static shipped instruction text and callable code are excluded.); Conversational-SFT JSONL and provenance sidecar: known — files [wired] (The artifact substrate is known independently of whether training subsequently occurs.); FileSystemKnowledge and knowledge reader/content-backend alternatives: uninspected (The report inventory includes filesystem retrieval, but canonical records do not establish a complete retained-object substrate inventory for that alternative. No separate positive is inferred from the class name.); Included media reconstruction, provider caches/opaque retained state, workflow checkpoints/forks, custom learning stores and remote knowledge alternatives: uninspected (The accepted inventory leaves these operative alternatives uninspected. Their storage substrate cannot be enumerated from the available records; known built-in routes do not classify them. Inference weights are excluded unless a use-fed retained update is established.)	partial coverage (Automatic software/model admission is established on explicit routes. No human-per-content manual finding is supported by generic caller identity; unresolved initial and direct doors remain included.); Automatic session/run retention, summary generation, compaction and result offload: known — automatic [wired] (Starting a run or manual fold request does not make software summary/payload admission manual.); Background MemoryManager and built-in conversation extraction: known — automatic [wired] (Background timing and final-trace coverage remain separate limits, not uncertainty about the inspected admission mechanism.); Foreground memory/entity/decision/learned-knowledge and add_to_knowledge tool writes: known — automatic [wired] (PROPOSE asks approval but writer checks no approval artifact. Optional Function confirmation is not assumed configured on these routes.); Curator removal/dedup, managed-page sync and SFT export admission: known — automatic [wired] (Maintenance invocation and source-sync initiation are not per-entry human approval.); Tool compression admission: known — automatic [wired] (Durable later replay remains unresolved, independently of the software admission finding.); Direct add/save/replace, AgentOS content edits and component-config requests/pointer moves: not-determinable (Generic API/caller identity does not say whether a human supplies, approves or replaces each retained item. Human control is afforded but cannot establish manual agency; neither physical DB I/O nor a supplied request alone establishes automatic agency for this admission door.); Initial user/profile/knowledge/state content and custom capture/admission: not-determinable (Known subsequent extraction does not resolve initial per-content control or custom hook/direct caller admission.); Included media reconstruction, provider caches/opaque retained state, workflow checkpoints/forks, custom learning stores and remote knowledge alternatives: uninspected (The accepted inventory leaves these operative alternatives uninspected. Their write agency cannot be enumerated from the available records; known built-in routes do not classify them. Inference weights are excluded unless a use-fed retained update is established.)	partial coverage (Automatic context supply and requested retrieval coexist. Pull-only subroutes do not make unresolved alternatives or the system push-selection axis inapplicable.); History, stored folds and session-summary automatic supply: known — push [wired] (A task request is not a separate request for each retained history/summary part.); MemoryManager user-scoped context: known — push [wired] (No query-targeted search_user_memories is performed on this inspected push.); LearningMachine automatic built-in recall and foreground searches: partial — push [wired], pull [wired] (Request fulfillment is not a second push. Custom stores and full alternative callers remain unresolved.); Caller-controlled manual LearningMachine context placement: known — pull [afforded] (The manual context/instructions door has a documented consumer role, not inspected external application wiring.); Knowledge current-query references and requested search: partial — push [wired], pull [wired] (Custom retriever content/selector policy is included but unresolved.); Managed-page chat search, compaction archive lookup and ResultStore reload: partial — pull [wired] (Page read/full-page/list/grep and all snapshot alternatives are not fully reconstructed.); Database-resolved component dispatch: known — push [wired] (Worker/client work initiation does not itself request each retained config instruction. Explicit pointer selection and host delivery are distinct operations.); SFT files at an external trainer: partial — pull [afforded] (External trainer is a documented afforded role, not an inspected running consumer.); Retained compressed-content at later history/provider invocation: not-determinable (Serialization permits retention but does not establish complete provider-specific later read-back; neither direction is inferred from current message mutation.); Included media reconstruction, provider caches/opaque retained state, workflow checkpoints/forks, custom learning stores and remote knowledge alternatives: uninspected (The accepted inventory leaves these operative alternatives uninspected. Their read-back direction cannot be enumerated from the available records; known built-in routes do not classify them. Inference weights are excluded unless a use-fed retained update is established.)	partial coverage (Push selectors are classified at trigger/input/selected-part level. Coarse limits, identity matching, lexical relevance and semantic-vector selection coexist. No inferred-judgment push is established by a write-side model or opaque reranker.); Automatic session-history/fold/summary push at invocation: known — identifier [wired], coarse [wired] (Trigger is message preparation; identities select actual retained parts, while budget/availability controls remain separate coarse selectors.); MemoryManager user push and LearningMachine identity stores: known — identifier [wired], coarse [wired] (Trigger is system-message preparation. User identity selection is not embedding or model-judged relevance.); LearnedKnowledge automatic current-message recall: partial — inferred-embedding [wired], identifier [wired], coarse [wired] (Recorded semantic/vector route supports embedding selection; custom backend/reranker and imported embedding provenance remain unresolved.); Entity automatic directory and relevance expansion: known — coarse [wired], identifier [wired], inferred-lexical [wired] (Trigger is recall before inference; selected entity text and aliases are actual selector inputs. Supersession judgment is a write-side evaluator, not inferred-judgment push selection.); DecisionLog automatic recent-decision supply: known — identifier [wired], coarse [wired] (Current session is deliberately not the selector. Stored outcome confidence is not a recorded judgment-based push policy.); Generic automatic Knowledge references: partial — inferred-embedding [wired], identifier [wired], coarse [wired] (Custom retrievers take precedence; optional query transformation and reranking are unresolved and cannot be upgraded by the inspected vector witness.); Retained component configuration at dispatch reconstruction: known — identifier [wired] (Trigger is component resolution for dispatch; this is identity selection of consumed configuration, not merely an ID appearing in a catalog.); Requested page/archive/result search and external SFT file reads: inapplicable (These identified operations fulfill explicit requests, so push-signal classification is inapplicable locally; automatic page/Knowledge selection is separately included above.); Compression replay selectors and custom/opaque push alternatives: not-determinable (Later compressed-content consumption and custom/opaque selector inputs are not recorded. Neither a known pull route nor a known identity push resolves these included alternatives.)	partial coverage (Several independent trace-fed durable context/guidance writes establish yes. Compression replay, external training and opaque capture alternatives prevent completeness. No local or system no is asserted without bounded absence records.); Automatic MemoryManager user-message extraction to later personalization: known — yes [wired] (Wired qualifying update is not a successful recall or improvement observation.); UserProfile/UserMemoryStore/SessionContext automatic updates to later system context: known — yes [wired] (Concurrent snapshots need not contain final current answer or tools; that limits provenance completeness, not the existing qualifying chain.); LearnedKnowledge conversation extraction to later query-selected guidance: known — yes [wired] (No new proposition, independent truth test or benefit is required for this route-level classification.); Session summary and compaction continuation fold: known — yes [wired] (Summary/boundary-only and transcript-bearing forms both qualify when retained/replayed. Unsupported archive writes do not establish separate durable recovery.); Tool compression retained replay: not-determinable (Tool traces feed automatic derived wording and serialization can retain it, but a complete later durable behavior-shaping provider consumer is not established.); SFT export through an external conversational trainer: not-determinable (Dataset production is wired and an external trainer role is afforded. Whether any trainer is configured, consumes the dataset or produces durable updated parameters is uninspected; export alone is not a closed qualifying training update.); Foreground entity/decision logging and direct/HTTP authored replacement as trace-learning mechanisms: not-determinable (Entity/decision process no-ops do not establish absence of learning on every foreground tool route. Tool-authored self-reports, caller edits and config requests lack an established trace-fed automatic derivation chain; no system/local no is inferred without a bounded absence record.); Raw session/offload trace retention and external-source sync as qualifying derived updates: not-determinable (Raw retention/reload and source import alone do not establish a derived behavior-shaping automatic update. The distinct summary/extraction chains carry the positive classifications; arbitrary trace-bearing source imports remain unresolved.); Included media reconstruction, provider caches/opaque retained state, workflow checkpoints/forks, custom learning stores and remote knowledge alternatives: uninspected (The accepted inventory leaves these operative alternatives uninspected. Their trace-learning mechanism cannot be enumerated from the available records; known built-in routes do not classify them. Inference weights are excluded unless a use-fed retained update is established.)	partial coverage (Each force is tied to a retained part and named consumer. Update learning, downstream guidance, ranking and symbolic vetoes coexist without licensing opaque consumers or observed benefit.); Retained conversational/fact/reference/summary text to later foreground model: known — knowledge [wired] (Delivery is wired but compliance is unobserved. Opaque payloads and custom consumers remain outside this resolved text unit.); Retained preferences, plans and organizational directives to later inference: known — instruction [wired] (Instruction is warranted by the content/effect, not simply by being placed in a system-message channel.); Retained messages/existing context to extraction and summary update consumers: known — learning [wired] (Learning force at the update consumer is distinct from output knowledge/instruction at later inference; no improved capacity is inferred.); Knowledge vectors/content and entity text/aliases at retrieval selectors: partial — ranking [wired] (Fixed scoring still makes retained content a ranking input. Complete page-search SQL and custom reranker effects are unresolved.); Fold boundaries, session identities, namespace and current-version state at reload/dispatch: known — routing [wired] (This does not attribute authority to unconsumed metadata. Detailed workflow checkpoint consumers remain unresolved.); Entity retirement/archive flags and ResultStore access-index scope at delivery checks — enforcement part: Retirement flags suppress ordinary live display; index scope mismatches refuse requested payload delivery.: known — enforcement [wired] (Retired facts remain searchable in named alternate routes. User matching requires both stored/current user IDs; this is not universal authentication.); Entity retirement/archive flags and ResultStore access-index scope at delivery checks — validation part: Retained index session/user fields define conformity criteria for an admissible read request.: known — validation [wired] (Retired facts remain searchable in named alternate routes. User matching requires both stored/current user IDs; this is not universal authentication.); Changed component instruction/resource config and retained published/current-stage state — instruction part: Reconstructed retained instruction content directs later foreground inference.: known — instruction [wired] (These values concern retained configuration parts, not static tool code or every HTTP/backend control.); Changed component instruction/resource config and retained published/current-stage state — routing part: Current-version pointer and model/tool references select later reconstruction/dispatch.: known — routing [wired] (These values concern retained configuration parts, not static tool code or every HTTP/backend control.); Changed component instruction/resource config and retained published/current-stage state — validation part: Retained stage/ownership/version state defines admissible updates and pointer moves.: known — validation [wired] (These values concern retained configuration parts, not static tool code or every HTTP/backend control.); Changed component instruction/resource config and retained published/current-stage state — enforcement part: Published-row immutability and draft-current rejection veto inadmissible changes in inspected PostgreSQL paths.: known — enforcement [wired] (These values concern retained configuration parts, not static tool code or every HTTP/backend control.); Scored attempt records at SFT exporter — validation part: Stored score and trace/exportability fields determine dataset eligibility.: known — validation [wired] (The effect is dataset selection, not factual acceptance. only_passed=False changes failed-attempt eligibility.); Scored attempt records at SFT exporter — enforcement part: Exporter excludes unscored, disallowed failed and tool-bearing attempts before file admission.: known — enforcement [wired] (The effect is dataset selection, not factual acceptance. only_passed=False changes failed-attempt eligibility.); Exported conversational examples at an external trainer: partial — learning [afforded] (The record names external training as an afforded consumer; no training execution or installed weight update is established.); Compressed-content field at retained later-history/provider consumer: not-determinable (Transformation and serialization are recorded, but provider-specific later replay is not established. Current-payload substitution alone does not determine retained authority.); Included media reconstruction, provider caches/opaque retained state, workflow checkpoints/forks, custom learning stores and remote knowledge alternatives: uninspected (The accepted inventory leaves these operative alternatives uninspected. Their behavioral authority cannot be enumerated from the available records; known built-in routes do not classify them. Inference weights are excluded unless a use-fed retained update is established.)	AAS-2026-10-13-agno-60e6e09dd4a6-01
AI Agents in Depth	2	Included: accumulated user-memory notes/basic and advanced cards, raw persisted dialogue and optional opaque Dify history, live context-agent history and its material parts, sequential derived states/checkpoint controls and reused receipts, generated retry-policy candidates and rejected-check feedback; all other included suite memory alternatives remain unresolved. Excluded: static shipped prompts/rules/stable source as accumulated memory, inference weights without recorded through-use updates, unconsumed terminal reports and judge evidence as later-agent memory, and external reproduction/host deployment loops outside the frozen source.	partial coverage (Positive classifications cover the named inspected routes only. Included alternatives remain unresolved; this is not a suite-complete inventory of values. Storage and remote operative-state coverage uncertainty is retained in the corresponding units.); User-memory note content and addressing metadata: known — in-memory [wired], files [wired] (Local retained objects and file stores are inspected; optional remote history is split below.); Basic hierarchical memory cards: known — in-memory [wired], files [wired] (Local retained objects and file stores are inspected; optional remote history is split below.); Advanced category cards: known — in-memory [wired], files [wired] (Local retained objects and file stores are inspected; optional remote history is split below.); Raw user-memory conversation turns: known — in-memory [wired], files [wired] (Local retained objects and file stores are inspected; optional remote history is split below.); Sequential derived states and checkpoint controls/receipts: known — in-memory [wired], files [wired] (Local retained objects and file stores are inspected; optional remote history is split below.); Live context-agent history and its PDF/currency/reply parts: known — in-memory [wired] (Optional single-task report files are excluded here: no inspected reload consumer is recorded.); Retry-policy generated source and rejected-check feedback in the campaign: known — in-memory [wired], files [wired] (Stable shipped source is an input, not accumulated memory; live installation is not established.); Optional Dify retained dialogue index and consumed search payload: not-determinable (Index requests carry text and metadata, but provider storage/embedding semantics and a shipped later model-search consumer are opaque. A vector store cannot be inferred from semantic-search parameters.); Other included suite memory routes: mem0, memobase, agentic-rag-for-user-memory, contextual-retrieval-for-user-memory; retrieval-pipeline, dense-embedding, sparse-embedding, structured-index, structured-knowledge-extraction, log-sanitization, agentic-rag, contextual-retrieval; other chapter loops, skills, fixtures and results.: uninspected (These alternatives remain uninspected in the accepted memory inventory. Their storage and remote operative-state coverage cannot be enumerated or classified from directory names; no negative or complete-value conclusion follows.)	partial coverage (Positive classifications cover the named inspected routes only. Included alternatives remain unresolved; this is not a suite-complete inventory of values. Initial content and admission mechanisms uncertainty is retained in the corresponding units.); Background extraction and model mutation admission: known — automatic [wired] (Manual process/save starts analysis; it does not make its ensuing fact admissions manual.); Sequential JSON state replacement and checkpoint admission: known — automatic [wired] (Judge grade reports results rather than vetoing prior memory admission.); Explicit human CLI content editing: known — manual [wired] (Manual is bounded to the operator route, not all calls to the manager.); Deterministic consolidation, eviction and raw interactive recording: known — automatic [wired], automatic [wired] (Human dialogue authorship is separate from automatic recording/admission.); Retry candidate generation and artifact admission: known — automatic [wired] (The launch is human-configured; production installation is not recorded.); Live context-history append: known — automatic [wired] (No per-message approval is recorded.); Initial files and generic direct apply/update callers: not-determinable (Caller-dependent initial content and exposed helper inputs do not identify human versus software control. Physical file I/O cannot decide admission agency.); Other included suite memory routes: mem0, memobase, agentic-rag-for-user-memory, contextual-retrieval-for-user-memory; retrieval-pipeline, dense-embedding, sparse-embedding, structured-index, structured-knowledge-extraction, log-sanitization, agentic-rag, contextual-retrieval; other chapter loops, skills, fixtures and results.: uninspected (These alternatives remain uninspected in the accepted memory inventory. Their initial content and admission mechanisms cannot be enumerated or classified from directory names; no negative or complete-value conclusion follows.)	partial coverage (Positive classifications cover the named inspected routes only. Included alternatives remain unresolved; this is not a suite-complete inventory of values. Later-consumer routes and unresolved search consumers uncertainty is retained in the corresponding units.); Conversation/analysis/context automatic prompt supply: known — push [wired] (The user requests current work, not the particular retained items supplied by the host.); Sequential latest-state supply and checkpoint response reuse: known — push [wired] (Requested current evaluation is distinct from automatic selection of past material.); CLI query/show of retained user memory: known — pull [wired] (The local request and requested return are one pull operation, not an extra push.); Rejected-control summary to subsequent coding proposal: known — push [wired] (This is a within-campaign later invocation, not an installed successor reload.); History-search API and opaque Dify later consumption: not-determinable (A storage/search API with unspecified caller does not establish the later consumer route. Neither an end-to-end pull nor a push can be classified from that capability alone.); Other included suite memory routes: mem0, memobase, agentic-rag-for-user-memory, contextual-retrieval-for-user-memory; retrieval-pipeline, dense-embedding, sparse-embedding, structured-index, structured-knowledge-extraction, log-sanitization, agentic-rag, contextual-retrieval; other chapter loops, skills, fixtures and results.: uninspected (These alternatives remain uninspected in the accepted memory inventory. Their later-consumer routes and unresolved search consumers cannot be enumerated or classified from directory names; no negative or complete-value conclusion follows.)	partial coverage (Positive classifications cover the named inspected routes only. Included alternatives remain unresolved; this is not a suite-complete inventory of values. Actual automatic selectors and delivered parts uncertainty is retained in the corresponding units.); All retained user facts and current-session raw history on chat: known — coarse [wired], identifier [wired] (Whole-store availability and actual identity-filtered raw-history selection are separate push operations.); Recent/unprocessed dialogue selection for background analysis: known — coarse [wired], identifier [wired] (IDs are marked before successful analysis; this selector is not a durability or correctness guarantee.); Live context-agent mode/history supply: known — coarse [wired] (NO_HISTORY withholds stored memory; symbolic tool IDs are associations, not an evidenced identity-targeted relevance selector.); Sequential checkpoint and latest-state/receipt selection: known — coarse [wired], identifier [wired] (Output token budget limits generated memory, not relevance selection of current-state input.); Rejected control feedback push: known — coarse [wired] (The candidate hash is supplied metadata, not evidence of a hash-match selector for delivered feedback.); CLI request-driven query/show only: inapplicable (Substring selection answers an explicit retained-material request. Push-signal classification is inapplicable to this pull operation, not to the broader mixed boundary.); Opaque history-search later-consumer branch: not-determinable (Provider semantic/reranking parameters do not establish a consumed push payload or its operative selector. Known pull CLI routes do not resolve this alternative.); Other included suite memory routes: mem0, memobase, agentic-rag-for-user-memory, contextual-retrieval-for-user-memory; retrieval-pipeline, dense-embedding, sparse-embedding, structured-index, structured-knowledge-extraction, log-sanitization, agentic-rag, contextual-retrieval; other chapter loops, skills, fixtures and results.: uninspected (These alternatives remain uninspected in the accepted memory inventory. Their actual automatic selectors and delivered parts cannot be enumerated or classified from directory names; no negative or complete-value conclusion follows.)	partial coverage (Positive classifications cover the named inspected routes only. Included alternatives remain unresolved; this is not a suite-complete inventory of values. Qualifying automatic trace-fed writes uncertainty is retained in the corresponding units.); Automatic user-memory dialogue extraction and later answer reload: known — yes [wired] (This qualifies structurally despite swallowed save failures and absent observed benefit.); Sequential session-fed checkpoint memories and later writer/answer: known — yes [wired] (Within-run and resumed consumption qualify; source-based judge grades need not feed a further update.); Automatic retry-policy candidate from supplied failure experience: known — yes [wired] (A durable formal behavior-shaping artifact qualifies without live installation; this does not establish exercised self-improvement or learned weights.); Live raw context-history capture without a recorded durable derived update: not-determinable (Raw logging and later visibility alone do not establish a durable trace-derived behavior-shaping write. No ABS record supports a local no; this route remains not-determinable for a stronger negative.); Generic initial/direct retained content and opaque remote branches: not-determinable (Initial/admission origins and remote update-to-later-consumer chains are unresolved. They cannot be made no by an inspected route elsewhere.); Other included suite memory routes: mem0, memobase, agentic-rag-for-user-memory, contextual-retrieval-for-user-memory; retrieval-pipeline, dense-embedding, sparse-embedding, structured-index, structured-knowledge-extraction, log-sanitization, agentic-rag, contextual-retrieval; other chapter loops, skills, fixtures and results.: uninspected (These alternatives remain uninspected in the accepted memory inventory. Their qualifying automatic trace-fed writes cannot be enumerated or classified from directory names; no negative or complete-value conclusion follows.)	partial coverage (Positive classifications cover the named inspected routes only. Included alternatives remain unresolved; this is not a suite-complete inventory of values. Actual consumers and forces uncertainty is retained in the corresponding units.); Retained notes/cards and live dialogue at answer/analysis models: known — knowledge [wired] (Preference content in user-message context does not itself establish instruction authority.); Retained dialogue/current state at memory-update consumers: known — learning [wired], knowledge [wired] (Update-consumer learning authority and downstream answer knowledge are distinct paths; neither proves benefit.); Checkpoint controls and stored call results at resume host: known — routing [wired], validation [wired], enforcement [wired] (These forces belong to control metadata, not arbitrary memory fact values; fixture/prompt freshness remains unchecked.); Retained note content/tags/time and card identifiers at maintenance selectors: known — routing [wired] (Selection is supported; no retained learned ranking policy or relative relevance-score route is established here.); Rejected-check experience and candidate code at generation/evaluation consumers: known — learning [wired], knowledge [wired], enforcement [wired] (The candidate rules have evaluator authority only; manifest eligibility does not extend them to live stable operation.); Optional Dify index/search output at a later model consumer: not-determinable (Search requests mention reranking but no shipped later model-search consumer is recorded and provider scoring is opaque. Relative ranking authority of retained remote content is not established.); Other included suite memory routes: mem0, memobase, agentic-rag-for-user-memory, contextual-retrieval-for-user-memory; retrieval-pipeline, dense-embedding, sparse-embedding, structured-index, structured-knowledge-extraction, log-sanitization, agentic-rag, contextual-retrieval; other chapter loops, skills, fixtures and results.: uninspected (These alternatives remain uninspected in the accepted memory inventory. Their actual consumers and forces cannot be enumerated or classified from directory names; no negative or complete-value conclusion follows.)	AAS-2026-10-12-ai-agent-book-d56e8121d636-01
AI Research OS Workshop	2	Retained topic-local catalog/navigation, raw documents/highlights/assets, source/topic/comparison/thesis/repository/question pages, rolling criticism/gaps and operation log; reusable GitHub clone; exported research appendix and prose/chart/canvas render parts. Includes their admission, maintenance and later-consumer paths, including unresolved query save/log-tally and appendix-delivery alternatives. Excludes static shipped instructions/helpers, deleted ordinary current-run discovery/lint scratch and immediate unsaved answers, host conversation/compaction and external model/source-service retention or embedding internals.	known coverage (Inventory covers topic-local retained files, exported appendix and reusable Git clone. Provider/host retention and static plugin files are excluded, not classified as file memory.); Catalog, navigation, all raw text/highlights/assets, wiki/criticism/question/log files, appendix, prose renders, chart program/PNG and canvas/sidecar: known — files [afforded] (Each operative payload or access part is retained in a filesystem file; YAML/JSON/Markdown/Python are encodings, not separate substrates. Model-written landing is afforded; conflicting destinations do not change the substrate.); Executable catalog and navigation persistence: known — files [wired] (Inspected helpers write canonical YAML and generated navigation directly to files.); Reusable GitHub source clone: known — repo [wired] (Clone helper retains and refreshes a Git repository; this mutable cache is included, unlike excluded provider storage. Curated specifications remain files in the preceding unit.)	partial coverage (Explicit source supply, bookmarks and overrides are manual; subsequent software/model content admission is automatic. Generic caller identities and original source histories do not resolve initialization control.); Explicit seed/source identity supply, deliberate wiki KEEP edits, optional propagation redirects and explicit query bookmarks: known — manual [afforded] (A human makes the content/scope supply or override decision. Starting the enclosing workflow alone is not the basis.); Model wiki/spec/comparison/criticism/appendix/render admission and two-source query autosave: known — automatic [afforded] (Model/software admits formulated content without per-content operator approval; format selection does not make the generated claims manual. Query landing remains conflicted but software admission is the requested mechanism.); Catalog/navigation normalization and persistence: known — automatic [wired] (Software projects and writes caller rows without per-row human approval.); PDF/asset extraction and reusable Git clone replacement: known — automatic [wired] (Executable conversion/cache refresh admits retained output computationally, distinct from human source selection.); Initial external source, service summary and highlight control: not-determinable (Upstream human highlight curation is asserted, but actual editing/admission history and caller control for other initial material are unresolved; local copy does not establish that history.)	partial coverage (Query/export/viewing, summary/judge self-directed reads and regeneration are pull. Append prior-page assignment and render input selection are afforded push. Appendix delivery and log tally remain unresolved.); Question/load/filter/drill/compose, missing-YAML fallback and corpus-relative distill requests: known — pull [afforded] (These consumers request retained research; fulfilling the request is pull, not an extra push.); Navigation generation reload of retained YAML: known — pull [wired] (Requested helper generation reads the supplied canonical catalog.); Append prior-page assignment and render delegated input selection: known — push [afforded] (Push is bounded to upstream append/render assignment, not summary or judge consumer requests.); Summary writer catalog/rolling-file/discretionary reads and semantic judge qualifying-page reads: known — pull [afforded] (Catalog/frontmatter, rolling files and discretionary or eligibility-based reads belong to consumer pull.); Viewer opening navigation/canvas/render and requested chart rerun: known — pull [afforded] (Opening a retained artifact or rerunning its snapshot program requests that material. No independent unsolicited trigger is established.); Exported appendix delivery to later writer and ingest-log chart tally: not-determinable (Later writer role is documented but delivery trigger is unspecified; canonical constructor omits ingest_log, so that alternate chart read path is not established.)	partial coverage (Signals describe push selectors only. A requested URI is not an identifier push; unresolved appendix delivery/log tally prevent a complete pull-only or signal-inapplicable classification.); Append propagation to entity/concept writer using slug and prior page identity: known — identifier [afforded] (Trigger: append/init propagation. Input: source-writer slug identities and counts plus assigned existing-page path. Selected retained part: matching prior entity/concept page supplied for rewrite. Identity matching is distinct from the mention-count availability gate.); Summary writer catalog/rolling-file/discretionary selection and semantic judge qualifying-page selection: inapplicable (These consumers request their own retained inputs. Availability, five/eight-page budgets and model page judgment constrain pull, not a recorded upstream push selector; no coarse or inferred-judgment push finding is established here.); Render delegated wiki input set: known — coarse [afforded] (Trigger: render request. Input: user framing and selected source-page list. Selected retained part: assigned wiki bodies delivered to format writers; upstream operator selection is allowed, but records establish no exact identity-match selector beyond the supplied availability list.); External appendix delivery and log-tally selection: not-determinable (No recorded delivery trigger/selector establishes a controlled push signal for these alternatives.)	not-determinable coverage (External research transformations are not automatically trace learning. No bounded absence record covers all included automatic writes; query save and the log-tally alternative leave qualifying trace input/consumer facts unresolved.); Automatic saved question/answer-content production: not-determinable (The two-source trigger can retain question/pointers and proposed answer-derived knowledge for future agents. Records do not establish consumption of a recorded session log as input to this write, and explicitly distinguish the result from the original conversation; no yes or system-wide no is licensed.); Durable operation summaries and optional chart log tally: not-determinable (Operation summaries are not original conversation/tool traces; no automatic log-to-guidance update is established. Missing ingest_log construction leaves the alternative consumer unresolved.); Acquisition/wiki/lint/appendix/render routes over external-source knowledge: not-determinable (Their supported chains use research content rather than established qualifying trace-fed inputs. Without bounded absence records this does not establish no for the whole inventory; host compaction and provider learning are excluded.)	partial coverage (Research, clone, asset, access and export consumers support local findings. Operation-log consumption remains unresolved, so authority coverage is partial. Afforded effects are not observed activation.); Source/wiki/thesis/specifications, catalog summaries, question pointers and criticism at later query/write/lint/distill/render models: known — knowledge [afforded] (Selected retained content supplies research evidence, reference or context; it is not a system instruction merely because it is read.); Reusable clone at specification writers and raw visual assets at source writers: known — knowledge [afforded] (Named retained inputs and consumers are distinct from derived specifications and raw text.); Operation log and proposed chart ingest-log tally: not-determinable (No actual later-model log consumer is established; the proposed tally lacks ingest_log omitted by canonical construction. Retention supplies no authority value.); Appendix at documented later writer; prose/canvas/PNG at viewer or draft consumer: known — knowledge [afforded] (Retained exports/presentations supply knowledge to their documented consumers; deployed delivery is unobserved.); Catalog text/tags/scores at query/filter and distill relevance decisions: known — ranking [afforded] (Retained strings and relevance scores influence relative source priority; no learned ranking policy is required.); Catalog URIs/identities, navigation links, saved answer pointers and canvas file nodes at later selection/display: known — routing [afforded] (These retained access parts select a source, document or display destination.); Prior catalog original_path set at subsequent ingest: known — validation [afforded], enforcement [afforded] (Known identities guide skipping already admitted sources; this is a prompted admission criterion and veto, not filesystem interception.); Prior comparison subjects/backing counts and existing rolling findings at writer/judge admission: known — validation [afforded], enforcement [afforded] (Retained subjects/support and prior findings supply sufficiency/duplicate criteria; refusal or skip requests veto candidate writes.); Existing render prompt/path metadata at noop or overwrite gate: known — validation [afforded], enforcement [afforded] (Retained metadata tests whether the candidate is the same requested render and can cause noop, overwrite choice or skip; changed source bodies are not checked.); Deliberate KEEP override blocks and aliases at later wiki/summary/comparison rewrites: known — instruction [afforded] (Retained user edits and markers direct preservation during subsequent revision; their force is instruction, not enforced interception.); Chart embedded constants at later Python rerun: known — knowledge [afforded] (Snapshot data supplies the retained numerical content from which the chart is produced; successful rendering would not certify it.)	AAS-2026-10-12-ai-research-os-workshop-1a93869f8e7f-01
AI-Context-OS	2	Retained canonical L0/L1/L2 and control metadata; usage aggregates; staged capture text, opaque attachments and proposals; generated discovery and scratch exports; journal, task, daily/session and observability stores; initialization, editing, derivation, governance and restore routes; and their later model, host, code and human consumers. Possible accumulated provider-side state remains an unresolved alternative, not an assertion of such storage. Excludes shipped static instructions/weights as such, ordinary component chat-turn accumulation, transient context/manifest values, private/later product revisions and external host implementations. Existing-content origins, opaque payloads, incomplete alternate mutations and missing consumers remain included uncertainties rather than disappearing from the inventory.	partial coverage (Local files and SQLite are established. Possible provider-side accumulation remains unresolved, preventing a complete storage inventory. Transient manifests, context packages and component turn state are not additional retained substrates.); Local canonical content, control/usage sidecars, capture items and attachments, proposals, generated discovery/handoff, journal/task/log files and ZIP backups: known — files [wired] (The declared local file objects cover canonical, derived, staged and operational file retention. Generated scratch output has file storage even though no later reader is established.); Request telemetry, health snapshots and optimization suggestions in the observability database: known — sqlite [wired] (Storage is resolved independently of the validity of health scores or suggested edits.); Possible accumulated provider-side state behind the configured inference interface: not-determinable (Provider internals are uninspected; the records do not establish whether accumulated state exists or its retention substrate. Shipped weights and ordinary forwarded turns do not settle this alternative.)	partial coverage (Automatic derivatives and specific manual governance actions are established. Generic command callers, initial content, alternate mutations and provider internals prevent a complete agency inventory.); Automatic access/telemetry accounting, generated discovery, journal derivatives, proposal production and optimization/health artifact admission: known — automatic [wired] (Known automatic derivatives are kept separate from their input authorship and any later apply decision.); Confirmed governance UI deletion and explicit importance boost: known — manual [wired] (Classification belongs to these explicit operator-decision mechanisms, not every invocation of the same save/delete APIs.); Onboarding selected-template admission: known — automatic [wired] (This unit covers derived template file admission only; arbitrary existing-vault admission and profile-field caller control remain separate.); Generic caller-controlled canonical/raw writes, capture edits, proposal apply/reject, direct operational commands, profile input, restore replacement and initial external-vault content: not-determinable (Commands admit caller values or selected artifacts, but generic caller identity is not an explicit human per-content decision. Serialization is not evidence of automatic admission control either.); Full rename/move/duplicate admission mechanisms: uninspected (The accepted memory annotation retains these as alternate mutations without a full control audit; their content-specific admission control is not supplied.); Possible accumulated provider-side state behind the configured inference interface: not-determinable (Provider internals are uninspected; the records do not establish whether accumulated state exists or its write/admission control. Shipped weights and ordinary forwarded turns do not settle this alternative.)	partial coverage (Pull has wired and afforded witnesses. Missing later consumers and opaque provider alternatives prevent a complete pull-only conclusion or system-wide absence of push.); Requested query context, exact skill/body reads and local operational API/UI reads: known — pull [wired] (Known local read-back is request-driven, including internal reads performed to satisfy the request. Query ranking does not independently establish push.); Documented host discovery and subsequent canonical-file reads: known — pull [afforded] (The documented external reader role supplies affordance rather than deployed host wiring.); Session logs, opaque attachments and connector scratch export beyond established delivery: not-determinable (The records establish storage/export but no later session-log, binary-extraction or specific scratch reader. Their direction cannot be inferred from file paths.); Possible accumulated provider-side state behind the configured inference interface: not-determinable (Provider internals are uninspected; the records do not establish whether accumulated state exists or its later read-back direction. Shipped weights and ordinary forwarded turns do not settle this alternative.)	not-determinable coverage (No controlled push signal is established in the accepted records. Inapplicability is local to assessed pull routes; unresolved directions prevent axis-wide inapplicability. This is not a negative claim about provider or host selection.); Inspected requested query/body/operational and documented direct-host read chains: inapplicable (These particular direction units are pull. Lexical query scoring, requested IDs and catalog paths are not push selectors, so no push signal is assigned to them.); Unestablished later session-log/attachment/scratch readers and possible provider memory selection: not-determinable (No complete direction assessment or push trigger, selector input and selected retained part is recorded for these alternatives. The known pull branches cannot make the whole signal axis inapplicable.)	partial coverage (Supported local yes findings give a supported system yes, with partial coverage. No local no is asserted: the accepted set supplies no evidenced-absence record warranting it. Missing semantic rewrites, observed improvement or a chat compactor do not negate trace-fed usage/advice routes.); Read/serve events automatically updating durable usage features for later ranking: known — yes [wired] (The raw-to-derived-to-later-consumer chain is recorded. Operation, changed rankings and successful capacity improvement remain unobserved.); Retained serving telemetry automatically transformed into persisted maintenance advice for later maintenance readers: known — yes [wired] (The qualifying output is durable maintenance advice, not the requested compression or an applied-status canonical revision.); Journal-derived tasks/date tags, caller-authored session summaries and arbitrary captured content used for proposals: not-determinable (Automatic derivatives exist, but original input content is not established uniformly as experience/traces; session logs additionally lack an established later consumer. No qualifying trace-fed chain or bounded no can be assigned across these alternatives.); Possible accumulated provider-side state behind the configured inference interface: not-determinable (Provider internals are uninspected; the records do not establish whether accumulated state exists or its trace-fed write and later-consumer chain. Shipped weights and ordinary forwarded turns do not settle this alternative.)	partial coverage (All seven forces have local supported consumers. Unknown consumers and opaque alternatives prevent a complete authority assessment; neither advisory delivery nor an Applied status proves activation or epistemic acceptance.); Descriptive canonical text at model/host readers and operational prose at journal/task/proposal/maintenance readers: known — knowledge [wired], knowledge [afforded] (Wired internal/API consumers and afforded external discovery retain different evidence bases.); Retained rule/skill content at MCP host readers and adapter-reading external hosts: known — instruction [wired], instruction [afforded] (Instruction force is advisory natural language, not a universal mutation gate.); Retained text/features, importance/tags/links and usage aggregates at scoring and dependency selection: known — ranking [wired], routing [wired] (These are actual fixed-code effects; confidence and provenance are not assumed to supply an acceptance threshold.); Protected metadata at guarded structured-save, MCP-save and deletion consumers: known — validation [wired], enforcement [wired] (The guard consumer is known. Broader protection is expressly not inferred.); Read/serve experience at the usage-update consumer and telemetry at the maintenance-advice generator: known — learning [wired], learning [wired] (Learning authority here names the update consumer, not improved capacity or execution of the suggested repair.); Session-log prose, opaque attachment bytes, unspecified later scratch-handoff consumers and incompletely inspected alternate mutation/host/provider paths: not-determinable (Storage or display summaries do not establish the missing later consumer and effect. Provider behavior and full alternate mutation handling are not resolved by one guarded or delivered route.)	AAS-2026-10-18-ai-context-os-82777c0c3926-01
ai-memex-cli	2	Local accumulated raw/session documents, wiki text and metadata, live-schema additions, catalog/action history, semantic findings, recomputed access index, scene manifests, context registry, local copies and installed host projections, including their admission, maintenance and later-consumer routes. Excludes unchanged shipped workflow text, ordinary transient invocation/lint state without durable projection, external agent/provider histories, weights and checkpoints, and qmd internal indexes; qmd-backed requested retrieval remains a pull route, not a classification of its internal storage.	known coverage (The local retained-object and access-structure inventory is covered. Files, not Markdown encoding or the product Git description, establish the substrate; external provider and qmd internals are outside this memory boundary.); Raw and session documents, wiki bodies and headers, live schema, catalog/log, local copies, scene manifests, registry and host bootstrap files: known — files [wired] (Each retained local artifact has a filesystem location in the accepted inventory.); Recomputed wiki index used to access retained pages: known — in-memory [wired] (An access structure is rebuilt from page files rather than persisted as a separate index.)	partial coverage (Software/model admission is recorded for deterministic and delegated paths. Generic CLI invocation is not human content approval. Explicit source selection and optional save decisions remain local, while manifest initialization and some caller-supplied content control are unresolved.); Mechanical session conversion, delegated distillation/ingestion/semantic repair, scaffold/template seeding, migration, projections and context regeneration: known — automatic [wired] (Known software/model decisions admit these outputs without a wired per-content human gate.); Interactive direct keyword-fetch result selection: known — manual [wired] (This branch requires an explicit human choice of returned sources, not verification of their claims.); Yes-all/direct rule-based and delegated capture admission: known — automatic [wired] (Automatic result admission is a distinct alternative to the interactive decision.); Foreground save/archive under installed query reference: known — manual [afforded] (An explicit user save decision is an afforded admission alternative, not a universal gate.); Foreground durability-based saving under vault-schema doctrine: known — automatic [afforded] (The conflicting schema alternative delegates the admission decision to model judgment.); Initial manifest creation, arbitrary source/manual edits and caller-supplied log entries: not-determinable (The records do not identify a human content decision for every generic caller or an initial manifest writer. Authorship or software I/O alone cannot settle admission control.)	known coverage (Coverage includes explicit retrieval requests and automatic preselection/projection supply. Delivery into files and prompts is wired; actual external host loading remains unobserved.); Search, foreground page reads, explicit injection and project-local retrieval/projection requests: known — pull [wired], pull [afforded] (These operations fulfill a consumer request for retained material; their return is not another independent push.); Automatic schema/catalog selection into maintenance-child prompts: known — push [wired] (A target workflow request does not itself request each retained guidance/catalog part; code preselects them for the child.); After-clean refresh into registered later-host context files: known — push [wired] (The automatic clean-triggered projection is distinct from a subsequent explicit host page lookup.)	known coverage (Signals describe only included push operations, not explicit query terms or addressed pull files. Registry host keys choose destinations; they do not by themselves identify selected page payloads.); Per-invocation maintenance schema/catalog preselection: known — coarse [wired] (Trigger is a management invocation; selection uses availability and the configured schema/catalog roles, not task matching.); Clean-triggered host digest construction: known — coarse [wired] (Trigger is a zero-issue report; selector uses the reloaded wiki inventory/order and configured mode; selected parts are count-bounded page names/descriptions.); Bound-scene manifest selection during host projection: known — identifier [wired] (Explicit scene names select corresponding retained manifests for rendered trigger guidance. This does not filter the full page digest and bulk refresh can lose bindings.)	partial coverage (The qualifying positive is an automatic session-fed schema update and reload chain, not demonstrated improvement. Mechanical conversion alone is not learning; open source-to-wiki alternatives are not made negative merely because their inputs are missing.); Delegated session extraction to durable schema learning note to later maintenance prompt: known — yes [wired] (The code requests a trace-fed durable guidance update and later reloads that artifact. The wired basis does not assert compliance or benefit.); Session-derived summaries/pages used as later knowledge context: partial — yes [wired] (The summary extraction/retention and ordinary later-consumer paths are wired; open ingestion can also take non-trace sources and resulting artifacts are not supplied.); Other source-to-wiki, semantic-repair and foreground-save writes: not-determinable (No complete input/output inventory identifies every additional qualifying automatic trace-fed update. Neither the absence of examples nor mechanical conversion by itself warrants a system-wide no.)	partial coverage (Recorded knowledge, instruction, ranking, routing, validation and learning paths remain distinct. Initial scene authorship is not needed to identify rendered guidance, but no actual host use or exhaustive model-internal effects are established.); Retained raw/session/wiki material and catalog/findings supplied to interpreting maintenance or answering consumers: known — knowledge [wired] (These parts supply context or evidence rather than a factual acceptance oracle.); Live-schema experience notes at later maintenance agents: known — instruction [wired] (The live schema is reloaded into later child prompts, including appended notes if written.); Rendered scene triggers and host bootstrap at later project hosts: known — instruction [wired] (The host-loading contract supplies guidance; its activation is not observed.); Page name/ID/tag/description fields at fixed keyword scoring: known — ranking [wired] (Ranking uses metadata rather than page bodies; no learned ranker is implied.); Include directives and context registry at file selection and refresh dispatch: known — routing [wired] (Access/control fields select files or destinations, not truth or user permission.); Retained page-ID inventory at link membership evaluation: known — validation [wired] (The retained IDs provide a bounded conformance criterion; advisory diagnostics do not enforce rejection of page content.); Recorded session experience at the delegated schema-note update consumer: known — learning [wired] (Learning authority at the update consumer is separate from instruction authority at the later maintenance consumer.); Arbitrary resulting model-authored pages, saved answers and opaque semantic judgments: not-determinable (The open model routes have no supplied output corpus or consumption trace; further standing directions, enforcement or other effects cannot be classified from page-type names.)	AAS-2026-10-15-ai-memex-cli-4324549b4bc0-01
ai-modules	2	Accumulated raw captures and metadata (including unresolved image assets), wiki declarative/procedural pages, locally customized schema and lint policies, index/action/audit metadata, live/archive task contracts and completion state, later-invocation relay evidence, incremental changelog, retained evaluator/benchmark artifacts, and unresolved remaining shipped-family retained-output consumers. Excludes static shipped skill/role/registration/deployment content, unchanged standing instructions, ordinary immediate call state, standalone summary outputs without a retained reload contract, and host inference/provider state with no established accumulated-memory write route.	partial coverage (The named file objects are covered. Assets and remaining retained-output families prevent an exhaustive inventory. External inference service/weights and static deployment stores are not substituted for accumulated memory.); Raw bodies and origin/digest parts; wiki page bodies/frontmatter; local schema; catalog and action log; task body/status; relay snapshots/returns; changelog: known — files [claimed], repo [afforded] (These named retained objects are filesystem files; this does not characterize accompanying image assets or every local instance history.); Retained lint policy declarations and evaluator benchmark artifacts: known — files [wired] (Inspected code reads schema policy and filesystem benchmark inputs; temporary Python values are not an additional retained-memory substrate.); Accompanying raw image assets: uninspected (Raw body record explicitly leaves image assets uncharacterized; their operative retained object/substrate inventory is unresolved.); Remaining shipped-family retained outputs and their later consumers: uninspected (Detailed output-consumer routes of remaining instruction-writing, portability, git, formatting and editorial families were not established by accepted records. Standalone summary delivery alone supplies no save/reload chain. This prevents exhaustive classification, not the positive local findings.)	partial coverage (Manual and automatic coexist at distinct admissions. Raw-maintenance conflict and initial/remaining-policy control prevent complete coverage.); Import/wrapup page selections and explicit human task edits/accepted findings: known — manual [claimed] (Manual is content approval, not merely starting a workflow.); Raw acquisition, direct hub/query filing, deterministic page maintenance and procedure capture: partial — automatic [claimed] (Model admits content without per-item human approval on these paths. Maintenance raw-body exception conflicts with general read-only constraint; operative precedence remains unresolved.); Schema initialization domain/tag decision versus generated conventions: partial — manual [claimed], automatic [wired] (Human supplies domain/starting tags; scaffold copying is code generation. Full convention/template admission inventory remains uninspected.); Readiness task repair, audit stamp, lint acceptance, navigation/log and changelog update: known — automatic [claimed] (Computational judgment admits these outputs; consequential intent changes remain separately human-owned.); Evaluation capture/benchmark generation: known — automatic [wired] (Caller selects experiment; software admits captured/generated artifacts.); Preexisting corpus initialization and remaining Lint-policy admission mechanisms: uninspected (Current files and fixed read consumers do not establish who approved original content or each exclusion/boilerplate declaration.); Remaining shipped-family retained outputs and their later consumers: uninspected (Detailed output-consumer routes of remaining instruction-writing, portability, git, formatting and editorial families were not established by accepted records. Standalone summary delivery alone supplies no save/reload chain. This prevents exhaustive classification, not the positive local findings.)	partial coverage (Inspected routes are pull; no bounded exhaustive absence of other directions is declared. Static installed guidance is excluded, not counted as accumulated-memory push.); Question/activated wiki orientation and page retrieval; named task lifecycle reads; procedure retrieval; pointer-admitted relay reads; incremental changelog read: known — pull [claimed] (These fulfill activated consumer work requests; subsequent tool delivery is not a second unsolicited push.); Schema policy reads by later linter and previous benchmark by later evaluation: known — pull [wired] (Fixed consumers load retained criteria/reference inputs on invocation.); Remaining shipped-family retained outputs and their later consumers: uninspected (Detailed output-consumer routes of remaining instruction-writing, portability, git, formatting and editorial families were not established by accepted records. Standalone summary delivery alone supplies no save/reload chain. This prevents exhaustive classification, not the positive local findings.)	uninspected coverage (No supported push operation is classified. Remaining-family direction coverage is unresolved, so the whole axis cannot be inapplicable or absent.); Completely classified requested wiki/task/relay/changelog and fixed lint/benchmark reads: inapplicable (These direction units are pull-only; push selector classification is inapplicable to them. Topic search and a named file do not establish push.); Remaining shipped-family retained outputs and their later consumers: uninspected (Detailed output-consumer routes of remaining instruction-writing, portability, git, formatting and editorial families were not established by accepted records. Standalone summary delivery alone supplies no save/reload chain. This prevents exhaustive classification, not the positive local findings.)	partial coverage (Positive automatic trace-fed routes coexist with manual extraction and unresolved consumers. No system-wide no, exercised improvement or model-weight learning is asserted.); Maintenance lint feedback to automatically retained finding declaration to later linter: known — yes [claimed] (Qualifying automatic trace-fed checking-policy write is prescribed; fixed downstream matching is wired but does not upgrade admission to wired.); Helper criticism/verification to automatically repaired task to later implementation: known — yes [claimed] (Task remains durable behavior-shaping contract, even though relay is cleared next run and improved capacity is unobserved.); Operational-pattern procedure and completed-work lesson writes: partial — yes [afforded] (Trace-fed use is afforded; generic direct authoring need not have trace origin, and lesson source mixtures are not fully characterized.); Selection-approved session wrapup: inapplicable (This admission mechanism requires explicit human content selection, so it is not the automatic write required by trace_learning; trace-extracted lineage remains positive.); Evaluation snapshots/benchmarks, Git changelog and audit baselines: not-determinable (Retention and comparison alone do not establish every qualifying durable behavior-shaping update. Benchmarks have no automatic later-worker update in the inspected path; there is no bounded absence record for all possible consumers.); Remaining shipped-family retained outputs and their later consumers: uninspected (Detailed output-consumer routes of remaining instruction-writing, portability, git, formatting and editorial families were not established by accepted records. Standalone summary delivery alone supplies no save/reload chain. This prevents exhaustive classification, not the positive local findings.)	partial coverage (Authority is consumer-specific. Learning at durable update consumers is not a benefit claim; no parametric adaptation or global enforcement/ranking authority is inferred.); Raw and wiki declarative text, catalog summaries and historical tasks at query/prior-art readers: known — knowledge [claimed] (Content informs answers, authoring or precedent use, without semantic acceptance or demonstrated activation.); Wiki procedures and schema prose at later operational/writing readers; task Goal/Approach at implementation: known — instruction [claimed] (Guidance authority is distinct from executable denial.); Schema enum/policy parts at lint parser and coverage/report consumer: known — validation [wired], enforcement [wired], routing [wired] (Acceptance bullets enforce info-only report partitioning; exclusions select checked trees; these do not certify source truth.); Audit baseline and task lifecycle/design flags at maintenance and finish consumers: known — routing [claimed] (Model consumers use durable state to choose work, rechecking or conditional architecture refresh.); Frozen task/intent and verifier returns at repair checks: known — validation [claimed], routing [claimed], learning [claimed] (Retained baseline criteria and licensed edits constrain later model checks/writes, with protocol rather than host enforcement.); Lint feedback at acceptance-declaration update consumer: known — learning [claimed] (Update-consumer learning force is separate from downstream symbolic suppression.); Catalog/page relevance and remaining selectors: partial — routing [claimed] (Index summaries influence model relevance selection; accepted records do not establish a relative-priority/ranking mechanism distinct from selection, so ranking is not asserted.); Previous benchmark at regression aggregation; changelog latest day at update consumer: known — knowledge [wired], routing [claimed] (These paths change report/processing decisions, not worker training.); Remaining shipped-family retained outputs and their later consumers: uninspected (Detailed output-consumer routes of remaining instruction-writing, portability, git, formatting and editorial families were not established by accepted records. Standalone summary delivery alone supplies no save/reload chain. This prevents exhaustive classification, not the positive local findings.)	AAS-2026-10-14-ai-modules-a7044d5627d2-01
Apache Maka	2	Retained local Markdown bodies/proposals and lifecycle metadata; long-term extracted facts and access/control structures; durable session/child history, effective projections, text and opaque provider checkpoints and tool-result reload; imported sessions; optional assistant index bodies/originals/coverage/progress, Matter snapshots/reviews, extraction files and daily archives; goal continuation and child patch artifacts. Included unresolved alternatives: remaining Goal/plan/Todo/WorkHub state, Feishu cache, through-use changed skill/workspace/plugin content, external-executor retention and client-specific later consumers. Excludes unchanged shipped static prompts/skills, ordinary unretained current-run state, serving weights not updated by these routes, and the whole runtime permission/storage surface except retained controls actually governing scoped memory.	partial coverage (Files and SQLite are established locally; checkpoint layouts and the remaining retained-state inventory prevent complete storage classification. Provider serving weights are not memory accumulated by the inspected routes.); Local entry bodies, pending proposals, bundle metadata and backups: known — files [wired] (The named Markdown documents and backups establish the local bundle substrate.); Long-term facts, keys, lifecycle, cursors and receipts: known — sqlite [wired] (The retained content and control parts share the inspected SQLite store.); Assistant index documents, original-reference/cache records and coverage notes: known — sqlite [wired] (Both text and access metadata persist in the assistant corpus.); Matter snapshots, materialized files and review journal: known — sqlite [wired], files [wired] (SQLite history and file objects are distinct retained parts of the same Matter route.); Independent history extraction output and receipt: known — files [wired] (The separate extraction result is not an index row.); Text checkpoint and encrypted provider-checkpoint persistence: not-determinable (A checkpoint writer is recorded, but physical storage is uninspected for both alternatives. Ciphertext is an encoding, not a service-object, model-weight or file substrate finding.); Runtime/child history and reload evidence, imported sessions, goal continuation, daily archives and patch artifacts: not-determinable (Durability and later readers are recorded, but these records do not establish each physical store layout. A runtime event address or artifact name is not a substrate.); Remaining included Goal/Todo/plan/WorkHub state, Feishu cache lifecycle, changed skills/workspace guidance, plugin state and client-specific retained-state routes: uninspected (The accepted inventory leaves physical storage unresolved for these paths. Goal continuation findings below do not resolve every goal or plan store; plugin composition does not expose arbitrary plugin state or external executor retention.)	partial coverage (Known computational transformations remain positive even though local manual/editor callers and other initial-content control are unresolved. No manual finding is licensed solely by user-labeled metadata.); Local bundle approve/manual-save and whole replacement mechanisms: not-determinable (Explicit review/edit surfaces and producer-written user labels do not establish a human decision at every caller. Manual is not asserted from API names or physical file I/O.); Fact proposal/canonicalization-to-commit: known — automatic [wired] (Starting remember/extraction does not approve each generated item.); Textual and provider compaction checkpoint admission: known — automatic [wired] (Structural/coverage checks admit generated state without per-summary operator approval.); Assistant index body edits and coverage/progress checkpoints: known — automatic [wired] (An initiating user criterion is not per-document approval.); Matter checkpoint and reviewed settlement: known — automatic [wired] (Intermediate checkpoint and final settlement gates are separate but both computational.); Independent extraction files; daily activity archive: known — automatic [wired] (Caller initiation does not change the generated-output admission mechanism.); Goal continuation durability and child patch publication: known — automatic [wired] (Automatic persistence is tied to selection/finalization, not arbitrary physical I/O.); External-session source selection and admission control: partial — automatic [wired] (Software conversion/publication is established, while the source-selection caller and per-message control are not.); Remaining included Goal/Todo/plan/WorkHub state, Feishu cache lifecycle, changed skills/workspace guidance, plugin state and client-specific retained-state routes: uninspected (The accepted inventory leaves per-content human versus computational admission unresolved for these paths. Goal continuation findings below do not resolve every goal or plan store; plugin composition does not expose arbitrary plugin state or external executor retention.)	partial coverage (Both pull and push routes exist, but opaque external executor retention, imported-history consumers and other included paths leave the direction inventory incomplete.); Local bundle foreground composition: known — push [wired] (The model does not request these stored entries; the composer supplies eligible text.); Requested fact/transcript/material and archived-tool-result retrieval: known — pull [wired] (Selection and delivery fulfill explicit consumer requests; delivery is not a second push.); Checkpoint replay into later provider requests: known — push [wired] (Text and opaque alternatives have different channels but automatic reuse is wired for both.); Index navigation/body reads and continuation guidance: known — pull [wired], push [wired] (Separate operations in one chain need separate directions.); Matter state and review replay: known — pull [wired], push [wired] (Wake paths are not body delivery; the documented workflow explicitly reads files.); Goal next-turn continuation: known — push [wired] (Persisted pending steering is selected by continuation machinery.); Daily-review archive query, child output and patch artifact listing: known — pull [wired] (These are recorded later requested reads, distinct from original immediate returns.); Independent extraction-file full-body reuse and imported-session downstream history use: partial — pull [afforded], pull [wired] (Files afford reading; imports expose a later catalog but exact history Recall/resume wiring is uninspected.); Remaining included Goal/Todo/plan/WorkHub state, Feishu cache lifecycle, changed skills/workspace guidance, plugin state and client-specific retained-state routes: uninspected (The accepted inventory leaves later request/selection/delivery operations unresolved for these paths. Goal continuation findings below do not resolve every goal or plan store; plugin composition does not expose arbitrary plugin state or external executor retention.)	partial coverage (Identifier and coarse selection are established on actual pushes. Recall lexical matching is not relabeled inferred-lexical push; remaining branches prevent a system-wide pull-only or complete signal assessment.); Local bundle push selection: known — coarse [wired], identifier [wired] (Trigger is foreground composition; selector receives policy, lifecycle, session ID and a character cap, selecting eligible stored entry bodies.); Checkpoint automatic reuse selection: known — coarse [wired], identifier [wired] (Trigger is a later provider send; stored coverage/source identity and model compatibility select reusable checkpoint state, with context-budget pruning.); Index maintenance continuation task: known — identifier [wired], coarse [wired] (Trigger is continued maintenance; exact index/range identity selects retained criterion/progress notes for that worker.); Matter saved-review reuse: known — identifier [wired] (Trigger is settlement replay/currentness handling; submission/fingerprint/revision identity selects the retained verdict, not every future narrative.); Goal pending continuation admission: known — identifier [wired], coarse [wired] (Trigger is later goal-turn admission; lane/checkpoint/control identity and current active/wait state select stored steering.); Inspected requested Recall, archive Read, index body reads and daily/child queries: inapplicable (These particular operations are pull: literal/key/address selection satisfies the consumer request. Their lexical ranking is not a push signal.); Independent extraction-file reuse and imported-history push alternatives: uninspected (No complete later full-body consumption inventory establishes pull-only use, so these alternatives cannot be declared inapplicable.); Remaining included Goal/Todo/plan/WorkHub state, Feishu cache lifecycle, changed skills/workspace guidance, plugin state and client-specific retained-state routes: uninspected (The accepted inventory leaves automatic push selectors unresolved for these paths. Goal continuation findings below do not resolve every goal or plan store; plugin composition does not expose arbitrary plugin state or external executor retention.)	partial coverage (Several wired yes routes establish architectural trace learning, not improved capacity or exercised behavior. No local or global no is asserted without bounded absence evidence; unresolved opaque and alternate writes remain included.); Runtime fact extraction to later Recall: known — yes [wired] (Trace-fed model/software writes create durable behavior-shaping fact context and an implemented later consumer.); Text history compaction to later provider invocation: known — yes [wired] (A generated continuation summary qualifies even within the same run; benefit and new knowledge are unnecessary.); Provider checkpoint alternative: not-determinable (Automatic durable history-fed production and replay are wired, but hidden payload semantics do not establish whether it is a qualifying behavior-shaping artifact rather than opaque transport alone.); Index maintenance trace-derived documents and progress: partial — yes [wired] (History-fed document writes and later readers qualify; arbitrary remote sources and other authoring paths remain unresolved.); Matter revised task state and review loop: known — yes [wired] (Execution/inbox experience feeds computationally admitted snapshots; later activations read state. Review-journal future narrative use remains narrower than state use.); Goal reason-derived durable steering: known — yes [wired] (The inspected continuation path, not all Goal persistence, is classified.); Independent extraction files, daily-review archives and patch results: not-determinable (Automatic trace/history generation and retention alone do not establish a durable behavior-shaping later consumer for every product: full extraction adoption, archive-to-agent revision and parent patch application are uninspected.); Local authored bundle and imported history admission alternatives: not-determinable (The records do not establish complete trace-fed proposal origins and later imported-history use. Authored/manual labels and raw-log import neither prove qualifying learning nor warrant no.); Remaining included Goal/Todo/plan/WorkHub state, Feishu cache lifecycle, changed skills/workspace guidance, plugin state and client-specific retained-state routes: uninspected (The accepted inventory leaves qualifying automatic trace-fed writes and later consumers unresolved for these paths. Goal continuation findings below do not resolve every goal or plan store; plugin composition does not expose arbitrary plugin state or external executor retention.)	partial coverage (All authority values describe named retained parts at actual consumers. Update-input learning and output knowledge/instruction are separate. No source wiring proves model obedience, benefit or universal retained-state authority.); Local active entry text at foreground model: known — knowledge [wired] (The composer labels retained entries untrusted context, not standing permission or higher-priority direction.); Bundle lifecycle/scope/revision metadata at projection and mutation code: known — enforcement [wired], validation [wired], routing [wired] (Eligibility and revision controls differ from the advisory body consumer.); Facts and past transcripts at Recall consumer: known — knowledge [wired] (Recall returns selected assertions/passages, not stored canonicalization criticism.); Fact keys/lifecycle and extraction cursors at lookup/admission consumers: known — ranking [wired], enforcement [wired], routing [wired], validation [wired] (The actual consumers use keys for matching, lifecycle for reachability and cursors for update windows.); Retained history at fact and textual-compaction update consumers: known — learning [wired] (Learning authority describes the update input, not a claim of measured improvement.); Text checkpoint at later model and replay validator: known — knowledge [wired], validation [wired], routing [wired] (The text and its coverage wrapper have different effects.); Opaque provider checkpoint at provider-boundary consumer: partial — routing [wired] (Compatibility routing is visible, but the encrypted consumed payload effect is unresolved; a knowledge/instruction/learning classification of that payload is not inferred from display text.); Index bodies, criterion/progress notes and coverage metadata at their distinct consumers: known — knowledge [wired], instruction [wired], learning [wired], routing [wired], validation [wired], enforcement [wired] (Body reads, continued maintenance guidance and structural checkpoint decisions remain separate.); Matter state and persisted review/currentness fields: known — knowledge [wired], learning [wired], validation [wired], enforcement [wired], routing [wired] (Later state-file reads, current correction feedback and software replay checks have different horizons.); Goal continuation prompt and durable checkpoint/lease state: known — instruction [wired], routing [wired], enforcement [wired] (The retained reason and symbolic state have distinct consumers.); Daily-review archive at later client query: known — knowledge [wired] (This establishes a reference product for a client, not automatic method revision.); Independent extraction files and patch artifacts at future adopter: not-determinable (A retained file/path or artifact reference does not establish a downstream full-body consumer or parent adoption; previews/current returns alone do not settle later authority.); External imported histories and changed workspace/skill/plugin content: partial — instruction [afforded] (Imported role filtering and actual changed-content adoption remain unresolved; no universal instruction authority is inferred.); Remaining included Goal/Todo/plan/WorkHub state, Feishu cache lifecycle, changed skills/workspace guidance, plugin state and client-specific retained-state routes: uninspected (The accepted inventory leaves consumer effects and later behavioral force unresolved for these paths. Goal continuation findings below do not resolve every goal or plan store; plugin composition does not expose arbitrary plugin state or external executor retention.)	AAS-2026-10-15-maka-4a8b73e41579-01
archie	2	The use-changed modelRoles configuration value, including its opaque retained backend and initial provenance/admission, the helper's default replacement, and subsequent helper read/modify/write consumption. External model-role interpretation is an unresolved downstream contract. Excludes static shipped shell definitions and launch flags as memory, the planned assistant as an operative memory system, unrelated desktop/contributor persistence, and external agent sessions, provider weights and internal maintenance routes.	not-determinable coverage (The records inventory one retained configuration object but identify no backend; files, databases and other substrate alternatives remain unresolved.); Retained modelRoles value behind external omp config get/set: not-determinable (The retained backend is opaque. JSON is encoding, and the transient shell variable is transport rather than the retained substrate. No controlled retained-substrate value follows.)	partial coverage (Automatic subsequent transformation is supported; initial and external admission control remain unresolved.); Helper replacement of the default and admission request to omp config set: known — automatic [wired] (The complete helper establishes this subsequent write mechanism, not external setter validation or successful retention.); Initial configuration admission and external content replacement: not-determinable (The getter's existing content has no recorded initialization control or external editing mechanism. Generic caller identity cannot establish manual agency.)	partial coverage (The helper supplies a pull witness, but the included intended downstream default-consumption contract prevents complete direction coverage.); Subsequent profile invocation requesting the whole retained modelRoles through config get: known — pull [wired] (The repeated getter/transform/setter chain is wired, conditional on external persistence; no execution establishes that it operated.); Intended later runtime consumption of the retained default, including direct omp invocation: not-determinable (The retained-default consumer contract is uninspected. Launch flags and a callable bypass do not resolve request, selection or delivery operations there.)	not-determinable coverage (A known local pull does not make the whole axis inapplicable while downstream delivery remains unresolved; no positive signal can be assigned.); Known helper pull operation: inapplicable (This complete local chain requests the whole value; its key identifies requested configuration, not a push selection match.); Unresolved downstream runtime default-consumption contract: not-determinable (No push operation, trigger or selector is established for this contract, but its direction is unresolved rather than proven pull-only.)	not-determinable coverage (No qualifying trace-fed write is established. No canonical bounded-absence record supports no, and initial provenance is unresolved. External sessions and provider training are outside this memory scope.); Automatic helper default replacement and later helper consumption: not-determinable (The route records configuration and a fixed literal as inputs, not a qualifying trace-fed update. No canonical evidenced-absence record establishes the bounded negative required for a controlled no finding. Automatic retention alone establishes neither yes nor a formally warranted no.); Initial or externally replaced retained configuration content: not-determinable (Initial production, admission and any trace inputs are unknown. The helper's local negative cannot establish how that earlier content was produced.)	not-determinable coverage (The records establish operational configuration requests, not a controlled authority value at an established retained-content consumer. Unknown external interpretation is not an evidenced absence of authority.); Existing modelRoles fields at the later helper's read/modify/write consumer: not-determinable (The helper replaces the prior default and mechanically carries other fields forward. No retained field is shown supplying knowledge, directions, a ranking input, a selection policy, an admissibility criterion or a veto at that consumer. Mechanical preservation alone establishes none of the controlled forces.); Retained default at the intended external runtime model-role consumer: not-determinable (Configuration authority is intended, but the actual retained-default consumer and precedence are uninspected. Explicit current-launch flags cannot establish routing by the retained value.)	AAS-2026-10-15-archie-1ab3d8b8ff03-01
AriGraph	2	Includes attempt/task-local facts, episodes, numerical access metadata, history, feedback plans, summary/full-history/smart-RAG baselines, snapshots and logger outputs; Hypergraph thesis/event/index and incompletely recorded inherited spatial state; auxiliary ObservedGraph and caller JSON alternatives; unresolved provider-internal experience retention at the adapter boundary. Excludes static shipped instructions, fixed model weights as such, current-only temporary action/answer proposals and diagnostic-only evaluation accumulators, external paper/site and bytecode-only variants. No restart persistence or external deployment is presumed.	partial coverage (Recorded local and file stores are classified. The undeclared inherited Hypergraph store and opaque provider retention prevent a complete inventory; numerical encoding is not a vector-database substrate.); Game/QA facts, episodes, history, baseline stores and feedback plans: known — in-memory [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Graph dense caches and episode vectors: known — in-memory [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Game episode snapshots and score histories: known — in-memory [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Shipped append-only text logs: known — files [wired] (The text logging branch is wired in shipped drivers; later reload, activation and benefit are not inferred.); Unused JSON serialization capability: known — files [afforded] (File storage is afforded by the inspected method, not wired by a shipped caller. Payload form, provenance, admission control and later consumption remain unresolved.); Hypergraph thesis/entity/event objects and incidence/vector fields: known — in-memory [afforded] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Auxiliary ObservedGraph retained network graph: known — graph [afforded] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Hypergraph inherited movement-triplet state beyond its recorded insertion: not-determinable (The runtime records reciprocal movement additions but leaves the relation to thesis/event state and later consumers unassessed. The missing separate part record prevents complete storage inventory classification; it is not silently excluded.); Unresolved provider-internal experience retention behind the generation adapters: uninspected (The adapters expose requests, not provider experiential state or its storage substrate. Static model weights are excluded; this uncertainty does not assert that a provider memory write exists.)	partial coverage (Computational admission is positive on inspected shipped routes and Hypergraph internals. Unknown generic caller control is not manual or automatic by default; authored seed provenance does not change automatic run admission.); Game/QA admission, recent/full/smart-RAG appends, summary and plan replacement: known — automatic [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Graph vectors and episode snapshot writes: known — automatic [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Shipped text logger admission: known — automatic [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Hypergraph content/vector/link admissions: known — automatic [afforded] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); ObservedGraph caller-supplied initialization/replacement and mutation: not-determinable (Generic caller identity does not distinguish a human decision supplying/replacing retained content from computational admission.); Unused arbitrary JSON logger payload admission: not-determinable (File replacement capability is inspected, but caller control of arbitrary content is unresolved and no shipped call exists.); Hypergraph inherited movement-triplet state beyond its recorded insertion: not-determinable (The runtime records reciprocal movement additions but leaves the relation to thesis/event state and later consumers unassessed. The missing separate part record prevents complete separate spatial-store admission coverage classification; it is not silently excluded.); Unresolved provider-internal experience retention behind the generation adapters: uninspected (The adapters expose requests, not provider experiential state or its write control. Static model weights are excluded; this uncertainty does not assert that a provider memory write exists.)	partial coverage (Pull and push occur on different chains. Requested graph delivery is not counted again as push. Opaque and undeployed alternatives prevent complete direction coverage.); Requested game/QA fact and episode retrieval: known — pull [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Main graph navigation/exploration reads: known — pull [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Driver-supplied recent/full history, feedback plan and summary: known — push [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Smart-RAG context construction: known — push [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Hypergraph requested read API: partial — pull [afforded] (API return is afforded; downstream behavioral delivery and deployment are not established.); Auxiliary state access, snapshots and log/JSON later consumption: uninspected (Display/access capabilities and no shipped snapshot replay do not identify external later requests or unsolicited delivery.); Hypergraph inherited movement-triplet state beyond its recorded insertion: not-determinable (The runtime records reciprocal movement additions but leaves the relation to thesis/event state and later consumers unassessed. The missing separate part record prevents complete spatial read-back classification; it is not silently excluded.); Unresolved provider-internal experience retention behind the generation adapters: uninspected (The adapters expose requests, not provider experiential state or its read-back direction. Static model weights are excluded; this uncertainty does not assert that a provider memory write exists.)	partial coverage (Known push routes use coarse windows and smart-RAG embedding/judgment inputs. No identifier push is inferred from dictionary/hash IDs. Pull-only local inapplicability cannot settle unknown alternatives.); History, summary and plan context pushes: known — coarse [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Smart-RAG ranked text/action push: known — inferred-embedding [wired], inferred-judgment [wired], coarse [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Completely inspected requested game/QA/navigation branches: inapplicable (These branches fulfill explicit caller requests rather than push. Embedding selection within a pull request is not a push signal.); Hypergraph API and possible downstream delivery: uninspected (The requested return is known, but downstream use is not assessed; no inferred push signal is assigned to event ranking.); Auxiliary/log/snapshot possible pushes: uninspected (No recorded external push operation supplies a trigger, selector and selected retained part.); Hypergraph inherited movement-triplet state beyond its recorded insertion: not-determinable (The runtime records reciprocal movement additions but leaves the relation to thesis/event state and later consumers unassessed. The missing separate part record prevents complete spatial push selection classification; it is not silently excluded.); Unresolved provider-internal experience retention behind the generation adapters: uninspected (The adapters expose requests, not provider experiential state or its push selector. Static model weights are excluded; this uncertainty does not assert that a provider memory write exists.)	partial coverage (The derived system union includes yes on supported routes. QA/raw-history routes lack canonical bounded-absence records for local no and remain unresolved rather than producing a system-wide negative. Trace learning here is retained behavior-shaping transformation, not demonstrated capacity improvement.); Game observation-to-fact/episode update with later planner/action consumption: known — yes [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Summary and feedback-plan experience-fed replacement: known — yes [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Smart-RAG experience-fed importance and vectors: known — yes [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); QA static paragraph admission: not-determinable (The recorded route extracts static paragraphs and terminates answer metrics diagnostically. No canonical bounded-absence record establishes local no for qualifying trace-fed writes, so a negative classification is withheld.); Full-history raw trace append alone: not-determinable (The recorded mechanism retains raw observation/action text rather than a derived artifact. Raw context delivery does not establish yes, and no canonical bounded-absence record warrants local no.); Hypergraph thesis/event API: not-determinable (Automatic derivation is available, but original caller trace provenance and qualifying later behavioral consumption are not established.); Auxiliary observed graph and arbitrary JSON content: not-determinable (Caller control/provenance and later behavioral consumption are unresolved; mutation or file I/O alone does not establish the qualifying chain.); Snapshot and text log external learning consumers: uninspected (These retain traces/copies without shipped replay. External trace-fed learning writes using them have not been inspected.); Hypergraph inherited movement-triplet state beyond its recorded insertion: not-determinable (The runtime records reciprocal movement additions but leaves the relation to thesis/event state and later consumers unassessed. The missing separate part record prevents complete spatial trace-to-later-consumer chain classification; it is not silently excluded.); Unresolved provider-internal experience retention behind the generation adapters: uninspected (The adapters expose requests, not provider experiential state or its trace-fed updates. Static model weights are excluded; this uncertainty does not assert that a provider memory write exists.)	partial coverage (Knowledge, learning, instruction, ranking, routing and a bounded exact-match enforcement path have local warrants. Static prompt rules and current-run parser criteria are excluded; they do not classify retained memory as validation. Unresolved auxiliary consumers prevent complete authority coverage.); Selected facts/episodes at game planner/action and QA answer models: known — knowledge [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Retained game facts at extraction/replacement update models: known — learning [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Retained fact/episode text and episode vectors at relevance selectors: known — ranking [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Spatial triples at path/exploration software: known — routing [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Recent history and full/smart-RAG text at planner/action models: known — knowledge [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Recent observation exclusion list at episodic filter: known — enforcement [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Summary at planner/action models and next summarizer: known — knowledge [wired], learning [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Feedback plan at later model requests: known — instruction [wired], ranking [wired], learning [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Smart-RAG numeric metadata at software ranking: known — ranking [wired] (The cited records identify the retained part, implemented path and consumers within this unit; operation and benefit are not inferred.); Hypergraph names/incidence/event vectors at API retrieval: partial — ranking [afforded] (API selection force is inspected; external planner consumption and its authority are uninspected.); Snapshots/logs at possible later consumers: uninspected (No shipped snapshot replay or log-reload behavioral consumer is established. External consumption and arbitrary JSON consumers are not inspected; mere retention grants no authority.); Auxiliary observed graph at later behavioral consumers: uninspected (Mutation/display is available, but no shipped later agent or model consumer establishes an authority path.); Hypergraph inherited movement-triplet state beyond its recorded insertion: not-determinable (The runtime records reciprocal movement additions but leaves the relation to thesis/event state and later consumers unassessed. The missing separate part record prevents complete consumer/authority classification; it is not silently excluded.); Unresolved provider-internal experience retention behind the generation adapters: uninspected (The adapters expose requests, not provider experiential state or its behavioral authority. Static model weights are excluded; this uncertainty does not assert that a provider memory write exists.)	AAS-2026-10-12-arigraph-fd0f6f1c1383-01
Ars Contexta	2	Accumulated or revised domain notes and source archives; topic maps and inline navigation; queue/task findings; derivation/configuration and generated or revised live context, skills and templates; methodology/observations; session lifecycle and substantive handoff alternatives; task priorities, recommendation and diagnostic history; optional numerical search/index/cache parts and unresolved progress/history alternatives. Includes registered plugin and generated-template write/read-back branches separately. Excludes unchanged bundled methodology/presets/templates as static guidance, external host/model weight state, and transient warnings/answers without retained writers. External search implementation remains uninspected rather than inferred from readable hits.	partial coverage (The declared file and Git surfaces are classified. Optional numerical memory and unresolved history/progress alternatives prevent a complete substrate inventory. External model weights are outside accumulated target memory, not a positive model-weights finding.); Vault content, navigation, operational evidence, checkpoint files, queue and generated live instructions: known — files [claimed] (All these declared objects are filesystem files; Markdown, YAML and JSON are encodings, not distinct substrates.); Optional Git persistence of accumulated vault and instruction files: known — repo [wired] (The registered commit callback establishes optional Git persistence. Copied skill versions and manual rollback establish file archival, not a separate repository mechanism.); Optional search index, embeddings and persisted search caches: uninspected (The records identify numerical/index alternatives but do not inspect their storage. A readable search hit cannot establish vector, sqlite or service-object storage.); Stats-history snapshot and alternate builder/progress initialization: uninspected (The YAML history writer and opaque tutorial-progress initialization are unresolved. Encoding and a named history file do not establish the full operative storage inventory.)	partial coverage (Both human-controlled and automatic admission exist. The approval/dontAsk conflict, initial content, summary alternatives and opaque writers prevent a universal write-agency classification.); Manual notes/captures, explicit remember guidance and user task-stack edits: known — manual [claimed] (Human content supply or direct editing orders are admission decisions, even when software performs I/O.); Contextual remember, interactive reweave and approved architecture/refactor/upgrade replacements: known — manual [claimed] (Each mechanism has its own explicit confirmation or replacement approval.); Reduce extracted-note admission under prescribed approval policy: partial — manual [claimed] (Reduce requires approval, including handoff; dontAsk and initial extraction eligibility leave the wrapper path unresolved.); Learn filing, reflect links, handoff reweave, session mining and next condition/history writes: known — automatic [claimed] (Starting a command is not approval of each admitted finding, link or recommendation.); Lifecycle tracking, marker migration and broad post-Write Git persistence: known — automatic [wired] (Automatic physical persistence is separate from the original content admission mechanisms.); Initial goals/identity, working-summary admission, stats-history and alternate progress writer: not-determinable (Instructions support human/model checkpoint alternatives but do not settle every admission control; installed initialization and opaque writers are unsupplied.); Optional numerical index/cache writes: uninspected (No recorded provider write/admission mechanism establishes human versus software control.)	partial coverage (Requested reads and unsolicited hook/host/lead supplies coexist. Generated and registered deployment retain separate evidence bases; unresolved alternatives prevent complete direction coverage.); Requested note/task/source/local-methodology reads, derivation rereads and advisory work-selection inputs: known — pull [claimed] (The consumer asks tools for retained material; returning the requested file is not an additional push.); Registered SessionStart unsolicited orientation: known — push [wired] (The callback selects existing goals/self bodies, recent headers and lifecycle JSON independently of a retained-content request.); Generated SessionStart unsolicited orientation: known — push [afforded] (This is the distinct generated-template route, not a proven installation.); Host-loaded accumulated context and lead-selected sibling content in worker prompts: known — push [claimed] (The lead selects sibling titles for workers before their own requested task/body reads.); Alternate copied/inherited delivery, optional search payload and unresolved summary/history/progress consumers: uninspected (Known pulls/pushes do not settle installed loading, opaque payload delivery, batch-summary consumption or all alternate progress/history routes.)	partial coverage (Push signals are coarse availability/recency/state gates and actual batch-identifier matching. Lexical/semantic requested retrieval is pull, not proof of an inferred push signal; the axis is not pull-only inapplicable.); Plugin SessionStart file-existence, preferred-goals and recent-header selection: known — coarse [wired] (Trigger is SessionStart; selector inputs are availability, configured flags and mtime; retained parts are full goals/self, current JSON and five header excerpts.); Generated SessionStart file-existence and recent-title selection: known — coarse [afforded] (Trigger is SessionStart; selector inputs are file presence, self-versus-ops preference and mtime; selected parts are full self/goals, current state and five titles.); Lead supplies already-created sibling titles to reflect/reweave and parallel workers: known — identifier [claimed], coarse [claimed] (Trigger is worker-prompt construction; selector inputs include batch identity and completed create status; selected retained parts are sibling titles/task paths, not all bodies.); Automatic project context loading: known — coarse [claimed] (Later session trigger supplies the live project context; requested skill loading is not used as an identifier-push witness.); Unresolved alternate loading/progress/history and opaque search supply: uninspected (Recorded requested search routes do not warrant inferred-embedding or inferred-lexical push. Missing selector/payload facts prevent classifying unresolved alternatives.)	partial coverage (Qualifying procedural yes routes coexist with unresolved automatic summary/history/alternate updates. Trace learning here needs neither accepted new knowledge nor demonstrated improvement. JSON lifecycle retention alone is not a substantive learning witness.); Remember session mining without per-finding human admission, followed by directive/observation consumers: known — yes [claimed] (This is a qualifying automatic trace-fed durable-guidance route at claimed basis, not a closed transcript-capture deployment.); Next automatic recommendation-history update and later anti-repetition consumer: known — yes [claimed] (Generated recommendation history shapes a later invocation without per-entry human approval.); Automatic worker phase findings and later task-file consumers: known — yes [claimed] (The phase-worker chain is conditional on eligible work and model compliance; extraction approval remains a separate admission.); Session-summary/handoff alternatives and missing automatic substantive transcript producer: not-determinable (Substantive summaries could qualify if automatic, retained and later consumed, but the recorded human/model alternatives and JSON-only hooks do not resolve every writer.); Other diagnostic-history, alternate generation/progress and numerical updates: uninspected (The reports do not fully trace every automatic activity-fed write and later behavior-shaping consumer. No system-wide no or bounded local no is inferred.)	partial coverage (The actual readers distinguish knowledge, instruction, routing, ranking, validation, configured vetoes and learning-input authority. Opaque search/history/progress alternatives remain unresolved; post-Write warnings do not establish content enforcement.); Notes, topic prose, sources, task findings and health/operational evidence at later knowledge-work and answering consumers: known — knowledge [claimed] (Knowledge authority is reference/context use, not acceptance as true.); Live generated context/skills and full accumulated methodology at later guidance consumers: known — instruction [claimed], instruction [afforded], instruction [wired] (Full bodies are distinct from recent title/header displays; host loading and model compliance are external.); Recent operational-methodology headers/titles and lifecycle state at session display consumer: known — knowledge [wired], knowledge [afforded] (A display summary has only the force of the selected part, not the omitted directive.); Curated topic links, queue phases, derivation configuration, mined/status markers and user priority targets at selection consumers: known — routing [claimed] (Routing names operation/input selection; it does not assert a compiled scheduler or universal withdrawal.); Configuration flags at automatic session tracking and Git callback gates: known — enforcement [wired], routing [wired] (Only configured hook operations are covered; content writes are not universally blocked.); Retained map framing, user priorities and recommendation history at relative-priority consumers: known — ranking [claimed] (This is priority influence, not an inspected numerical ranking implementation.); Retained configuration-dependent templates at note quality checking: known — validation [claimed] (The unit covers regenerated domain criteria, not the bundled static kernel or transient warning output.); Session evidence, observations and recommendation history at durable guidance/update consumers: known — learning [claimed] (Learning authority is the update-consumer effect; it is independent of downstream instruction and measured improvement.); Numerical ranking inputs and unresolved history/progress consumers: uninspected (Search ranking implementation, stats-history production and alternate progress consumption remain uninspected; no force is assigned to unused metadata.)	AAS-2026-10-12-arscontexta-f9e5f2f34166-01
ATLAS	2	Included: use-created or changed Markdown knowledge and its path index, operational database payloads, extraction logs, connector/module checkpoints, browser conversation text/display metadata, and changed configuration insofar as it routes their acquisition or consumption. Provider-internal retention/context maintenance and unavailable external-module memory remain opaque included alternatives. Excluded: static shipped prompts, schema descriptions, unchanged setup instructions, external MCP-host memory, transient current-request tool blocks except as inputs to retained chat findings, nonpersisted derived query/diagnostic results, and model weights not established as accumulated ATLAS memory.	partial coverage (The files, SQLite variants, browser array and changed configuration cover the recorded local stores. Opaque provider and external-module alternatives prevent complete storage enumeration.); Markdown knowledge, attribution and contradiction parts; maintained path index: known — files [wired] (Filesystem wiring resolves this store, not deployed mounts or survival.); Operational payloads, extraction-log rows and connector/module checkpoints on shipped entry points: known — sqlite [wired], in-memory [wired] (Both source-wired storage configurations are retained; the available PostgreSQL method is not a shipped bootstrap path.); Browser conversational text and display metadata: known — in-memory [wired] (No reload or cross-device persistence is established.); Changed memory-operation configuration reloaded by later startup or current consumers: known — files [wired] (Only changed configuration affecting retained-data acquisition and consumption is included, not every shipped setting.); Provider-internal accumulated state and context maintenance: not-determinable (Provider retention, caching and compaction are opaque; storage substrate cannot be classified from the non-training client interface.); Additional retained state of unavailable external modules: not-determinable (ModuleContext permits additional state and native effects, but external implementations are unavailable; storage substrate remains unresolved.)	partial coverage (Automatic model/software routes and explicit browser human decisions are distinct findings. Generic callers and initial/opaque admissions remain unresolved rather than inheriting another route’s agency.); Engine and scheduled-metadata knowledge admission: known — automatic [wired] (Starting enrichment or supplying text does not make returned updates manually admitted.); Chat save_knowledge admission: known — automatic [wired] (No human content approver intervenes.); Browser editor save/delete admission: known — manual [wired] (This warranted human decision is confined to the browser route.); Generic direct KB API content admission: not-determinable (The same endpoint accepts generic callers; no record resolves a human decision versus software admission on this alternative.); Mapped/imported and model-extracted record admission: known — automatic [wired] (An explicit seed load or user upload initiates automatic subsequent admission.); Pipeline-log and built-in checkpoint admission: known — automatic [wired] (Checkpoint use is not substituted for its evidenced generated-write mechanism.); Returned browser assistant/reply-error retention: known — automatic [wired] (Software retention differs from explicit editor adoption.); Browser user-turn retention: known — manual [wired] (User control over this content does not extend to returned replies or tool saves.); Administrative record pruning: known — automatic [wired] (Human initiation of cleanup is not manual per-row admission/withdrawal.); Initial retained content and generic changed-configuration admission: not-determinable (Initial deployment control and generic configuration caller identity are unavailable; manual/automatic cannot be assigned.); Provider-internal accumulated state and context maintenance: not-determinable (Provider retention, caching and compaction are opaque; write/admission agency cannot be classified from the non-training client interface.); Additional retained state of unavailable external modules: not-determinable (ModuleContext permits additional state and native effects, but external implementations are unavailable; write/admission agency remains unresolved.)	partial coverage (Recorded local routes establish both push and pull. Unknown provider and plugin alternatives remain included and prevent complete direction coverage.); Engine relevant-file selection and provider delivery: known — push [wired] (Explicit upstream workflow initiation does not make this selection a requested file return.); Chat automatic knowledge path-index delivery: known — push [wired] (The index is selected/coarse availability metadata, not all file contents.); Chat requested full-file/tree/data reads and editor file reads: known — pull [wired] (Returning a requested file is not another push just because code delivers it.); Browser automatic history projection into a later chat invocation: known — push [wired] (The current new user message is not itself read-back; earlier tool-result payloads are omitted.); Scheduled automatic extraction-log selection and input delivery: known — push [wired] (Cursor selection is an upstream operation distinct from engine selection of knowledge files.); Built-in acquisition accessors and next-run checkpoint reads: known — pull [wired] (These storage requests return control data to actual later software consumers.); Later current/startup configuration reload: known — pull [wired] (Reload does not automatically refresh every already-constructed module reference.); Provider-internal accumulated state and context maintenance: not-determinable (Provider retention, caching and compaction are opaque; read-back operations cannot be classified from the non-training client interface.); Additional retained state of unavailable external modules: not-determinable (ModuleContext permits additional state and native effects, but external implementations are unavailable; read-back operations remains unresolved.)	partial coverage (Coarse availability/window selection and inferred lexical targeting are wired. Pull-specific signals are inapplicable locally, but unresolved push alternatives prevent axis inapplicability or complete coverage.); Enrichment push of selected knowledge contents: known — inferred-lexical [wired], coarse [wired] (Lexical targeting and coarse budget filtering are separate aspects of the same push. No actual identity equality match or embedding selector is established.); Chat automatic retained-path inventory push: known — coarse [wired] (Literal identifiers in catalog entries do not establish identifier-based push selection.); Browser prior user/assistant text push: known — coarse [wired] (Role projection excludes previous tool logs; no explicit history-size budget is recorded.); Scheduled metadata push to enrichment: known — coarse [wired] (Eligibility/time/window filters do not become identifier or inferred-judgment signals. Downstream knowledge selection is classified separately.); Requested knowledge/tree/database/checkpoint/configuration read returns: inapplicable (These operations fulfill explicit consumer requests; push signal does not classify a requested literal-path or ID read.); Provider-internal accumulated state and context maintenance: not-determinable (Provider retention, caching and compaction are opaque; push selection signal cannot be classified from the non-training client interface.); Additional retained state of unavailable external modules: not-determinable (ModuleContext permits additional state and native effects, but external implementations are unavailable; push selection signal remains unresolved.)	partial coverage (Qualifying scheduled metadata, derived progress and trace-fed chat-save branches establish positive existence. Raw replay supplies no qualifying positive or canonical bounded-absence warrant. Unknown origins/provider/plugins prevent a complete yes/no inventory; positive existence is not demonstrated learning benefit.); Scheduled extraction-log-to-Markdown update and later file consumer: known — yes [wired] (This qualifying trace-fed artifact route requires neither new facts nor observed accuracy/improvement; actual execution remains unobserved.); Built-in trace-derived progress checkpoints at later acquisition consumers: known — yes [wired] (Derived control artifacts shape later behavior. This is distinct from retaining raw logs and from learned parameters or improved acquisition capacity.); Chat-generated saved findings from submitted history/current tool feedback: partial — yes [wired] (The trace-fed save branch is wired. A save may instead contain other supplied content; actual proposal provenance and content are unavailable, so this finding does not classify every chat save as trace learning.); Raw browser conversation retention and replay: not-determinable (The recorded mechanism retains and replays raw text rather than establishing a qualifying derived artifact. No canonical bounded-absence record warrants a resolved local no; the separate chat-save branch and opaque provider compaction cannot be excluded by this raw-replay account.); Entity/raw-text enrichment and operational document extraction with unspecified original input provenance: not-determinable (The original documents/text are not established as activity traces for every route. Durable outputs are known, but qualifying trace-fed origin and any mixed inputs cannot be completely classified.); Generic API/editor/configuration content histories and initial content: not-determinable (Manual browser adoption is known locally; generic caller provenance and any trace-derived automatic alternative are unavailable.); Provider-internal accumulated state and context maintenance: not-determinable (Provider retention, caching and compaction are opaque; qualifying automatic trace-fed write and later consumer cannot be classified from the non-training client interface.); Additional retained state of unavailable external modules: not-determinable (ModuleContext permits additional state and native effects, but external implementations are unavailable; qualifying automatic trace-fed write and later consumer remains unresolved.)	partial coverage (Knowledge, ranking, learning, routing and enforced skips are independently wired at named consumers. Static prompt/schema rules are excluded, so their instruction/validation force is not attributed to accumulated memory. Unknown provider/plugin consumers prevent a complete authority inventory.); Retained knowledge prose and conflict warnings delivered to chat, enrichment and editor: known — knowledge [wired] (No trusted-source parser promotes arbitrary file prose into standing instruction or enforcing invalidation.); Retained knowledge text and paths at lexical enrichment selection: known — ranking [wired] (Retained features influence ranking despite the fixed, unlearned scoring algorithm.); Retained knowledge context at automatic file-update generation: known — learning [wired] (Authority at this update consumer is separate from knowledge authority at subsequent readers; no improved capacity is claimed.); Operational payloads at query/model consumers: known — knowledge [wired], ranking [wired] (Fixed result rules do not endorse source truth or establish an external host action.); Extraction-log hash/error/timestamp at processing and scheduled-enrichment selectors: known — routing [wired], enforcement [wired] (The enforced skip is not a factuality check or content acceptance criterion.); Extraction-log metadata at durable knowledge-update consumer: known — learning [wired] (This learning authority names the update path, not demonstrated improvement.); Connector/module checkpoints at next acquisition and maintenance invocation: known — routing [wired], enforcement [wired] (Markers influence fixed control flow, not model compliance; cursor failure semantics remain consequential.); Replayed browser history at subsequent chat model: known — knowledge [wired] (History is not a verified evidence ledger or automatically revised standing guidance.); Changed acquisition/model configuration at current and later retained-data operations: known — routing [wired] (Old module references can survive reload, preventing universal immediate routing propagation.); Provider-internal accumulated state and context maintenance: not-determinable (Provider retention, caching and compaction are opaque; actual retained-part consumer and force cannot be classified from the non-training client interface.); Additional retained state of unavailable external modules: not-determinable (ModuleContext permits additional state and native effects, but external implementations are unavailable; actual retained-part consumer and force remains unresolved.)	AAS-2026-10-12-atlas-b28ea1ea613f-01
Atomic	2	Included: accumulated atoms/chunks/vectors/tags, conversation text and trace metadata, live/pending/archived wiki parts, report findings/provenance, changed prompts/configuration, derived graph/control state and their shipped admission/maintenance/later-consumer alternatives. Client mappings, task/cache/lease state, alternate backends/strategies, reset/withdrawal and backup paths remain included but incompletely inspected. Excluded: static shipped instructions/assets, current-run-only tool registries/citation ledgers and ephemeral retrieval scores, external provider weights/state and external MCP host internals; benchmark metrics without a production update consumer are not production memory.	partial coverage (Positive database, vector, graph and key/value witnesses do not resolve all shipped client/control/backup state or physical backend alternatives.); Atoms excluding separately classified report findings: known — sqlite [wired], rdbms [wired] (Database retention is recorded for this operative part; adapter transaction and cleanup guarantees are not inferred.); Atom text chunks: inspected SQLite storage and unresolved alternative backend branches: partial — sqlite [wired] (SQLite is the inspected storage witness. The record identifies database chunk rows but does not settle the included alternative backend substrate inventory; complete SQLite-only classification is prevented.); Tag hierarchy and associations: inspected SQLite storage and unresolved alternative backend branches: partial — sqlite [wired] (SQLite is the inspected storage witness. Alternative database branches remain included, with PostgreSQL merge application uninspected; their substrate coverage is unresolved rather than excluded by this finding.); Conversation text and stored tool/citation metadata: inspected SQLite storage and unresolved alternative backend branches: partial — sqlite [wired] (The inspected branch retains chat tables in SQLite. The accepted record does not complete the alternative backend inventory for text and trace metadata; no complete SQLite-only object classification follows.); Live wiki prose and citations: inspected SQLite storage and unresolved alternative backend branches: partial — sqlite [wired] (SQLite publication storage is inspected. Included PostgreSQL/wiki alternatives are not fully classified in the accepted account; the positive branch witness cannot resolve the whole live-object substrate inventory.); Pending wiki candidates and acceptance metadata: inspected SQLite storage and unresolved StorageBackend alternatives: partial — sqlite [wired] (The SQLite branch supplies a positive witness, while the record's StorageBackend proposal storage leaves its full adapter inventory unresolved. Candidate database retention does not establish complete SQLite-only coverage.); Historical wiki checkpoints: inspected SQLite archive and unresolved PostgreSQL branch: partial — sqlite [wired] (SQLite archive rows are directly inspected; the record explicitly leaves PostgreSQL archive semantics uninspected. That included branch remains unresolved for substrate classification as well as parity; only the SQLite positive is asserted.); Report prose, provenance and citations: inspected SQLite storage and unresolved PostgreSQL branch: partial — sqlite [wired] (SQLite retains finding prose/provenance/citations in the inspected transaction. The included PostgreSQL implementation remains uninspected; this witness does not complete the object-wide substrate inventory.); Retained chunk vectors and their search indexes: inspected database/vector branch and unresolved alternative backends: partial — vector [wired], sqlite [wired] (Vectors are stored access state, not service-side model weights. Vector and SQLite findings classify the inspected branch; included alternative database/vector storage is not completely located by the accepted records.); Derived graph edges and centroids: partial — graph [wired] (The dispatch record establishes derived graph retention but does not completely locate every adapter representation.); Changed prompts and configuration: partial — kv [wired] (Registry/global/per-database setting routes establish key/value retention without fully classifying every physical adapter.); Task ledger, report cache, dirty markers, client mappings, import/reset and backup alternatives: uninspected (Accepted records leave physical storage and retained-part inventory of these included alternatives incompletely inspected. Static shipped files, current-run registries/ledgers and excluded provider weights are not added as memory substrates.)	partial coverage (Automatic derived admissions are wired. Manual admission is not asserted without a recorded human decision; unresolved caller-controlled branches remain included.); Model create/edit handler: known — automatic [wired] (Software/model admits each derived write without an independent per-content operator approval in this inspected mechanism.); Single-atom chunk/vector/tag admission: known — automatic [wired] (Software/model admits each derived write without an independent per-content operator approval in this inspected mechanism.); Ordinary model tag-merge application: known — automatic [wired] (Software/model admits each derived write without an independent per-content operator approval in this inspected mechanism.); Wiki candidate persistence: known — automatic [wired] (Software/model admits each derived write without an independent per-content operator approval in this inspected mechanism.); Prior wiki snapshot admission: known — automatic [wired] (Software/model admits each derived write without an independent per-content operator approval in this inspected mechanism.); Direct/immediate wiki publication: known — automatic [wired] (Software/model admits each derived write without an independent per-content operator approval in this inspected mechanism.); Report final finding admission: known — automatic [wired] (Software/model admits each derived write without an independent per-content operator approval in this inspected mechanism.); Scheduled graph/access-state refresh: known — automatic [wired] (Software/model admits each derived write without an independent per-content operator approval in this inspected mechanism.); Core/MCP/import/client initial capture; caller edits/settings; explicit tag pairs; reviewed wiki promotion; saved-answer invocation: not-determinable (Generic caller identity and uninspected UI do not establish an explicit human content decision or exclude one. Physical software I/O does not resolve control.)	partial coverage (Both requested returns and automatic provider supplies exist. A requested enclosing workflow does not turn each internal automatic context selection into pull.); Requested query/ID corpus and live-product readers: known — pull [wired] (Consumer asks for retained material; formatting and response delivery do not add another push.); Requested pending-candidate and archived-version readers: known — pull [wired] (Caller explicitly requests retained candidate/history material.); Automatic history selection inside a new requested chat turn: known — push [wired] (The request asks for a new answer, not retained historical text; software independently supplies selected old message content.); Initial report source batch before research tools: known — push [wired] (Software resolves scope/window/cap and supplies source snippets; later model-requested research reads remain pull.); Research-selected corpus chunks supplied to wiki synthesis: known — push [wired] (Research selects chunks for a later synthesis consumer without that consumer requesting the retained chunks; budget filtering precedes delivery.); Retained prompt and selected tag-tree supply: known — push [wired] (The provider does not request these retained settings/taxonomy; host assembles guidance/context.); Alternative wiki strategies, client automatic supplies and remaining maintenance/cache consumers: uninspected (Uninspected selection/delivery alternatives remain included; known pull readers cannot establish a pull-only system.)	partial coverage (Actual identity matches, budget/scope filters and research judgments classify inspected automatic supplies. No query embedding or identifier on a catalog alone is promoted to push selection.); History push on next chat turn: known — identifier [wired], coarse [wired] (Trigger: new turn. Input: requested conversation ID plus ordered retained groups and token estimates. Selected part: old role/content text.); Initial report-source push: known — coarse [wired], identifier [wired] (Trigger: manual/scheduled report. Input: configured tag/kind/time/source cap and watermark; selected part: retained source snippets. Packing keeps a first-item exception.); Prompt/taxonomy push to chat/tagging: known — identifier [wired], coarse [wired] (Trigger: chat or tagging operation. Inputs: effective setting keys and configured target tree/status. Selected parts: retained custom prompt and eligible tag hierarchy.); Research chunk selection for later wiki synthesis: known — inferred-judgment [wired], identifier [wired], coarse [wired] (Trigger: completed research. Input: retained search results and model-selected session chunk indices; selected parts: retained corpus chunks delivered to synthesis.); Remaining wiki strategies, report seed ordering, alternate client/maintenance pushes: uninspected (Known keyword/vector request retrieval does not itself establish lexical/embedding push. Alternate selectors and retained payload coverage remain unresolved.)	partial coverage (A local report trace-to-product chain supports yes. Missing client/admission and alternate-route facts preclude complete coverage; no unsupported local or system-wide no is asserted.); Report research transcript to automatic durable finding and later product readers: known — yes [wired] (Research messages/tool results feed a separate final writer; structural checks admit the durable finding consumed by later chat/MCP/search. No improvement or weight update is required.); Saved-answer content admission and its later enrichment: not-determinable (The answer-derived artifact and later consumers are wired, but caller UI/per-save human content control is unresolved; no independent unconditional automatic trace-write verdict is assigned.); Other imported session content, client paths, alternative wiki/maintenance and evaluator-fed updates: uninspected (Raw conversation retention, non-summary truncation and diagnostic scores do not by themselves supply a qualifying derived update. Remaining route coverage prevents bounded system absence.)	partial coverage (Local knowledge, instruction, ranking, routing, validation, enforcement and research-update learning paths coexist; chat and tagging have distinct retained instruction consumers. No blanket force is assigned to every stored object.); Selected atom/chunk prose at chat/wiki/report consumers: known — knowledge [wired] (Requested evidence supplies context rather than enforced directions.); Conversation text at later chat invocation: known — knowledge [wired] (Role/content replay is established; historical tool-call metadata is not replayed.); Live wiki prose: known — knowledge [wired] (Authority belongs to this actual reader path, not automatically to every possible model consumer.); Report finding prose: known — knowledge [wired] (Authority belongs to this actual reader path, not automatically to every possible model consumer.); Wiki candidate prose at requesting reviewer: known — knowledge [wired] (Authority belongs to this actual reader path, not automatically to every possible model consumer.); Historical wiki prose at requesting version reader: known — knowledge [wired] (Authority belongs to this actual reader path, not automatically to every possible model consumer.); Corpus text and vectors at query ranking: known — ranking [wired] (Scored retained features influence relative priority even though fixed code supplies the scoring policy.); Tag membership and hierarchy at scope selector: known — routing [wired], enforcement [wired] (Membership changes eligibility under include/require/exclude predicates.); Proposal base ID/timestamp at acceptance: known — validation [wired], enforcement [wired] (The stale-base gate is structural admissibility, not factual acceptance.); Retained custom chat prefix at system-message consumer: known — instruction [wired] (The changed prefix has standing guidance force, not guaranteed compliance.); Retained custom tagging prompt at tagging system-message consumer: known — instruction [wired] (Retained changed guidance directs subsequent tag production; model compliance and classification correctness remain unobserved.); Configuration and retained report/maintenance control markers: partial — routing [wired] (Settings choose branches and report watermarks choose later source windows; detailed lease/cache consumers remain unresolved.); Recorded report research transcript at durable final-product writer: known — learning [wired] (The final writer consumes research messages/tool results and retains prose reused by later readers; no benefit is established.); Graph/centroid details, unconsumed citation metadata, client/backup state and alternate strategy consumers: uninspected (No stronger consumer/force inventory is recorded; metadata retention alone grants no authority.)	AAS-2026-10-12-atomic-f98683232831-01
Auto-claude-code-research-in-sleep	2	Accumulated project wiki pages, relationships and query pack; reviewer responses and derived memory; compact recovery checkpoints; usage and iteration history; trace/audit artifacts; trace-guided proposal/ledger and human-adopted skill revisions; retained phase/thread access metadata and opaque external conversation state on their later-use routes. Includes unresolved global-log consumers, style-profile extraction, trace writers and platform maintenance identified by accepted records. Excludes static unmodified shipped instructions, model inference weights, ordinary unretained working turns, and host scheduling/eviction implementation.	partial coverage (File storage has wired witnesses for helper-produced artifacts and receipts, and claimed witnesses for procedural reviewer, checkpoint, proposal/ledger and skill-replacement writes. External conversation state and the remaining partial-loop inventory prevent complete enumeration.); Wiki pages and cached projection: known — files [wired] (Markdown pages and query pack are operative disk files.); Relationship triples and quarantine history: known — files [wired] (JSONL triples are files; the graph is their logical organization, not an evidenced separate graph database.); Reviewer memory and compact recovery checkpoints: known — files [claimed] (The procedures specify retained Markdown files; their production and replay are not observed.); Usage events and iteration history: known — files [wired] (Event summaries and iteration records are JSONL disk files.); Edited skill body, staged patches and cycle ledger: known — files [claimed] (File storage is classified for the specified proposal/ledger and skill-replacement branches at their claimed procedural basis, without borrowing receipt implementation strength.); Provenance receipt sidecars: known — files [wired] (Wired file storage applies to receipt stamping, not the separate model-followed proposal, ledger or skill-replacement procedures.); Review request/response and acquittal artifacts: known — files [claimed] (The trace protocol specifies filesystem artifacts; serialization helper is uninspected.); Phase and per-thread access metadata: known — files [wired] (Run JSON and per-thread option files persist symbolic access/recovery metadata.); External CLI/provider conversation payload: uninspected (The bridge delegates conversation retention; neither its storage substrate nor opaque payload internals is inspected.); Included remaining maintenance and platform routes: uninspected (Global-log consumers, style-profile extraction, trace writers and platform alternatives remain uninspected; known file witnesses do not resolve their operative storage.)	partial coverage (Admission control is separate from authorship and physical I/O. Model/software admissions are automatic, explicit patch selection/landing is manual, and generic helper callers remain unresolved.); Hook event append and iteration derived-state append: known — automatic [wired] (Once invoked/deployed, software admits each summary/count without a per-content human verdict.); Deterministic pack rebuild: known — automatic [wired] (Code admits the selected extracts; an initiating request is not approval of every extract.); Skill-driven paper capture/enrichment and evaluation write-back: known — automatic [claimed] (Models admit page content and outcome changes without mandatory per-content human approval.); Manual-field paper metadata admission: not-determinable (Manual metadata fields are an alternative input format, but the generic caller is not authenticated as a human making a per-content admission decision.); Generic direct helper/API writes: not-determinable (Caller identity/control is unspecified; physical Python I/O does not determine whether admission is human or computational.); Reviewer append, derived memory and compact production: known — automatic [claimed] (The model-followed procedures admit outputs without a required per-append or per-summary human decision.); Meta diagnosis, pending diff and bottleneck report production: known — automatic [claimed] (Producer model creates proposals and durable cycle evidence before human selection; this is separate from live skill adoption.); Landed skill replacement: known — manual [claimed] (Human selects patches and separately invokes landing; jury PASS is additionally required. Applier I/O does not erase the explicit content-adoption decision.); Trace-writer, external conversation and remaining maintenance admission: uninspected (Uninspected writers and provider internals prevent a complete admission-control account.)	partial coverage (Requested retrieval and automatic later-context selection coexist. File rebuild alone is not push; opaque payload internals and uninspected routes prevent complete coverage.); Idea-creator request for cached/rebuilt wiki pack: known — pull [claimed] (Later consumer requests retained context, scans and Reads it; mutation-time projection is production, not unsolicited delivery.); Explicit reviewer thread continuation: known — pull [wired] (Caller supplies a thread ID to request read-back; normal requested tool return is not an additional push.); Hard-mode reviewer full-memory injection: known — push [claimed] (The next round triggers automatic supply of full retained reviewer memory into the assigned prompt, without a new reviewer request for that content.); Reviewer file-path access and medium-thread continuation: known — pull [claimed] (Reviewer can request the named memory file or continue its thread; path supply alone does not prove its content was read.); Compact checkpoint selection for downstream context: known — push [claimed] (A later experiment/review invocation automatically selects available compact summaries under the COMPACT/recovery procedure.); Next-cycle optimizer reading prior ledger and project events: known — pull [claimed] (The maintenance invocation requests retained usage and prior-cycle evidence; no idle-host context injector is established.); Iteration logger prior counter and explicit direction-history show: known — pull [wired] (Next note reads the last valid retained count as input; explicit show requests the recorded ledger.); Phase reload and per-thread options: known — pull [wired] (Resume/reply requests cause loading retained metadata.); Later host loads changed skill: known — pull [claimed] (A later skill invocation requests its installed procedural body; static unmodified instructions are outside this memory boundary.); Opaque conversation internal selection: uninspected (Known requested continuation does not resolve internal delivery/selection operations.); Included remaining maintenance and platform routes: uninspected (Global-log consumer, style-profile reload, trace consumers and platform alternatives retain uninspected read-back paths.)	partial coverage (Only push selection is classified. Coarse route/availability selectors are not semantic or identity matching; pull-only local operations do not settle opaque alternatives.); Hard reviewer-memory prompt supply: known — coarse [claimed] (Trigger: next review round. Selector inputs: backend, difficulty and round. Selected part: entire reviewer memory file, without relevance filtering or truncation.); Compact recovery context supply: partial — coarse [claimed] (Trigger: later experiment/review initialization. Selector inputs: COMPACT flag and file availability. Selected parts: candidate checkpoint or findings plus experiment log. File names are not targeted identity matches among retained content.); Requested wiki pack read: inapplicable (This local delivery fulfills a consumer request. Filename order/outcomes/budgets select the requested projection, not unsolicited push context.); Explicit phase/thread/ledger/skill requests: inapplicable (These local requested returns need no push signal; this does not classify provider-internal alternatives.); Provider-internal context supply: uninspected (The bridge does not reveal how conversation parts are selected or whether additional push operations exist.); Included remaining maintenance and platform routes: uninspected (Uninspected style-profile, global-event and platform consumers prevent a complete inventory of push selectors.)	partial coverage (Positive classifications concern automatic trace-fed durable context/guidance, not demonstrated improvement. Manual skill adoption is distinct from automatic proposal/ledger production, and raw logging alone is insufficient.); Compact experiment summary to later review: known — yes [claimed] (Experiment activity feeds automatic model-authored retained results/verdict/reproduction summaries, later selected as review context. This qualifies by design even though runs are unobserved.); Reviewer-authored memory update to later reviewer: known — yes [claimed] (Recorded review experience feeds automatically admitted durable suspicions/judgments, subsequently delivered to later rounds. Verbatim append alone would not establish this finding.); Experiment outcomes/lessons to later ideation: known — yes [claimed] (The instructed automatic verdict/outcome update is retained and projected into later ideation as guidance; generic API content is not thereby classified.); Usage evidence to retained diagnosis/bottleneck ledger and next optimizer: known — yes [claimed] (Trace-fed model diagnosis and durable cycle ledger are produced before human landing, then requested by a later maintenance cycle. The qualifying automatic artifact is the proposal/ledger, not the manually adopted skill.); Human-approved live skill replacement: inapplicable (Explicit human patch selection/adoption makes this admission manual. It cannot supply an automatic trace-learning witness, even though earlier proposals can.); Counter/phase/options persistence in isolation: not-determinable (Counters and ordinary recovery/access metadata do not by themselves establish a durable behavior-shaping artifact update. No bounded system-wide negative is inferred.); Candidate checkpoint and findings production: not-determinable (Candidate condensation is specified, but complete trace origin and findings production are missing; the known experiment-log witness does not resolve these alternatives.); External conversation maintenance and trace consumers: uninspected (Opaque retained history and uninspected serialization/consumer internals prevent qualifying or excluding their update routes.); Included remaining maintenance and platform routes: uninspected (Style extraction, global-log use and platform maintenance may add qualifying routes; raw event logging alone is not counted.)	partial coverage (Classifications name actual consumers, not authority inferred from a status name. Requested use remains claimed; opaque and uninspected consumers remain unresolved.); Wiki excerpts at later idea model: known — knowledge [claimed] (Papers and gaps provide reference/context to ideation through a scanned file read.); Failed-idea outcomes at later ideation choice: known — ranking [claimed] (Failed ideas bias candidate priority/exclusion as a model-interpreted banlist. This is not a code-enforced action veto.); Outcome fields at projection selector: known — routing [wired] (Retained negative/mixed outcomes select which idea snippets enter the pack.); Reviewer responses and derived memory at later reviewer: known — knowledge [claimed] (Past critiques and unresolved suspicions provide context for subsequent assessment.); Reviewer experience at memory-update producer: known — learning [claimed] (Reviewer-owned accumulated review experience feeds a durable memory update for later critique; no improved review accuracy is inferred.); Compact summaries at later experiment/review consumer: known — knowledge [claimed] (Selected ideas, results and reproduction inputs guide later invocations as context rather than standing host rules.); Iteration stale counter at next logger call: known — routing [wired] (Last valid stale count determines next pivot/escalation decision.); Iteration direction ledger at subsequent host tick: known — knowledge [claimed] (Recorded tried directions supply context to avoid repeating the same frame.); Usage summaries and prior bottleneck ledger at optimizer: known — learning [claimed] (Accumulated experience supplies evidence for a durable diagnosis/patch/ledger update. The later skill admission remains human-gated.); Revised skill body at future host invocation: known — instruction [claimed] (Landed procedural text supplies standing directions when loaded by a later host.); Phase status/order/policy at resume selector: known — routing [wired] (Retained status/policy determines the next unfinished stage, not scientific acceptance.); Saved model/config/sandbox/cwd at continuation launcher: known — routing [wired] (Saved options choose continuation configuration and scope; external CLI owns its enforcement.); Provenance receipt/hash at is_auto_curatable: known — validation [wired], enforcement [wired] (Stored accepted fields and current hash are criteria for automatic-curation eligibility, and a failing check refuses eligibility when the consumer invokes it.); Run-scoped acquittal artifacts at review procedure: partial — validation [claimed] (Current receipts/evidence pointers support stop eligibility checks; writer fidelity and native validator internals remain unresolved.); Quarantine raw history later consumer: not-determinable (Human review is afforded, but no actual automatic later consumer of raw quarantine history is established; retention alone has no authority.); Opaque provider conversation at resumed reviewer: uninspected (The requested continuation is known, but what retained internal parts exert which force is not inspected.); Included remaining maintenance and platform routes: uninspected (Global log consumers, style-profile use and remaining platform maintenance are uninspected; no blanket authority union is complete.)	AAS-2026-10-15-auto-claude-code-research-in-sleep-a8abf5f80cbb-01
auto-harness	2	Retained Terminal/BIRD diagnostic copies and their source-job lifecycle; requested learning entries; through-use candidate Python/instruction revisions and their later adoption; last CLI results, score history, suite membership/threshold/diagnostics and baseline-derived task splits. Includes unresolved caller-controlled content admission and requested note reuse. Excludes unchanged shipped program/templates, ordinary Terminal/tau/BIRD task/session accumulation and diagnostic previews as supposed model checkpoints, external coding-host memory, provider state and external framework internals.	known coverage (Complete substrate inventory within the local retained boundary; excluded ordinary sessions and external models do not add in-memory or model-weight memory values.); Diagnostic copies/source jobs, learning log, CLI results, score history, suite parts and task splits: known — files [wired] (These records enumerate the local retained noncandidate surfaces; suite parts share their container's file substrate.); Through-use candidate code and instruction revisions: known — files [afforded], repo [claimed] (File and prescribed Git retention classify this revision surface, not the coding model's own storage.)	partial coverage (Automatic local transformations and prescribed model admission do not resolve the separate caller/configuration and preserved-content mechanisms.); Adapter result/trace writes, baseline history/splits, suite initialization, diagnostic overwrite and post-check membership promotion: known — automatic [wired] (Starting preparation or evaluation does not make subsequent deterministic admission manual.); Requested coding-agent learning entries and candidate changes: known — automatic [claimed] (Classification of the requested admission mechanism, not demonstrated autonomous editing.); Score-number submission to recorder and configuration/previous-content admission: not-determinable (Recorder serialization is software, but the substantive supplying/approval decision is caller-controlled and unresolved. Existing workspace contents are preserved. No explicit human decision establishes manual admission, and generic caller identity does not establish automatic admission of those contents.)	partial coverage (Known pull and automatic instruction push coexist with unresolved note/diagnostic read-back.); Gate loaders, split lookup and executable candidate adoption: known — pull [wired] (Decoding and returning requested data are not an additional unsolicited push.); Foreground failure-trace diagnosis: known — pull [afforded] (No exercised read or local loader is supplied.); Revised instruction constants supplied to task models: known — push [wired] (Automatic instruction delivery is separate from the upstream requested Python import; cached runners qualify replacement timing.); Learning-log reuse and historical diagnostic consumption beyond serialization: not-determinable (The log's memory label supplies no reader/selector; suite diagnostic loading supplies no consequential historical-diagnostic reader. Direction cannot be assigned to these unresolved later-consumer paths.)	partial coverage (A coarse instruction-supply witness does not resolve unknown note reuse or convert task-identified pull into targeted push.); Automatic delivery of revised candidate instruction constants: known — coarse [wired] (Mode/configuration availability is coarse selection; task IDs are not a match selecting historical prose.); Gate/split/import requests and foreground trace inspection: inapplicable (These operations fulfill explicit consumer requests and are pull; filenames and task IDs on them do not establish identifier-based push.); Unresolved learning-log and historical-diagnostic delivery: not-determinable (No established selection/delivery mechanism decides whether these are additional pushes or which signal would select their contents.)	partial coverage (Known wired selection updates and an afforded candidate-update route coexist with unresolved log reuse and caller/upstream alternatives. No system-wide no is asserted.); Outcome-fed regression-suite promotion and baseline split generation: known — yes [wired] (The raw feedback-to-durable-selection-to-later-consumer chain is recorded; actual operation is unobserved.); Trace-responsive candidate revision and later imports/model instruction delivery: known — yes [afforded] (Standing trace-fed revision capacity is independent of the report's uninspected future-capacity learning and causal-benefit conclusions.); Learning-log entries and their proposed later consumer: not-determinable (Automatic note writing is requested, but an established later consumption route is missing. Raw trace retention and a memory label do not close this chain.); Caller-supplied score content and external payload transformations: not-determinable (Caller control and opaque upstream processing prevent a complete classification of additional trace-fed durable updates. No bounded global negative follows.)	partial coverage (Preserves implemented gate/selection and instruction channels without upgrading arbitrary executable revisions, unconsumed diagnostics or requested note reuse.); Retained diagnostic traces at outer diagnosis and subsequent revision: known — knowledge [afforded], learning [afforded] (Separate evidence and update consumers/effects within the same prescribed route.); Suite IDs, split membership and prior CLI failures at future task selection: known — routing [wired] (Selection is fixed code, not semantic ranking.); Suite threshold, suite membership denominator and history maximum at gate decision: known — validation [wired], enforcement [wired] (Validation and enforced gate-return consequences coexist but do not authenticate inputs.); Outcome feedback at suite promotion and baseline split construction: known — learning [wired] (Update-consumer authority is distinct from the downstream task-selector authority.); Revised instruction constants at later task-model calls: known — instruction [wired] (This classifies the instruction part, not all executable Python as instruction or enforcement.); Arbitrary revised Python's particular downstream effects and learning-log later use: not-determinable (Imports are wired but absent revisions do not identify particular retained controls or their force. The learning log has no established loader or actual reader, preventing a definite authority classification at reuse.); Historical suite diagnostics and accounting metadata beyond consumed score fields: not-determinable (Merge/serialization and commit/timestamp bookkeeping establish no further consequential consumer; no additional force is assigned.)	AAS-2026-10-15-auto-harness-92303e444cb2-01
autocontext	2	Retained material accumulated or changed through use in Python and npm: immutable context-bundle parts, mutable playbooks/hints and SKILL lessons, generation reports/replays/metrics and SQLite history, package payloads and snapshots, compaction entries and their access structures, training checkpoints, serving bindings and evolving training/kernel source. Includes their admission, maintenance, prompt delivery, requested MCP/Pi retrieval and afforded host-file consumption. Included unresolved alternatives are notebook/report/hint writers, evidence and production-trace/dataset processing, npm maintenance and memory packs, session/mission reload, configured hooks, provider/host retained state and model-serving reload. Excludes unchanged shipped instructions and ordinary transient strategy/round state unless retained for a later invocation. Session extraction-only returns are not treated as demonstrated durable promotion.	partial coverage (The accepted inventory supports file, SQLite, repository and cache storage, while opaque and uninspected stores prevent complete enumeration.); Bundle advice, routing JSON, four mutation payloads, reference helper source and digest-scoped executable validators: known — files [wired] (The container enumerates eight operative parts; each is retained in the file-backed bundle.); Mutable lessons, playbooks, hints, package-installed source, snapshots and compaction ledger: known — files [wired] (Covers inspected file objects, not complete opaque host state reached through skill discovery.); Generation artifacts and imported attributed history: known — files [wired], sqlite [wired] (File bodies and database rows are separate supported substrates within this inventory.); Python component-compaction lookup cache: known — in-memory [wired] (This is an access structure, not an opaque full-model checkpoint.); Local training checkpoints, registry artifacts and evolving training/kernel source: partial — files [wired], repo [wired] (Backend/serving internals and complete registry reload are not inspected; file paths do not establish immutable parameter identities.); Opaque provider/host sessions, npm memory packs, mission/session state and remaining notebook/evidence/trace stores: uninspected (Adapters/loaders establish included seams, not their complete operative storage inventory. No service, vector or other substrate is inferred from those seams.)	partial coverage (Automatic implementation and afforded manual package admission coexist; approval labels do not resolve caller-dependent branches.); Software bootstrap/overlay and matched bundle admission, lesson accumulation/curator replacement, artifact/ledger persistence and default npm publication: known — automatic [wired] (Known subsequent admission remains valid even when legacy authorship is unknown.); Explicit operator package selection through CLI: known — manual [afforded] (This is the operator-controlled branch, not evidence of an actual import or of human control at every caller.); Training source/checkpoint keep-discard, automatic publication and eligible ambient activation; kernel champion admission: known — automatic [wired] (Ambient approval-required policy is a separate held branch, not an automatic-activation witness.); Lifecycle action callers, MCP/library import, optional npm approval, ambient required approval and direct/hook-modified writers: not-determinable (Generic caller identity, an approve method or a configuration flag does not establish an explicit human content decision; configured hook control is unresolved.); Initial legacy content and remaining notebook/hint/report/npm/session/mission admissions: uninspected (Initial authorship/control and included untraced writer mechanisms prevent complete admission coverage.)	partial coverage (Both wired push and pull exist, plus an afforded file pull; unresolved alternatives prevent a complete direction inventory.); Generation setup selects retained advice/reports and active routing/validator parts before later role/evaluation invocations: known — push [wired] (Internal storage reads are steps within automatic supply, not independently requested foreground retrieval.); Registered MCP knowledge/search and Pi runtime snapshot tools: known — pull [wired] (Handler delivery fulfills that request; it is not a second push merely because code returns data.); Documented external file-consumer role: known — pull [afforded] (Symlinks and resource links do not establish automatic host injection.); Training/kernel campaign context and ambient registry/serving selection: partial — push [wired] (Complete model-serving reload and downstream inference delivery are uninspected.); Session/mission reload, opaque provider/host continuation, npm memory packs and extraction-result callers: uninspected (Included alternatives lack complete retention-to-later-consumer routes; known tool retrieval cannot make them pull-only.)	partial coverage (Push has identifier, coarse and lexical selectors; pull units are locally inapplicable, not evidence that the whole axis is inapplicable.); Generation-triggered active-bundle and scenario/run/role-specific retained-part selection: known — identifier [wired], coarse [wired] (Identity matching selects actual delivered parts; budget/availability filters remain separate from targeting.); Component-compaction keyword/structure selector within automatic prompt delivery: known — inferred-lexical [wired], coarse [wired] (Deterministic lexical selection is not embedding retrieval or model relevance judgment.); Campaign primary-feedback disclosure and ambient active-target choice: partial — coarse [wired], identifier [wired] (Complete training-data/serving reload selection and model-internal memory selection are unresolved.); MCP/Pi requested artifact/search return and documented requested file consumption: inapplicable (These completely assessed direction units are pull. IDs and lexical top-k in a requested return do not establish push signals.); Opaque host continuation, remaining notebook/evidence/npm packs and session-result callers: uninspected (Untraced delivery leaves possible selectors unresolved; no inferred-embedding or inferred-judgment value is established for them.)	partial coverage (Qualifying automatic trace-fed paths exist. Unresolved durable callers, trace provenance and opaque continuation remain included; no system-wide no or observed learning benefit is asserted.); Automatic attempt/replay-to-SKILL warning and lesson/curator update followed by later role prompts: known — yes [wired] (Support roles precede the current tournament; the deterministic nonadvance warning consumes current results at persistence. The positive does not require a coach to review an as-yet-unrun tournament.); Training/kernel iterative retained source and feedback-to-successor path: partial — yes [wired] (The source-update chain qualifies; complete dataset-to-weight trace provenance and exercised serving are not established by checkpoint publication.); Component compaction and durable audit summary as a possible continuation artifact: not-determinable (Trace-containing text is reduced and a summary retained, but later ordinary generation consumes source components, not demonstrably the retained ledger summary as behavior-shaping continuation. Requested inspection alone does not resolve that path.); Session extraction returns and possible durable callers: not-determinable (The inspected functions return candidates without durable promotion. Uninspected callers prevent a bounded negative for the complete route.); Remaining bundle/role-proposal trace provenance, ambient datasets, production traces, notebook/hint/report writers and opaque host continuation: uninspected (Generation requests and static imports alone do not establish an original trace-to-durable-artifact-to-later-consumer chain for every included alternative.)	partial coverage (All seven authority values have local supported consumers; this is not complete authority coverage or an observed-effect claim.); Retained playbook, hints, lessons and mutation advice at later role-model prompts: known — instruction [wired] (Scope is standing directions actually selected into prompts; delivery is not proof of compliance.); Prior reports, analysis, trajectory and noncallable helper reference at role consumers; requested retained artifacts at host inspection consumers: known — knowledge [wired] (Knowledge force does not authenticate source truth; installed imported validator source is not granted this route's authority.); Emitted skill/playbook files at the documented external file consumer: known — instruction [afforded], knowledge [afforded] (Keeps afforded external consumption distinct from wired generation and MCP/Pi paths.); Bundle routing configuration at orchestrator synchronization: known — routing [wired] (This is fixed-rule configuration, unlike context-policy prose interpreted by models.); Bundle executable validator at enabled validity gate and retry consumer: known — validation [wired], enforcement [wired] (Conditional wiring is known; actual validator execution, correctness and security are unobserved. Prompt checklist strings are excluded from this unit.); Explicit stale markers at the SKILL lesson reader: known — enforcement [wired] (The force belongs to retained status at this particular reader, not to universal withdrawal.); Retained text at requested lexical search selector: known — ranking [wired] (Search returns metadata, not automatic injection of the highest-ranked playbook.); Retained experience/lessons at curator update and campaign history at training/kernel proposal consumers: known — learning [wired] (Update-consumer force is separate from downstream instruction force and does not establish improved capacity.); Registry/probation/evaluation state at ambient candidate selection and serving binding: known — ranking [wired], routing [wired], validation [wired] (Scope is metadata consumers, not an assertion of reliable learned behavior in opaque weights.); Imported harness execution, checkpoint behavioral contents, notebook/evidence/npm memory packs and opaque host consumers: uninspected (Installed source or a readable summary cannot establish force at untraced consumers; parameter-level knowledge and host retention remain unresolved.)	AAS-2026-10-15-autocontext-b7f59f8405c4-01
AutoSci	2	Included: accumulated raw/source-derived wiki pages and outputs, semantic and projected relations, citations, index, gaps, context brief, audit-tail reuse, initialization/discovery/research checkpoints and conditional Git fan-in, process-local review threads, and Hiera contracts, candidates, attempts, ledger, replay, experience reporting and final reports. Includes their admission, maintenance and later-consumer alternatives, including partially inspected source preparation, regeneration, proposal/tuning and dissemination. Excluded: shipped static skills/schema/policy and setup copies as memory, host conversation compaction and provider-private retention/weights. Those excluded dependencies are not classified from the local implementation.	known coverage (Known for the explicitly scoped stores, including volatile and conditional repository parts. This does not classify the excluded host/provider storage or assert observed persistence.); Local raw/wiki content, derived navigation/graphs, audit log and recovery files: known — files [wired] (Recorded stores cover this local file inventory; incomplete extraction or regeneration inspection does not obscure the named substrate.); Hiera contract, candidate/attempt payloads, ledger and final report: known — files [wired] (Full experience dictionaries are not included as durable file payloads merely because a report stores their revision.); Review conversation history and reconstructed Hiera experience/state: known — in-memory [wired] (Volatile objects are distinct from their file-backed inputs and report digests; provider-private state is excluded.); Actually committed initialization and daily-intake artifacts: known — repo [afforded], repo [wired] (Classification is conditional repository retention, not a claim that all local file memory is versioned or transported.)	partial coverage (Both agencies occur on separate mechanisms. Initial and generic-caller admissions are not resolved by downstream automatic writes or human-gate doctrine.); Compiler, gap/projection/dedup, automatic lint repair and ledger/reconciliation writes: known — automatic [wired] (Physical file I/O is not the basis: code owns content selection and reconciliation.); Model admission during ingest, seeding, eliminated-idea retention, refinement and auto evaluation: known — automatic [afforded] (Afforded model writes do not become observed admissions or deterministic compliance.); Unattended CI intake admission: known — automatic [wired] (Repository retention remains conditional on staging/tracking, independently of agency.); Explicit human crystallization authorization, user-specified edit and interactive verdict approval: known — manual [claimed] (Does not extend manual classification to basic ingest, auto verdict or arbitrary browser callers.); Generic browser/direct CLI writes, checkpoint initialization, reset flags and Hiera approval/admission caller identity: partial — manual [claimed] (Executable caller identity and --yes/approve commands do not authenticate a person; initial content control and generic direct-write agency remain unresolved.)	partial coverage (Both requested access and automatic recommendation-context supply are supported, but not a complete delivery inventory.); Requested wiki/brief/gap/index/page reads, rebuild returns, recovery and browser reads: known — pull [afforded] (File names and top-K relevance selection do not by themselves add a push delivery operation.); Later review replies and Hiera loop/status/tune replay requests: known — pull [wired] (Requested reloads are separate from downstream automatic selection of the next execution.); Scheduled/manual daily intake preparation supplies retained wiki/audit context to recommendation: known — push [wired] (Operator workflow invocation is upstream of automatic selection; compatible model receives further compacted evidence, not the full corpus.); Other host-context, source-preparation and dissemination alternatives within recorded local routes: uninspected (Records do not establish every later delivery operation; intended worktree transport is additionally conditional on actual committed payloads. Excluded host compaction is not inferred.)	partial coverage (Coarse profile selection is supported; no inferred embedding, lexical relevance, judgment or identifier push is inferred from mere file addressing or action selection. Other push alternatives remain unresolved.); Daily intake wiki-profile push to recommendation: known — coarse [wired] (Keyword/gap extraction further transforms selected text. Selection of new external papers by lexical scores is not itself read-back of retained memory.); Requested wiki/recovery/thread/ledger reads and compilation returns: inapplicable (These units fulfill explicit retained-material requests; filenames, thread IDs and checkpoint keys are not evidence of targeted push.); Additional automatic selection/delivery inside Hiera and uninspected local consumer alternatives: not-determinable (Score-directed choice of an execution is established, but records do not separately establish every retained-payload delivery/identity-match operation or other automatic injection route. Cannot use known pull paths to declare this unit inapplicable.)	partial coverage (Supported yes findings coexist with unresolved branches. This axis is automatic trace-fed artifact updating, not demonstrated improved capacity, scientific validity or self-improvement.); Automatic pilot/formal outcome-to-negative-memory update and later ideation: known — yes [afforded] (Filter rejection without execution feedback is not silently equated to experiment trace learning. Actual use or improvement is unobserved.); Failure-summary compilation into a retained brief for later agents: known — yes [afforded] (Non-ampliative summarization still qualifies; detailed reasons may be omitted under budgets.); Hiera feedback and attempt-bound durable reconciliation consumed by later loop/tune: known — yes [wired] (This names the operative ledger-to-decision route, not mere retention of stdout or the separately reported experience revision.); Audit-tail recommendation preparation, process-local review history, report-only experience digest and remaining opaque write alternatives: not-determinable (Context use/count reporting is known, but these records do not establish for every branch an automatic trace-fed durable behavior-shaping output and later consumer. No bounded absence record warrants a local no.)	partial coverage (Seven forces are supported on named consumers, with local bases preserved; uninspected alternatives prevent a complete inventory.); Full wiki content and compressed text at later answering/research/review consumers: known — knowledge [afforded], instruction [claimed] (Different forces belong to evidence text and negative guidance even when the same brief carries both.); Retained content/features consumed by context and intake selectors: known — ranking [wired], ranking [claimed] (Fixed scoring still gives its retained inputs ranking force; it need not be a learned ranker.); Recovery metadata and Hiera retained execution controls at their code/agent consumers: known — routing [afforded], routing [wired], ranking [wired], validation [wired], enforcement [wired] (These are accumulated per-run controls, not the excluded shipped writer/schema policy.); Retained experiment feedback at the durable negative-guidance update consumer: known — learning [afforded] (This does not establish improved capacity, calibrated science or exercised self-improvement.); Review-thread messages at later provider calls and experience revision at report consumers: known — knowledge [wired] (Display/report consumers are not substituted for the distinct replay/scheduler path.); Uninspected source conversion, regeneration, remaining dissemination and opaque payload consumers: uninspected (Records do not enumerate every such consumer/effect; known paths cannot establish complete authority coverage.)	AAS-2026-10-15-autosci-430df584df8f-01
Basic Memory	2	Included: accepted and materialized notes; graph/search/vector access structures; note revision/publication state; raw lifecycle archives; Pi direct and Python-hook checkpoints; Tau handoffs and persisted receipts; Hermes continuity; OpenClaw captured notes and cross-invocation recall cache; retained schema criteria and stored-case evaluator consumers. Included unresolved alternatives: filesystem/import/CI/cloud admission, other read/maintenance paths, opaque host/provider accumulated state and evaluator feedback/persistence. Excluded: static shipped instructions/skills as memory objects, current request drafts and prepared current-mutation publication payloads without established later memory consumers, inference-only model weights as accumulated memory, external host internals and uncaptured live infrastructure. Shipped skill prescriptions remain evidence about potential writes and maintenance, not evidence of exercised operation.	partial coverage (Positive local substrates do not enumerate the partial-loop memory inventory. Host receipts, vector storage, remote alternatives and evaluator persistence remain unresolved; no sqlite, repo or model-weights finding follows from packaging or static model availability.); Accepted note bodies and note identity/version/publication state: known — rdbms [wired], files [wired] (The accepted database body and materialized file are distinct completion points, both established locally.); Published graph relations and observations at later context traversal: known — graph [wired] (Classifies published graph memory consumed by later traversal, not every graph repository implementation. The process-local prepared payload belongs to current-mutation publication machinery; the records do not establish its qualifying retention and later memory consumption, so it supplies no substrate finding.); Raw lifecycle inbox and processed envelopes: known — files [wired] (The local JSON archive substrate is explicit; it is not a conversation-note graph.); OpenClaw recall cache between bootstrap and assemble: known — in-memory [wired] (This short-horizon accumulated selection is separate from durable note storage.); Tau persisted host branch receipts and branch trace storage: not-determinable (Host custom entries are persisted, but the backing substrate and guarantees are not inspected; a host entry label does not establish files, a database or service-object.); Persisted embeddings, semantic chunks, caches and opaque provider-side accumulated state: uninspected (Vector work and numerical provider outputs are established, but accepted records do not trace vector-table/backend/cache storage or accumulated opaque provider state. Inference weights are not automatically accumulated memory.); Other importer/watch/CI/cloud stores and retained evaluator artifacts: uninspected (The recorded inventory leaves alternate storage paths and evaluator persistence incomplete. QA in-process objects cannot be joined to the separately described benchmark files without a recorded path.)	partial coverage (Automatic checkpoint and downstream state admission coexist with a claimed human research gate. Generic caller requests and untraced alternatives prevent complete agency classification.); Pi direct and Python hook automatic checkpoint admission: known — automatic [wired] (Workspace trust and project configuration enable automation; they are not decisions approving each checkpoint.); Tau, Hermes and OpenClaw enabled capture/update admission: known — automatic [wired] (Lifecycle/manual triggering does not by itself change subsequent content admission to manual.); Derived publication, sync state, raw envelopes and Tau receipts: known — automatic [wired] (Classifies these downstream state writes separately from unresolved upstream generic note admission.); Research workflow saving decision: known — manual [claimed] (Doctrine names a per-content human decision; ordinary service admission does not enforce this workflow.); Reflection content admission: known — automatic [claimed] (Model-facing prescription, not an observed or autonomously scheduled execution.); Generic note/API callers and direct filesystem, importer, CI/cloud and initial-state admission: not-determinable (Generic caller identity cannot resolve human versus model control. Alternate admission bodies and initial control remain incomplete.); Provider-side accumulated writes and evaluator persistence/adoption: uninspected (Inference calls and stored-case availability do not establish every provider write or how evaluator artifacts are admitted into later operation.)	partial coverage (Both requested retrieval and automatic host supply are wired. Unresolved read-back alternatives remain included; known pull routes do not turn the signal axis inapplicable.); Requested note/graph reads and explicit Pi/Hermes recall: known — pull [wired] (These are established requested operations, not proof of every alternative reader.); Pi direct automatic recall and Python hook session briefs: known — push [wired] (Automatic selection is distinct from explicit recall commands within the same integrations.); Tau orientation and post-compaction supply chain: known — pull [wired], push [wired] (The internal requested read and outer automatic context supply are distinct operations, not double-counted delivery.); OpenClaw bootstrap and later assemble: known — push [wired] (Automatic retained-context selection is push even though underlying search calls return requested rows.); Tau previous-handoff update input, archive replay keys and sync/recovery reads: known — pull [wired] (Read-back includes later deterministic consumers; not every archive is read by a model.); Stored QA case rejudging: known — pull [wired] (No untraced production injection or artifact-file path is assumed.); Hermes outer injection, OpenClaw child-note readers, alternate tools, cloud/provider and evaluator consumers: uninspected (Known pull and push witnesses do not resolve the incomplete outer-host and alternative consumer inventory.)	partial coverage (Signals characterize automatic push selection only. Prompt query, identity match and coarse eligibility remain distinct; underlying requested reads and catalog identifiers do not themselves establish a push signal.); Pi direct before-agent-start checkpoint push: partial — coarse [wired] (The current prompt or fallback query drives search, but the records leave ranking/provider behavior uninspected. Query presence alone does not resolve lexical versus embedding or judgment selection.); Python hook session-start brief push: partial — coarse [wired], identifier [wired] (Memory annotation supplies the task/decision/session queries and repository scope; other adapters and exact search internals remain incomplete.); Tau orientation and successful-compaction pointer push: partial — identifier [wired], coarse [wired] (Identity selection is actual branch-to-capture selection, not merely an identifier printed on a note. Topic discovery search details and complete placement policy remain unresolved.); OpenClaw bootstrap recall push: known — coarse [wired] (Fixed class/status/time selection is coarse; no prompt-specific semantic selector is established here.); Hermes possible background delivery and other host/remote automatic supply: uninspected (Outer Hermes context injection and alternative automatic consumers are incomplete; their selector inputs and selected retained parts cannot be exhaustively classified.)	partial coverage (Positive trace-fed durable-context routes are local wired yes findings, independent of the reports’ unestablished improved-capacity learning property. Opaque and untraced writes remain unresolved rather than being classified no.); Pi direct and Python hook durable checkpoints with later recall: known — yes [wired] (A wired trace-fed durable context route qualifies even without parameter training, semantic novelty or observed benefit.); Tau incremental handoff with subsequent summary/orient consumers: known — yes [wired] (The external model is opaque; the trace-to-durable-handoff-to-later-consumer chain is wired.); Hermes extractive summary/session notes with later prefetch: known — yes [wired] (The code-wired later search/preview path qualifies without guaranteeing selection of a particular note or outer-host use.); Prescribed reflection of recorded experience: partial — yes [claimed] (Claimed trace-fed behavior-shaping output is not an implemented scheduler or observed learning benefit.); OpenClaw daily conversation and child handoff learning routes: not-determinable (Automatic trace writes are wired, but bootstrap reads tasks/recent titles rather than conversation bodies and a child handoff reader is not established. Tools afford access without completing every qualifying later-consumer chain.); Raw archive alternatives, generic imported traces, provider updates and evaluator-to-policy routes: uninspected (The inspected raw sweep is bookkeeping, not a behavior-shaping update; broader archive readers, generic content origins and further automatic update consumers remain uninspected. No system-wide no is warranted.)	partial coverage (The recorded consumer paths support multiple independent forces. Static shipped directions are excluded; retained payload placement is not itself instruction or enforcement. Unknown consumers prevent an exhaustive authority union.); Note/checkpoint prose delivered to later readers: known — knowledge [wired] (Delivery establishes a knowledge channel, not compliance, factual endorsement or benefit.); Prior Tau handoff at the next durable handoff update: known — learning [wired] (Names the update consumer separately from downstream knowledge supply; it does not claim improved capacity.); Retained text and indexed representations at query ranking/selection: partial — ranking [wired] (Local query/hit selection is recorded; full ranking algorithms, optional reranking and all consumers remain uninspected.); Sync/version metadata at mutation and materialization gates: known — validation [wired], enforcement [wired], routing [wired] (These forces are operational consistency controls, not truth criteria.); Tau persisted receipts at orient and uncertain-write recovery: known — routing [wired], validation [wired], enforcement [wired] (Distinct receipt consumers give distinct effects; host persistence guarantees remain outside inspection.); Retained schema definitions at requested structural validation: known — validation [wired] (The check returns a report, not a core factual endorsement or automatic write veto.); Processed envelope keys at later archive replay accounting: known — routing [wired] (Counting duplicates is an operational branch, not knowledge acceptance or semantic deduplication.); Other host child-note readers, provider state, schema observers and evaluator-to-product consumers: uninspected (No authority is inferred for unconsumed metadata or untraced external observers/child notes. Optional ranking and other note consumers are incomplete.)	AAS-2026-10-16-basic-memory-46fe4c6cfdc3-01
beads_rust	2	Retained issue prose and symbolic work state; mutable ancestor agent_context; local mutation events, interaction audit entries, scoped gate history and saved filters; JSONL transfer and recovery generations; installed mutable project instruction artifacts. Includes unresolved auxiliary capacity/acceptance evidence, recovery consumers and non-update instruction-distribution variants. Excludes the static shipped instruction string as accumulated memory, executable upgrades, external host conversations/weights and Agent Mail internals. No opaque external provider store is attributed to br.	partial coverage (Known retained issue, local-evidence and file stores coexist with unresolved auxiliary/distribution/recovery coverage.); Issue prose, symbolic issue state, ancestry context and local event/gate/filter stores: known — sqlite [wired] (SQLite holds these parts; JSONL transfers issue state but not the complete local evidence store.); Issue JSONL, interaction log, recovery snapshots and mutable installed agent instruction files: known — files [wired] (These are retained files. JSON/JSONL encoding alone is not a graph or database substrate.); Other included instruction-distribution, auxiliary-policy/evidence and recovery parts: uninspected (The accepted account leaves non-update installation variants and auxiliary/recovery branches unaudited. Their classifications cannot be supplied by the inspected witnesses.)	partial coverage (Automatic bookkeeping and maintenance are established; generic caller identity cannot decide manual versus automatic content admission.); Mutation-event generation, automatic pre-export backup and configured rotation/omission: known — automatic [wired] (Software generates or selects these retained changes without per-content operator approval.); Create/update/context, import, reports, audit labels, saved queries and instruction replacement admission: not-determinable (Callers may be humans or software. Confirmation/force alternatives do not prove an explicit human content decision on any exercised path.); Other included instruction-distribution, auxiliary-policy/evidence and recovery parts: uninspected (The accepted account leaves non-update installation variants and auxiliary/recovery branches unaudited. Their classifications cannot be supplied by the inspected witnesses.)	partial coverage (Requested returns coexist with automatic selection of ancestor guidance. Included unaudited consumers prevent an exhaustive direction inventory.); Requested issue/event/gate reads, ready/scheduler, saved-query results, MCP prompts and history restore: known — pull [wired] (Returning requested retained material is pull even when code filters or projects it.); Opt-in automatic ancestor enrichment during child show or text start-work update: known — push [wired] (The child request is upstream; code independently selects additional retained ancestor instructions, not merely the requested child fields.); Future installed-file instruction loading and other unaudited consumers: uninspected (The host load/selection protocol, audit semantic consumers and other distribution variants are not established. No pull-only global conclusion follows.)	partial coverage (An identifier-selected ancestor enrichment is established; host instruction loading and other unresolved delivery alternatives remain unclassified.); Ancestry-selected ancestor agent_context blocks: known — identifier [wired] (Trigger: opted-in child show or eligible text start-work transition. Selector input: child ancestry identities and source eligibility. Selected part: at most root and immediate-parent governing blocks.); Explicitly requested state returns: inapplicable (These requested returns are pull; their IDs/filters do not independently establish a push signal.); Host project-file instruction load and unresolved delivery alternatives: uninspected (No complete automatic selector contract is recorded for these included alternatives.)	not-determinable coverage (No accepted bounded-absence record establishes a controlled no, and no recorded complete automatic trace-to-durable-behavior-shaping chain establishes yes. Raw event retention is insufficient.); Mutation events and explicit interaction/coordination records: not-determinable (Raw trace recording and review labels are established, but a qualifying automatic derived artifact and later semantic consumer are not. No system-wide negative is inferred.); Instruction replacement, candidate compaction and other maintenance/evidence routes: not-determinable (Versioned replacement is not shown to consume experience; compaction terminology is not a trace summarizer. Remaining branches prevent complete negative coverage.); Other included instruction-distribution, auxiliary-policy/evidence and recovery parts: uninspected (The accepted account leaves non-update installation variants and auxiliary/recovery branches unaudited. Their classifications cannot be supplied by the inspected witnesses.)	partial coverage (Authority follows actual retained-part consumers, not all permissions of the runtime. No learning authority is inferred from retaining traces.); Issue prose, comments, events, audit review material and gate notes at later readers: partial — knowledge [wired], knowledge [afforded] (Detail/event displays supply context. Audit review is afforded, but its semantic planning consumer remains uninspected.); Retained ancestor instructions at the child-output reader: known — instruction [wired] (Directions are delivered, not enforced against the external agent.); Installed mutable project instructions at a future host instruction load: known — instruction [afforded] (The write is wired; downstream host loading is an affordance, not observed compliance.); Retained issue status, priority and relations at readiness/scheduler consumers: known — ranking [wired], routing [wired] (Fixed software ranks candidates using retained inputs; no learned ranking policy is required.); Saved filters at the later list-query consumer: known — routing [wired] (The stored selector changes query inputs, rather than supplying a prior result snapshot.); Scoped gate verdicts at gate checking and guarded transition admission: known — validation [wired], enforcement [wired] (The sufficiency rule is validation and its rejection consumer enforces a veto on covered non-bypass transitions.); Retained interaction IDs at label admission: known — validation [wired], enforcement [wired] (Identity existence is a narrow admissibility criterion, not semantic review of the parent.); Other included instruction-distribution, auxiliary-policy/evidence and recovery parts: uninspected (The accepted account leaves non-update installation variants and auxiliary/recovery branches unaudited. Their classifications cannot be supplied by the inspected witnesses.)	AAS-2026-10-16-beads-rust-e82d5506d2ff-01
beever-atlas	2	Included: accumulated normalized messages; extracted facts/provenance/vectors and graph registry; topic/channel summaries; wiki content, cache, archives and curation controls; HTTP and reused MCP dialogue; retained ingestion checkpoints; use-modified model routing and retained evaluation/feedback state. Included unresolved alternatives: import/media origins, full compiler/modules/auto-builder, reconsolidation, migration and alternate adapters, dependency-owned session transformations and diagnostic consumers. Excluded: static shipped prompts, ordinary current-invocation working state, external clients' memory policies and provider-owned model weights/internal training.	partial coverage (The inventory covers messages, derived knowledge, controls, conversation/recovery state and diagnostics. Positive substrate witnesses do not establish backend or archive completeness.); Normalized channel messages, HTTP chat history, wiki page/cache/version objects and ingestion recovery checkpoints: known — service-object [wired] (The accepted records identify MongoDB persistence for these operative parts, including the memory annotation on checkpointed runtime state.); Retained fact prose, fact metadata, text vectors, topic/channel summaries and searchable QA history: known — vector [wired] (Classification follows the inspected vector-backed retained parts.); Persisted graph entities and relationships: known — graph [wired] (Graph persistence is established; adapter parity remains a separate unresolved alternative.); MCP conversation state reused by a later turn: known — in-memory [wired] (Only later-turn session reuse is included; ordinary current-run state is excluded.); Persisted model routing assignments and drift reports: known — service-object [wired] (Configuration accumulated through use and retained diagnostic objects are included, not model weights.); Entity-registry backing details, alternate storage adapters, migration snapshots and all archived/shared copies: uninspected (The records locate these included parts but do not classify every backing object or backend alternative. No extra substrate is inferred.)	partial coverage (Automatic derived-content and host-retention mechanisms are supported. Initial/caller-origin decisions and incompletely inspected writers prevent complete agency classification.); Derived fact/graph admission, summary retention, incremental wiki updates and generated full-build persistence: partial — automatic [wired] (Initiating sync/refresh is not per-content human approval; full compiler branch coverage remains incomplete.); HTTP dialogue/answer persistence and selected-stage checkpoint saves: known — automatic [wired] (Known host-owned writes are separate from the human or computational origin of user messages.); Fact supersession and drift-report storage: known — automatic [wired] (Software admission is established even though actual execution is unobserved.); Initial raw acquisition, file upload/commit, curation/structural changes, purge and model-assignment requests: not-determinable (Explicit caller selection is established, but caller identity is not enough to decide human versus software admission. No manual value is inferred.); Remaining compiler/auto-builder, embedding-migration and retry-provider admission variants: uninspected (Inspected automatic witnesses do not resolve all alternative admission mechanisms.)	partial coverage (Pull tool fulfillment coexists with automatic history, maintenance, summary and recovery supply. Included dependency/compiler alternatives retain unresolved coverage.); Requested facts, provenance, summaries, wiki pages and past QA through QA tools/API/MCP: known — pull [wired] (The inspected requested tool family is resolved; granular selector internals remain a separate included limitation.); HTTP prior-dialogue assembly into a fresh QA turn: known — push [wired] (An upstream supplied session identity determines the automatic selection input; only content from up to ten stored messages is replayed.); Event-driven incremental wiki completion input: known — push [wired] (Push is into the maintenance consumer, not unsolicited knowledge injection into every future QA turn.); Summary context production from dirty/touched clusters: known — push [wired] (The context builder caps fetched members and selected facts before supplying the model.); Checkpoint restoration into a fresh retry session: known — push [wired] (Known processor retries do not establish cross-worker-job checkpoint reuse.); MCP dependency-owned context reuse/compaction details, granular tool payload routes, full compiler/module and migration consumers: uninspected (Known requested and automatic deliveries do not completely resolve every included later-consumer operation.)	partial coverage (Signals classify actual automatic supply operations, not identifiers or embeddings used only in pull retrieval. Included alternatives prevent exhaustive push-selector coverage.); HTTP history push on a new turn: known — identifier [wired], coarse [wired] (Session identity actually selects delivered dialogue; the tail budget is a distinct coarse selection operation.); Incremental maintainer event push: known — identifier [wired], coarse [wired] (Identifier matching and coarse eligibility are separate selector contributions; routing implementation is not inferred semantic judgment.); Summary-model context push: partial — identifier [wired], coarse [wired] (Graph enrichment is included in the route, but a complete targeting inventory for every enrichment branch is unresolved.); Checkpoint restoration push: known — identifier [wired] (The memory and workflow annotations establish matching and restoration, not cross-tick equivalence.); Full compiler/auto-builder, granular delivery alternatives and dependency-managed MCP session selection: uninspected (Untraced automatic delivery/selection alternatives cannot be made inapplicable by a known pull route. Query embeddings and lexical MMR on requested reads are not push signals.)	partial coverage (Positive automatic trace-to-durable-artifact-to-later-consumer chains coexist with unresolved writes. No system-wide no or capacity-improvement claim is made.); Retained conversation messages through automatic fact/graph extraction and later summary/wiki/QA consumers: partial — yes [wired] (The recorded conversation-derived branch qualifies without weight training or demonstrated improvement. Non-conversation imported alternatives retain source uncertainty.); HTTP recorded question/context through generated retained answer and later QA context/search: known — yes [wired] (This finding concerns the generated durable answer artifact, not raw log retention alone. Stored tool/thinking fields are not all replayed.); Other import/media content origins, alternative full compilers/checkpoints and feedback/evaluation-driven updates: not-determinable (No complete bounded negative is supported. Unresolved sources or downstream update consumers prevent classifying all included writes as qualifying or nonqualifying.)	partial coverage (Separate retained parts and actual consumers support knowledge, learning, ranking, enforcement, validation and routing. Static shipped instructions are outside scope; no instruction finding is inferred from delivery alone.); Message/fact/graph/summary/wiki content at extraction, consolidation, wiki generation and QA/human consumers: known — knowledge [wired] (Knowledge authority is advisory/contextual; it does not establish source truth or model compliance.); Retained conversation evidence and derived facts at automatic fact/summary/wiki update consumers: partial — learning [wired] (Learning authority is at the update consumer; downstream output knowledge authority is a different path. No capacity improvement is claimed.); Fact prose/vectors, name vectors and fact quality scores at retrieval, dedup candidate selection and summary context ordering: known — ranking [wired] (Fixed ranking code still gives retained scoring inputs ranking authority.); Fact invalidation and supersession fields at guarded retrieval/consolidation consumers: known — enforcement [wired] (BM25 and explicit superseded/ID reads remain counterexamples to universal exclusion.); Retained page curation controls at incremental maintainer: known — enforcement [wired], routing [wired] (These forces apply at the incremental consumer, not all full-refresh writers.); Retained rejected entity pairs at cosine dedup admission: known — validation [wired], enforcement [wired] (Exact-name collisions bypass this registry lookup; the veto is branch-specific.); Checkpoint stage identity, persisted model assignments and message extraction status at their later executors: known — routing [wired] (This is selection of operations/input windows, not retained natural-language instruction.); Retained HTTP dialogue and searchable prior answers at subsequent QA model: known — knowledge [wired] (Only dialogue content is reconstructed by the HTTP loader; stored thinking/tool traces are not all supplied.); Feedback/drift downstream adoption, complete module payload consumers, and dependency-managed session transformations: uninspected (Retained reports do not establish instruction, learning or other force at an untraced later consumer. Full consumer inventory is unresolved.)	AAS-2026-10-15-beever-atlas-10a3512c094c-01
binder	2	Retained journal observations, summaries, scores, goals and plans; task findings, specifications, criteria, status and links; mutable schema/view/navigation configuration; transaction, redo and repair history; and selected disk projections with synchronization metadata. Includes their shipped capture, import, edit, maintenance, request/read and subscribed-delivery paths. Opaque journal data-sync injection and incompletely recorded workflow/app/browser alternatives remain unresolved where relevant. Excludes shipped static skills as accumulated memory, ephemeral importer proposals, fresh completion contexts and process-local caches as independent memory stores. External host/provider memory, weights and unshipped extension implementations are outside this boundary; shipped extension delivery and its limits remain included.	partial coverage (SQLite and files are wired; opaque injection and incompletely assessed consumer alternatives prevent complete inventory coverage. Provider storage and model weights are excluded, not classified as absent.); Retained graph records, mutable config and transaction state: known — sqlite [wired] (Covers the recorded database container and its journal/task/config parts; transient caches are not separate accumulated objects.); Rendered documents, snapshot metadata, transaction logs, backups and redo staging: known — files [wired], sqlite [wired] (Separates retained disk content from its database access metadata.); Journal extra data-sync payload before graph admission and unassessed app/browser retained alternatives: not-determinable (The records name injected health/calendar/activity context and uninspected frontend details but do not establish their retained-object inventory or storage before delivery.)	partial coverage (Manual and automatic mechanisms coexist; generic caller writes and original/injected content remain unresolved rather than inheriting either classification.); Explicitly supplied day-log content and user-approved journal summary/planning, task capture/refinement/design and import mapping/application: known — manual [claimed] (Approval is prompt policy, not a universal graph write invariant.); Day-update compact conversation capture, day-start post-answer capture and investigation findings without separate content approval: known — automatic [claimed] (These are instructed agent writes, not observed automatic operation.); Structural journal initialization, document/metadata generation and software-selected history reconciliation/staging: known — automatic [wired] (The initializer's executable structural writes are separated from semantic journal prompt policy.); Generic CLI/MCP/HTTP/library input, disk authoring, initial live content and opaque injected-data admission: not-determinable (Parsers and software I/O do not resolve whether a human decided the admitted content. Caller identity, disk authorship and live initialization are not established.)	partial coverage (Both pull and non-model push routes exist; unresolved injection prevents calling the whole memory boundary pull-only or completely inventoried.); Requested graph reads/searches and named journal/task later-use commands: known — pull [wired], pull [claimed] (Requested skill-time context and explicit rereads establish intended later consumers, not always-on injection.); Commit/subscription delivery and automatic document projection for later consumers: known — push [wired], push [wired], pull [wired] (These pushes are to extension/filesystem/editor consumers, not an established automatic model-memory injector.); Extra journal data-sync delivery and incompletely recorded app/browser consumers: not-determinable (Payload producer, request boundary and selector wiring are not established for the extra injection or every consumer alternative.)	partial coverage (Coarse push selection is supported locally; opaque injection and bounded projection inspection prevent complete signal coverage. Requested identifiers establish no identifier-based push.); Subscribed transaction push to configured hooks: known — coarse [wired] (Covers the recorded hook supply operation, not the unknown extension's subsequent selection or retention.); Automatic graph-to-document projection selection: partial — coarse [wired] (Recorded config/filter selection supports coarse delivery. Full query/slot algorithms are not recorded; paths and entity UIDs in metadata alone do not establish identifier-matched targeted push.); Requested journal/task/graph read operations: inapplicable (These named operations fulfill requests, so their keys/filters are pull selectors, not push-signal findings.); Opaque extra journal injection and other unassessed delivery alternatives: not-determinable (No recorded trigger/input-to-selected-payload chain classifies an additional signal. Unknown direction is not pull-only inapplicability.)	partial coverage (Yes is supported at the claimed layer for automatic conversational capture and investigation retention. Manual/structural routes and unresolved alternatives remain local; no observed or causal learning benefit is asserted.); Automatic agent conversation capture in day-update/day-start followed by later journal review/planning: known — yes [claimed] (Only the named automatic capture mechanisms are classified; human-approved day/period summaries do not inherit their write agency.); Automatic investigation findings written after reproduction/tracing and consumed on task reread: known — yes [claimed] (Doctrine establishes the chain as claimed, not execution, causal correctness or improved future capacity.); User-approved summaries, plans, task design and direct user-supplied content: inapplicable (These explicitly manual admission mechanisms do not meet automatic trace-fed write conditions, even when trace inputs inform human decisions. This is not a system-wide no.); Structural initialization, projection and recovery plumbing as described: inapplicable (Recorded automatic projection/restore/initialization does not derive a new behavior-shaping artifact from experience; raw recovery traces and fixed structural copies alone do not qualify.); Generic caller/disk admissions, opaque injected input and unassessed remaining transformation alternatives: not-determinable (Content control, trace-fed production or later-consumer chains are unresolved for these alternatives; no bounded absence record licenses a global no.)	partial coverage (Knowledge, instruction, learning, routing, validation and enforcement are locally supported, with claimed semantic and wired structural force kept separate. Unresolved consumers prevent a complete-value assertion.); Retained journal observations/summaries/scores and task diagnoses at briefing, review and implementation consumers: known — knowledge [claimed], learning [claimed] (Separates context at the later work consumer from experience feeding durable behavior-shaping plans at the update consumer.); Retained journal goals/plans and task specifications/criteria at planning and product-work consumers: known — instruction [claimed], validation [claimed] (The criteria's completion role is separate from the specification's directive force; neither is a runtime-enforced release gate.); Retained task status, field presence and links at guide selection and duplicate screening: known — routing [claimed], validation [claimed] (Prompt screening has no demonstrated exhaustive duplicate oracle or mechanically enforced semantic veto.); Mutable schema/config at structural admission and view/navigation projection consumers: known — validation [wired], enforcement [wired], routing [wired] (Fresh preparation and projection are covered; prepared replay is not assigned fresh schema validation.); Snapshot metadata and retained history at synchronization, redo and repair consumers: known — routing [wired], validation [wired], enforcement [wired] (Assigns force only to consumed metadata/history, not to unused author labels or absent rationale.); Injected journal payload and incompletely assessed app/browser or remaining guide effects: not-determinable (Records do not inventory every selected part and consumer effect; no ranking authority is inferred merely from search, scores or ordered results.)	AAS-2026-10-15-binder-a35c41531d40-01
browzy.ai	2	Material accumulated or changed through use: raw sources and images, wiki articles, SQLite source/search state, JSON index/backlinks, saved and live session turns reaching later calls, digest files and metadata, drafts, activity records, query/HTML caches, saved outputs and owner-edited schema. Includes their admission, maintenance and later-consumer paths. Provider-side use-dependent retention and personalization/history/autocomplete/cost state remain unresolved included alternatives. Excludes unchanged shipped prompts, comment-only schema defaults, bundled examples before admission, credentials, general runtime configuration, and a discarded compaction return. Static inference weights are not accumulated memory; possible provider-side changes through use are unresolved, not classified from the inference component alone.	partial coverage (Recorded local stores are classified; unresolved included alternatives prevent complete inventory coverage.); Disk raw content/images, wiki Markdown, JSON manifest/index, session JSON, digests, drafts, logs, outputs and edited schema: known — files [wired] (Covers the recorded disk inventory, including non-reinjected drafts and outputs.); SQLite sources, article copies and FTS5 access tables: known — sqlite [wired], rdbms [wired] (FTS is lexical indexing, not vector storage.); Live session turns reused on later calls and process-local answer/HTML caches: known — in-memory [wired], kv [wired] (Volatile retention is not disk persistence.); Use-dependent provider retention and uninspected personalization/history/autocomplete/cost state: uninspected (Provider internals and the named host internals are not resolved by accepted records; their accumulated objects and substrates cannot be enumerated.)	partial coverage (Known manual schema editing coexists with automatic host admission and unresolved other controls.); Raw conversion/admission, parser/template wiki writes, metadata maintenance, host session/log/cache writes, digest/draft generation and demo seeding: known — automatic [wired] (Partial skip is not passed as a compiler veto; a clipboard attempt need not successfully admit arbitrary strings.); Owner edits/replaces schema content: known — manual [wired] (Excludes unchanged generated defaults; human-selected reset is replacement, not model self-revision.); Generic exported storage callers and initial external/bundled content production: not-determinable (Storage capability and imported authorship do not identify human versus software control at unspecified admission mechanisms. Known seeding remains automatic.); Provider-side use-dependent writes and host personalization/history/autocomplete/cost updates: uninspected (These controls were not established in accepted records.)	partial coverage (Both operation directions exist. Unestablished alternative read-back paths remain included rather than declared absent.); Requested article retrieval, human search/log display, identical-key answer reuse and requested-URL HTML reuse: known — pull [wired] (These request operations coexist with independent context-supply operations below.); Host-selected model context/history/schema, startup digest/banner input and post-answer source prefixes supplied to crystallizer: known — push [wired] (Question-driven retrieval is pull at the retrieval API and automatic supply at the answer model; startup digest display is a distinct human push.); Compilation request and host-selected raw/existing-article inputs supplied to the compiler model: known — pull [wired], push [wired] (Request fulfillment at storage and automatic supply at inference are separate operations in the same chain.); Draft/image/exported-output later consumers, optional digest query admission and provider/host opaque alternatives: not-determinable (Retention/capability alone supplies no direction; incomplete digest indexing and unresolved host/provider routes prevent complete coverage.)	partial coverage (Each positive belongs to an actual automatic supply operation; request keys on pull-only cache reads do not add identifier push.); Automatic wiki subset supplied to answer inference after a question: known — inferred-lexical [wired], coarse [wired] (The ineffective matched-section branch does not erase the actual oversized-article selection route.); History and schema supplied automatically on applicable inference calls: known — coarse [wired] (A schema filename is not a targeted identity match among delivered memories.); Startup prior-session/digest selection and post-answer crystallizer article supply: known — identifier [wired], coarse [wired] (Model novelty judgment disposes a new candidate; it is not an inferred-judgment selector of retained memory delivered on these paths.); Compiler-model input supply on requested compilation: known — identifier [wired], coarse [wired] (Identity-based eligibility and coarse prefix/batch selection are distinct; source IDs alone would not establish targeted supply without the membership predicate.); Provider/host opaque push alternatives and other unestablished later consumers: not-determinable (Missing directions/selectors prevent a pull-only or complete signal conclusion for these included parts.)	partial coverage (A positive wired digest route coexists with unresolved other routes. Missing later consumers are limits, not canonical evidenced absences; the positive union cannot imply complete coverage or measured learning.); Automatic saved-session digest write followed by startup human-banner context: known — yes [wired] (Wiring is qualified by caught runtime-module failures and incomplete wiki indexing; success and improvement are unobserved.); Answer-derived draft write in the inspected terminal/crystallizer route: not-determinable (The scoped search establishes a writer without an established later adoption/retrieval consumer. Proposal retention alone cannot establish yes; no canonical bounded-absence record warrants no. Human promotion and external hosts remain unresolved.); Activity-log/system-notice feedback to inspected history/digest consumers: not-determinable (Display consumers and filtering of system notices are established, but no automatic durable corrective update is established. Without a canonical bounded-absence record these limits do not warrant no.); Provider-side updates, other imported trace content and uninspected personalization/feedback writes: uninspected (No supplied corpus or inspected internal write chain resolves additional qualifying routes; absence of adapter training calls is not system-wide absence.)	partial coverage (Authority is local to actual consumers; parser gates and static linter code are not retained validation criteria. Unknown consumers prevent a complete authority inventory.); Raw text to compiler; wiki text/history/schema to models; search snippets/cache answers/digest banner/activity display to people: known — knowledge [wired], instruction [wired] (Actual consumer paths are established, including human-only digest display; this does not attribute authority to unconsumed draft metadata.); Article text/tags/recency/backlinks and FTS copies at retrieval scoring: known — ranking [wired] (Ranking is not factual validation.); Source membership/manifest, cache keys/generation/times, latest-session metadata and demo marker at code selectors: known — routing [wired], enforcement [wired] (These are operational controls, not authorization principals or source-truth criteria.); Saved conversation at automatic digest transformation and retained question/answer at draft proposal generation: partial — learning [wired] (Draft production lacks an established later behavior-shaping consumer; optional digest-to-answer admission is incomplete. Neither establishes improved capacity.); Unconsumed draft/image/export parts, provider state and uninspected host personalization/feedback consumers: not-determinable (No consequential consumer is established for the first parts; provider/host internals are uninspected. No authority value follows merely from retention.)	AAS-2026-10-15-browzy-ai-0a823a13a7b6-01
ByteRover CLI	2	Included: accumulated local topic bodies and selection metadata; derived summaries, abstracts, overviews and manifest; query caches; runtime signals; archive/review/recovery state; retained compaction summaries and task-sandbox checkpoints; imported bundle/skill admission and unresolved registered alternative persistence/consumer routes. Excluded: static shipped workflow/schema/hook text, ordinary unretained protocol state, provider weights and internals, Cloud server storage, excluded external host and dependency implementations. Installation and caller adapters remain included without assuming host activation.	partial coverage (File, KV and process-local witnesses coexist. Unresolved backing and alternative state routes prevent a complete substrate inventory; provider weights and remote Cloud internals are excluded, not classified as accumulated local memory.); Live topic bodies and selection markup: known — files [wired] (Topics are local HTML/markdown files; markup is encoding, not a graph substrate.); Hierarchical summaries, abstract/overview siblings, manifest and archive cue/full/dream files: known — files [wired] (These retained parts use filesystem files. The archive service API does not establish an ordinary automatic archive cycle.); Runtime signal sidecar: logical store: known — kv [wired] (Signals use composite keys in IKeyStorage.); Runtime signal sidecar: daemon backing: known — files [wired] (Daemon construction injects FileKeyStorage; logical KV and physical files describe separate layers of the same object.); Process-local query result caches: known — in-memory [wired] (Two bounded process-local caches retain response/envelope strings for later requests.); Rolling checkpoint variable within the task sandbox: known — in-memory [wired] (Progress JSON survives history clear in the task sandbox, not demonstrated process restart or task deletion.); Full compaction boundary logical message store: known — kv [wired] (StoredMessage retains summary and boundary fields in message key storage; no physical backend classification is inferred here.); Compaction message-store physical backing and complete recovery/log storage inventory: not-determinable (Key-storage use and saved recovery records do not settle all physical backing or log persistence details.); Other registered ingest/map/swarm persistence, version/worktree/migration replacements, alternate compression and downstream log consumers: uninspected (Accepted records do not close these alternate write, replacement and consumer branches. Their unresolved object inventory prevents a complete boundary-wide classification; no absence is inferred.)	partial coverage (Automatic admissions are positive witnesses, not a statement that humans never author memory. Explicit human control and several alternative write mechanisms remain unresolved, preventing manual or automatic-only completeness.); Daemon topic admission after CLI/MCP continuation and internal generation: known — automatic [wired] (Software/model admission requires no per-content human decision; an initiating request or overwrite flag is not human semantic approval.); Human control of caller-authored/direct-editor content before admission: not-determinable (Generic caller/client identity does not establish explicit human supplying, editing or approval; no manual finding follows.); Signal updates, compaction boundaries and rolling checkpoints: known — automatic [wired] (Code/model constructs and admits derived material without per-content operator approval.); Cache insertion, manifest rebuild and inspected summary/abstract queue writes: known — automatic [wired] (Software/model generation or assembly admits these derived artifacts; not every current HTML write is shown to invoke every generator.); Hub bundle/skill installation and Cloud snapshot admission: known — automatic [wired] (Selecting an install/sync workflow is not per-file content approval; the service admits downloaded/snapshot content under guards.); Review replacement, dream archive decisions and initial archive-service content admission: not-determinable (Reviewer/caller decisions are explicit but principal identity is unresolved; archive writer capacity alone does not identify its ordinary admission owner.); Other registered ingest/map/swarm persistence, version/worktree/migration replacements, alternate compression and downstream log consumers: uninspected (Accepted records do not close these alternate write, replacement and consumer branches. Their unresolved object inventory prevents a complete boundary-wide classification; no absence is inferred.)	partial coverage (Pull requests coexist with automatic context/metadata supply. Operations within each chain are distinguished; the full-summary wording conflict and alternative delivery inventory prevent known coverage.); Requested public query/cache, ranked search, focused topic read and archive/overview expansion returns: known — pull [wired] (Consumers request retained material. Code selecting and delivering the requested response is not a second push.); Broad manifest context automatically selected inside internal requested query: known — push [wired] (The query requests execution, while automatic lane selection supplies broad retained context without a separate request for those parts. The resolver ignores query text; this is not unsolicited execution.); Retained signals automatically loaded into later search/prune selectors: known — push [wired] (Selector code supplies persisted numerical metadata to later ranking/maintenance consumers without a separate request for those signal records.); Full compaction summary supplied to subsequent model invocation: partial — push [wired] (Message-loader implementation includes the retained boundary automatically. The runtime wording conflicts about exactly what reload contains, so this is a local memory-record witness, not a resolved complete reload account.); Rolling checkpoint payload retrieved through explicit code access: known — pull [wired] (Continuation requests the model read the variable; the JSON payload is not automatically dereferenced. Wired access does not establish observed reload.); Complete alternative compression, worker and host delivery inventory: uninspected (Other registered routes and excluded host placement prevent exhaustive direction coverage; a known pull branch cannot make alternatives inapplicable.)	partial coverage (Push signals classify their actual operations only. Lexical ranking on a requested query is not an inferred-lexical push witness. Known pull branches do not justify boundary-wide inapplicability.); Manifest context push: known — coarse [wired] (Trigger: internal query branch. Inputs: available paths/types, condensation order, default importance and lane budgets. Selected parts: summary/context/stub entries, preferring abstracts. No query relevance or actual identity-match targeting is established.); Signal metadata push into later search/prune: known — identifier [wired] (Trigger: later search/index acquisition or scan. Input: result/topic path. Composite signal-key lookup selects the corresponding retained metadata delivered to the selector; this is an actual identity match, not merely a catalog identifier.); Compaction-boundary context push: partial — coarse [wired] (Trigger: post-compaction reload/next invocation. Inputs: session traversal and boundary stop rule. Selected retained part: summary boundary with the retained tail. Exact cross-report reload wording is unresolved; no lexical or embedding relevance selector is established.); Explicit requested query/search/read/expansion and checkpoint payload returns: inapplicable (These named delivery operations are pull, so their query/path/variable selectors do not classify push signal. This local inapplicability says nothing about alternative push routes.); Alternative registered worker/compression and host push selectors: uninspected (No complete accepted inventory identifies each alternative automatic selector, its input and the retained part delivered.)	partial coverage (Wired trace-fed behavior-shaping artifacts establish yes without observed improvement. No system-wide or local no is inferred from uninspected writers; source and consumer coverage remain incomplete.); Returned-hit activity to persisted signals to later search/prune consumers: known — yes [wired] (Automatic usage-fed updates produce durable behavior-shaping ranking/maintenance metadata. Confirmed reading, factual acceptance and improved performance are not required or established.); Conversation history to retained full-compaction summary to later model context: partial — yes [wired] (The memory object records a wired trace-summary-later-invocation chain. Runtime reload wording conflicts; it does not erase summary retention or this local positive, but prevents complete route coverage.); Conversation/tool activity to sandbox progress to later explicit task-model read: known — yes [wired] (Automatic checkpoint retains counts, tool names and assistant snippets across history clear for continuation. Horizon is later task invocations, not demonstrated restart persistence or observed use.); Topic authoring/import/replacement and derived content provenance outside the three qualifying routes: not-determinable (These routes do not establish original trace input for each content write. No bounded absence record supports a local no; unknown provenance is not a negative.); Other registered ingest/map/swarm persistence, version/worktree/migration replacements, alternate compression and downstream log consumers: uninspected (Accepted records do not close these alternate write, replacement and consumer branches. Their unresolved object inventory prevents a complete boundary-wide classification; no absence is inferred.)	partial coverage (Authority belongs to actual part/consumer/effect paths. Static shipped schemas and hook instructions are outside accumulated-memory scope, so their validation/instruction effects are not classified here; unresolved consumers remain included.); Topic descriptive statements and reasons at query/read answering or update authors: known — knowledge [wired] (Returned prose supplies reference/context; delivery does not establish factual acceptance or actual reliance.); Retained rule/decision directions at later calling agents: known — instruction [wired] (Retained directions can guide later work through rendered text. Their force is advisory; no host enforcement is inferred.); Retained topic/summary/cue text scored by search: known — ranking [wired] (Retained content influences relative priority under deterministic lexical search; a retained learned ranking policy is unnecessary.); Runtime signal fields at later search and prune selectors: known — ranking [wired] (Importance/access/maturity influence scoring and candidate ordering. Validated is a tier label, not truth acceptance.); Retained usage traces at the signal-update consumer: known — learning [wired] (Access activity feeds durable behavior-shaping ranking metadata; no improved capacity or semantic knowledge revision is asserted.); Retained conversation/tool history at compaction and checkpoint builders: known — learning [wired] (Recorded experience feeds a durable continuation artifact for later model invocation. This authority is at the update, distinct from downstream context use.); Summary/checkpoint text at subsequent internal task model calls: known — knowledge [wired] (Summary/snippet content supplies continuation context. The full-summary delivery conflict and explicit checkpoint reload requirement remain local limits.); Manifest paths/lane entries at context resolver; archive pointers and boundary links at reload consumers: known — routing [wired] (Retained structure selects files, reload destinations or history windows. It is not all user-facing knowledge.); Summary/abstract/overview and archived prose at internal prompt or explicit expansion consumers: known — knowledge [wired] (Selected readable content supplies context; abstract preference and lossy overview are not full-source equivalence.); Saved review statuses at legacy Cloud export filter: known — enforcement [wired] (Pending/rejected status blocks selected exports. This does not gate local query or establish universal push policy.); Saved review/dream recovery state at rollback consumers: known — routing [wired] (Task/path, operation kind and backup availability choose restoration/deletion actions; factual reacceptance is not implied.); Installed hub skill files at external host loading/placement: not-determinable (Installation is wired but actual host loading and resulting authority are unresolved; standing instruction force is not upgraded from availability alone.); Other registered ingest/map/swarm persistence, version/worktree/migration replacements, alternate compression and downstream log consumers: uninspected (Accepted records do not close these alternate write, replacement and consumer branches. Their unresolved object inventory prevents a complete boundary-wide classification; no absence is inferred.)	AAS-2026-10-15-byterover-cli-71610aaf716b-01
cass_memory_system	2	CM's accumulated diaries, playbook content/reasons/standing, feedback and outcome/delivery logs, processed/onboarding checkpoints, embedding cache, withdrawal logs, trauma patterns/explanations and exported instruction snapshots; their admitted writes, maintenance and later CLI, HTTP, reflector, scorer and guard consumers. External cass history is included as an opaque retained input/read surface, not as CM-owned indexing. Imported/custom origin and generic caller control remain unresolved. Excludes static shipped method text, ordinary transient validator results, provider/model weights and hidden host/provider state whose retention is outside the supplied boundary; no deployed corpus, activation or benefit is established.	partial coverage (Files are established for CM-owned state; the included external historical input prevents complete substrate coverage.); CM playbooks, diary JSON, feedback/reasons, delivery/outcome logs, processed/onboarding state, cache, blocked logs, trauma JSONL and project snapshots: known — files [wired] (These records cover the named local containers; numerical vectors are encoded in a file, not an established vector database.); External cass historical corpus/index used by context and reflection: uninspected (The adapter supplies retained history, but the external index implementation and operative substrate are outside inspection; snippets do not establish that substrate.)	partial coverage (Automatic derived admission and manual batch approval coexist with unresolved caller-controlled and external branches.); Reflection diary/delta admission, inline/outcome grading, deterministic curation and embedding-cache adoption: known — automatic [wired] (Applies to the admitted derived content and metadata, independent of upstream authorship.); Automatic delivery logging, reflection recovery bookkeeping and outcome-derived event replay: known — automatic [wired] (Configuration or workflow initiation is not approval of each generated log/checkpoint/event.); Initializer trauma scan batch approval: known — manual [wired] (Manual admission is supported for this batch branch only; subsequent file I/O does not change its agency.); Generic single/batch add, onboarding caller, direct mark, import/starter replacement, explicit withdrawal/undo/scrub, project export and manual trauma add/import/heal/remove: not-determinable (Explicit caller selections/replace flags do not resolve whether a human supplied or approved the content versus a computational caller. Software persistence alone cannot classify that decision.); External cass history admission and custom/imported content initialization before CM acquisition: uninspected (Originating logs, index writers and external authors are not inspected; no upstream manual or automatic finding is licensed.)	partial coverage (Pull and push coexist at different operations; host adoption and opaque external chains prevent complete boundary enumeration.); Requested task context, why/get/resource/search and portable/project export returns: known — pull [wired] (Known requested operations; each reader has its own projection/budget, not context's universal guarantees.); Reflection's automatic selection/supply of diary, existing bullets and related history to the update model: known — push [wired] (This is an update-consumer push, not unsolicited injection into every enclosing coding-agent session.); Automatic guard/task-warning reads and internal checkpoint, attribution, cache, blocked/active-state reload consumers: known — push [wired] (Requested high-level operations can contain automatic control supply; their requested terminal result remains pull.); Exported instruction snapshot at a later external host: not-determinable (Agent-file purpose affords later guidance, but no recorded host loader, trigger or request/selection/delivery relation establishes pull or push at that later consumer. The separately recorded requested export remains pull; its direction does not resolve the later host read.); External corpus-internal and unspecified opaque consumer alternatives: uninspected (CM request/automatic-supply operations are classifiable; the external index and hosts do not supply complete read-chain coverage.)	partial coverage (Coarse, identifier and lexical push have actual automatic consumers; unresolved related-history and external host selectors keep coverage partial rather than inapplicable.); Reflection trigger to selected diary/existing-playbook prompt fields: known — coarse [wired] (Scope covers inspected prompt supply, not the opaque selector for every related-history snippet.); Related-history selection into reflection and exported instruction snapshot into a host: not-determinable (At most three related snippets and an afforded agent-file consumer are recorded, but complete selector inputs and host loading conditions do not establish a controlled signal for these alternatives.); Outcome application trigger to latest retained delivery log and existing rule targets; semantic-cache reuse trigger to retained vectors: known — identifier [wired] (Identifiers act on retained parts delivered to software consumers, not merely as catalog labels.); Discovery/retry and loader triggers to progress/active/blocked control parts: known — coarse [wired], identifier [wired], inferred-lexical [wired] (Coarse availability and exact/token targeting remain separate mechanisms.); Pre-tool/Git or CLI task trigger to active trauma controls and matching explanation: known — coarse [wired], inferred-lexical [wired] (CLI warnings and host denial differ in force; fixed regex selectors establish neither host activation nor complete tool coverage.); Requested task-context lexical/embedding ranking and requested diagnostic IDs/resources: inapplicable (These selectors fulfill pull requests; they do not establish inferred-embedding or identifier push merely because an automatic reader computes the return.)	partial coverage (Two automatic durable trace-fed chains establish yes; unresolved caller/external alternatives prevent complete route coverage. No local or global no is asserted without a bounded absence warrant.); Automatic session export to retained diary/rule/reason to next reflector or requested context: known — yes [wired] (Includes reduced LLM-free diaries when later supplied as context; the no-LLM reflector's lack of new deltas does not negate the retained-summary route.); Inline/session-outcome grading to durable feedback/standing to later scoring and selection: known — yes [wired] (Raw delivery-log retention alone is insufficient; the subsequent durable event/standing update supplies the qualifying result.); Onboarding external extraction, explicit caller reports/maintenance/import and approval-gated trauma scan: not-determinable (Generic caller identity and external content provenance do not establish automatic trace-fed admission. The inspected trauma scan requires batch human approval, so it does not itself supply an automatic trace-learning yes; there is no bounded absence record licensing a system-wide no.); External history production and opaque provider/host alternatives: uninspected (External producers and hidden retention are outside inspected implementations; no additional learning route or negative can be inferred.)	partial coverage (All seven forces have named local consumers; inventory/granularity limits prevent complete coverage, not the source-wired positive paths.); Rule words at later context recipient, next reflector and exported agent-file consumer: known — instruction [wired], instruction [afforded] (Distinct delivery strengths remain distinct findings; absence of compliance evidence does not erase instructional force.); Rule content/vectors and standing at context ranking and duplicate selection: known — ranking [wired], validation [wired], enforcement [wired] (Named fixed consumers establish priority and duplicate controls, not rule truth.); Diary/session experience and outcome/delivery attribution at durable rule or standing update consumers: known — learning [wired], knowledge [wired] (Update-consumer learning and recipient knowledge are separate effects, not inferred from each other.); Reasoning, feedback explanations, withdrawn-rule diagnostics and trauma human_message at display recipients: known — knowledge [wired] (No blanket instructional or enforcing force is assigned to explanation text; stale/omitted reasoning is not evidence of iterative criticism.); Processed/onboarding checkpoint state and outcome session/rule references at next-work selection: known — routing [wired] (Control over the next input/target is not content advice.); Retired/blocked rule state at merged loaders and active selectors: known — enforcement [wired], validation [wired] (Eligibility checks and exclusion coexist; independent exports and legacy copies constrain completeness of withdrawal.); Trauma regex/status at generated host/Git matcher and CLI task-warning matcher: known — validation [wired], enforcement [wired], knowledge [wired] (Actual consumer paths are known at source level, with host effect explicitly uninspected.); Undeclared diary-status/validator linguistic parts and opaque external history consumers beyond recorded projections: not-determinable (The recorded granularity and hidden corpus do not establish independent authority paths for every part; current-run validator permission is not automatically retained memory authority.)	AAS-2026-10-13-cass-memory-system-30c18219cf68-01
Claude Context Guard	2	Included: use-derived/current project handoff, session history, tasks, user comments, decisions, lessons, QA state, archived plans, audit reports, code markers/sidecars, approved helper artifacts, their archive/access metadata, mutable project guidance and declared custom-context inputs; PreCompact file copies and the explicitly opaque host compaction retained summary/tail alternative. Excluded: shipped static commands/templates/settings as independent memory, current-run transient tool returns and worker frames without retained later consumption, host model weights/provider implementation, application code except retained navigation/helper artifacts, and comparison images. Optional Git persistence is included; generated deployments and opaque alternatives remain unresolved where records cannot classify them.	partial coverage (Named local retained objects are covered, including optional Git and wired backups; opaque host compaction prevents complete substrate coverage.); Generated ledgers, handoff, archive pages/indexes, plans, audits, markers, sidecars and lesson helpers: known — files [claimed], repo [claimed] (The records enumerate filesystem objects and optional project Git persistence. This classifies the specified storage, not successful writes.); PreCompact copies: known — files [wired] (The inspected shell copies existing named files to a timestamped directory; no restoration is inferred.); Use-modified CLAUDE instructions and declared custom files: known — files [claimed] (The supplied interface names local instruction and custom-file paths; unavailable content does not obscure that file-based interface.); Host compaction retained summary/tail and its later delivery: uninspected (The records explicitly leave host compaction state and processing uninspected; The retained payload substrate cannot be classified. This is not absence and does not include the host model weights.)	partial coverage (Automatic model/shell admission and explicit manual content gates remain distinct; generic custom-file provenance and opaque host compaction do not inherit either value.); Ordinary incremental/checkpoint logging, decisions, LB, handoff and clean-end replacement: known — automatic [claimed] (Starting /save or /end is not per-content human approval; clean-marker replacement has its own model admission, not just a subsequent read.); Sidecar/marker row admission, description maintenance, audit-report retention and plan-copy attribution: known — automatic [claimed] (Human itemisation scope approval is not approval of every label/description; audit-report creation precedes human repair choices.); Human QA status and lesson installation; chosen audit repairs: known — manual [claimed] (Software physical writes are separate from explicit human decisions on verification, proposals and selected fixes.); First-run user details, Git mode and deliberately human-authored sidecar descriptions: known — manual [claimed] (This unit is limited to explicitly human-supplied/edited content, not all template initialization or custom-file creation.); PreCompact file-copy admission: known — automatic [wired] (The fixed-list existence selector determines copying without per-copy approval.); Declared consumer custom-context file bodies: not-determinable (The declaration/read interface was inspected, but bodies and their histories were not supplied; Who supplied, edited or approved the body cannot be determined from the declaration alone.); Host compaction retained summary/tail and its later delivery: uninspected (The records explicitly leave host compaction state and processing uninspected; Write/admission control cannot be classified. This is not absence and does not include the host model weights.)	partial coverage (Requested consumer reads are pull; claimed auto-loading of mutable project guidance is push. Backup delivery and host compaction are unresolved, so neither a pull-only inventory nor complete delivery is asserted.); Requested /start and task-demand archives, plans and code-navigation reads: known — pull [claimed] (The human startup request and agent demand reads are pull operations. Delivery of requested contents is not a second push.); Requested /audit and /lessons update-consumer reads: known — pull [claimed] (Expanded evaluation/dedup reads satisfy these workflows; unspecified external audit tools are not wired consumers.); Use-modified CLAUDE guidance automatically loaded by the host: known — push [claimed] (Only the mutable instructions component belongs here; static shipped doctrine is excluded. Auto-read is a template claim and host delivery remains uninspected.); Exported helper at a later code caller: known — pull [afforded] (The documented later caller can request helper execution; populated call sites remain unavailable.); PreCompact backups: not-determinable (Storage exists but the hook specifies no later restore/read route; no direction is established and no bounded absence record permits a system-wide no-read finding.); Host compaction retained summary/tail and its later delivery: uninspected (The records explicitly leave host compaction state and processing uninspected; Request, selection and delivery direction cannot be classified. This is not absence and does not include the host model weights.)	partial coverage (Only coarse claimed auto-loading is established for push. Identifiers and model relevance on explicit file reads are not push signals; unresolved host/backups prevent an axis-wide inapplicable assessment.); Host auto-loading of use-modified CLAUDE guidance: known — coarse [claimed] (Trigger: session start. Selector input: installed project instruction availability. Selected part: project guidance body. No record establishes targeted identity matching among retained entries.); Requested startup, archives, audit/harvest and numbered code reads: inapplicable (These classified operations are pull. Archive relevance judgment and grep identity matching select requested material, not independent push delivery, so they do not establish inferred-judgment or identifier push signals.); Later helper invocation: inapplicable (The afforded future caller requests execution; the recorded call interface establishes no automatic push selector.); PreCompact backups at possible later delivery: not-determinable (A fixed-list/existence selector chooses files to write, not retained material to deliver. Without a later delivery route no push signal can be determined.); Host compaction retained summary/tail and its later delivery: uninspected (The records explicitly leave host compaction state and processing uninspected; Push trigger, selector input and selected payload cannot be classified. This is not absence and does not include the host model weights.)	partial coverage (The ordinary checkpoint chain supports claimed yes locally. Manual harvest/QA and pure byte copies do not supply a contrary system-wide no; unresolved custom, diagnostic and host routes remain included.); Ordinary automatic checkpoint/incremental summaries and guidance to later /start: known — yes [claimed] (Conversation/work feeds model admission to durable behavior-shaping artifacts and specified later readers. Semantic novelty or observed improvement is not required; execution is not supplied.); Human-approved harvested lessons/helpers and human QA admissions: inapplicable (These admission mechanisms require explicit human content decisions, so they do not qualify as automatic trace-fed writes. This does not negate automatic ordinary LB admission or experience authority at harvest.); Pure PreCompact copies: inapplicable (The hook copies existing files unchanged, does not consume raw conversation to derive behavior-shaping memory, and specifies no later copy consumer; it is not a qualifying learning write.); Code sidecar generation and audit report production: not-determinable (Automatic durable writes and later consumers are specified, but code bodies and consistency inputs are not necessarily activity traces; no supplied input instance settles whether a qualifying trace-fed write exists on these branches.); Declared consumer custom-context file bodies: not-determinable (The declaration/read interface was inspected, but bodies and their histories were not supplied; Whether automatic trace-fed admission has a later behavior-shaping consumer cannot be determined from the declaration alone.); Host compaction retained summary/tail and its later delivery: uninspected (The records explicitly leave host compaction state and processing uninspected; Whether an automatic trace-fed durable update and later consumer qualify cannot be classified. This is not absence and does not include the host model weights.)	partial coverage (Knowledge, instruction, learning, routing, validation and afforded helper enforcement have local warrants. No ranking or guaranteed enforcement by natural-language rules is inferred; opaque and missing consumers keep coverage partial.); Resume/session history, task provenance, comments, QA, plans and audit bodies at later project agents: known — knowledge [claimed] (These files supply context or reference to later model work; retention and delivery do not demonstrate activation.); Unactioned user directives, decisions and use-modified CLAUDE rules at later work agents: known — instruction [claimed] (The force is requested standing guidance, not shell enforcement or an observed veto.); Lessons at later work agents: known — knowledge [claimed], instruction [claimed] (Scoped failures, workarounds and reasons can supply advice; adopted behavioral directions are a separate force at the same consumer.); Retained session/work experience at ordinary checkpoint update consumer: known — learning [claimed] (Update authority is distinct from the downstream knowledge/instruction force. This is a requested trace-fed transformation, not improved capacity.); Script experience at harvest update consumer: known — learning [claimed] (Human approval controls installation; it does not erase the experience-fed update consumer.); Clean-state/category/resolution metadata and archive/sidecar descriptions at read selection: known — routing [claimed] (The records describe branch and input-window selection, not relative scoring; no ranking value is inferred from relevance alone.); Retained task/decision/LB/helper inventory at duplication and audit consumers: known — validation [claimed] (Existing records and address sets are reference criteria for new entries, coverage and correspondence; their truth is not established by parity.); Exported helpers at future product callers: partial — enforcement [afforded] (A future callable guard is afforded, but actual helper implementations and call sites are unavailable; other effects and complete authority coverage are not determinable.); PreCompact backup copies at a later consumer: not-determinable (The shipped hook specifies persistence but no restoration or delivery consumer; copied content has no established later authority merely by existing.); Declared consumer custom-context file bodies: not-determinable (The declaration/read interface was inspected, but bodies and their histories were not supplied; The content and force at its consumer cannot be determined from the declaration alone.); Host compaction retained summary/tail and its later delivery: uninspected (The records explicitly leave host compaction state and processing uninspected; The later consumer effect cannot be classified. This is not absence and does not include the host model weights.)	AAS-2026-10-13-claude-context-guard-357cd90664df-01
claude-obsidian	2	Package-owned retained user-vault pages/navigation/hot, raw payloads and provenance, operation log/folds, Canvas/Bases views, manifest and filing/threshold configuration, chunks/BM25 and rerank/tiling caches, capture queue/predecessor, transaction backups/journals/results and optional Git checkpoint history/metadata; includes unresolved initial/adopted content, raw encodings and caller payloads. Excludes static installed skills/templates before use, transient unadmitted worker drafts and current host turns, external host compaction/checkpoints, provider-internal caches/weights and external renderer/service implementation.	known coverage (The recorded package-owned inventory is file-backed; optional Git checkpoint history adds a repo substrate. Opaque raw bytes do not make their file substrate unknown. External host compaction and provider stores are excluded, not classified from service labels.); Vault notes, orientation, navigation, logs/folds, raw payloads and ledgers, manifest, views, configuration, derived chunks/index/vectors, queue/predecessor and transaction/checkpoint metadata: known — files [wired], files [claimed] (Records identify these operative retained parts as local Markdown, JSON/YAML, binary files or original-byte backups; optional checkpoint metadata is local state.); Optional operation-scoped Git checkpoint history: known — repo [wired] (The memory annotation traces durable Git history as an alternative audit/recovery representation; it is not a model loader.)	partial coverage (Explicit reviewed content paths and automatic code admission coexist. Approval hashes do not authenticate human review. Direct-capture callers, initial adopted state, queue submissions and optional checkpoint control remain unresolved.); Explicit selected-content Save, reviewed canonical replacement/Research merge, forced fold replacement, mode and view changes: known — manual [claimed] (This unit is the prescribed operator decision, not a claim that every CLI invocation actually received human approval.); Chunk/prefix admission, rerank vector cache and tiling cache publication: known — automatic [wired] (Workflow/egress initiation is separate from per-entry content admission.); Transaction journal/result/backup persistence and queue runtime transition bookkeeping: known — automatic [wired] (Software derives and admits operation evidence and state changes; this does not classify who supplied proposal/result content.); Direct capture content decision; queue enqueue/result payload admission; initial adoption; CLI hash without authenticated review; optional checkpoint content choice: not-determinable (Generic caller identity and a correct approval digest do not establish an explicit human per-content decision. No supplied trace resolves initialization or checkpoint human control.)	partial coverage (Requested vault/code consumption and lifecycle hot supply are distinct operations. Fold/view discovery is afforded. Queue external delivery, opaque raw extraction and external host state do not become known from these witnesses.); Requested Query/Save/Ingest/Research pages, ledgers, manifest and mode selection: known — pull [wired], pull [claimed] (A request for retained material remains pull even when software ranks or delivers the response.); Requested rerank/tiling, queue list/claim/resume/recover and transaction/checkpoint recovery/replay: known — pull [wired] (These later command invocations request and consume retained caches/control state.); Index-discoverable folds and inspectable/renderable views: known — pull [afforded] (This is an afforded reader route, not an automatic model loader.); Opted-in later SessionStart hot emission and Stop retained-journal warnings: known — push [wired] (Lifecycle triggers supply retained orientation/status without a current request for that material; external host delivery is unobserved.); Raw opaque payload extraction and queued external executor delivery: not-determinable (Raw ingest is documented but not all payload extraction paths or queue executors/consumption are established; state completion is not delivery evidence.)	partial coverage (Lifecycle push uses coarse availability/consent/budget selection, not targeted relevance. Pull routes have no push selector. Unresolved external execution/delivery cannot be removed by the known hook witness.); SessionStart startup/resume/clear/compact hot context and Stop journal/status warnings: known — coarse [wired] (Trigger is lifecycle event; inputs are opt-in/vault eligibility, hot availability/byte cap or bounded journal/status scan; selected parts are hot bytes or recovery warnings.); Requested retrieval, configuration, caches, queue and recovery/fold/view reads: inapplicable (These assessed operations fulfill requests and therefore have no push-selection signal; lexical/cosine request ranking is not a push signal.); Unresolved queued external payload delivery: not-determinable (No executor/delivery chain establishes whether later payload supply is requested or pushed, much less its selector.)	not-determinable coverage (No canonical bounded-absence record supports a negative, and the accepted records establish no qualifying automatic trace-fed durable behavior-shaping chain. Manual retention and automatic page-derived caches must not be combined into such a chain.); Conversation-derived Save notes/hot/log and operation-log folds: not-determinable (Trace inputs and potential later context/discovery are recorded, but explicit reviewed content retention does not establish automatic trace-fed admission; the hash does not resolve actual human control. No qualifying automatic route or bounded absence is established.); Automatic page-derived prefixes/chunks and embedding caches: not-determinable (Automatic writes and later consumers are known, but pages are the immediate input. Their recorded mixed/unknown initial lineage does not establish a trace-origin input for these writes or bounded absence of one.); Queue transition/results, transaction history and optional checkpoint state: not-determinable (Runtime bookkeeping is retained, but no automatic trace-to-durable behavior-shaping artifact and later contextual/update consumer is established. Arbitrary result provenance/control remains unresolved; this is not a system-wide no.); Explicit configuration, raw acquisition, adopted notes and views: not-determinable (Configured or imported content alone establishes neither qualifying automatic trace learning nor bounded absence across unknown initial/caller content.)	partial coverage (The inventory establishes contextual, ranking, routing and computational gate effects at named consumers. Opaque raw/caller payload consumers and some downstream withdrawal/rendering effects remain unresolved. Static installed instructions and provider weights are outside this memory scope.); Wiki/hot/claim text and source locators at later Query and drafting models: known — knowledge [claimed], knowledge [wired] (The force is evidence/context under the installed rules, not standing directions or demonstrated obedience.); Fold text at later discoverers and view text at afforded human/model readers: known — knowledge [afforded] (Cataloguing and organization afford contextual use, not a mandatory deployed loader.); Contextualized chunk text/postings and rerank cached vectors at retrieval scoring; tiling vectors and bands at diagnostic scoring: known — ranking [wired] (Fixed algorithms use retained text/features/vectors to prioritize candidates or diagnostic attention.); Manifest source/hash identity, mode/folder config, queue state and checkpoint operation-path state at next-operation consumers: known — routing [claimed], routing [wired] (These parts select work, destination, delta scope or checkpoint inputs rather than supply factual warrant.); Source/claim support metadata at ledger validator and prospective transaction gate: known — validation [wired], enforcement [wired] (Retained status, freshness, review, independence and evidence links supply admission criteria.); Journal/backups/results and queue tokens/state at recovery, replay and transition code: known — validation [wired], enforcement [wired] (Original note backups are restoration bytes here, not Query evidence.); Opaque raw payload consumption, arbitrary queue result consumers and rejected/deprecated downstream presentation: not-determinable (Locators and state labels do not establish every actual payload consumer/effect; generic current-support rules do not prove complete semantic retraction.)	AAS-2026-10-13-claude-obsidian-c5e56e25c8bb-01
claude-workstream-kit	2	Included: accumulated ACTIVE/workstream parts, optional research notes, consumer-edited PROJECT selectors, handoff inbox, archive ledger and tagged snapshots, adopted standing-guidance edits, retained telemetry and evaluator session output, and their kit write/maintenance/later-consumer routes. Excluded: unchanged shipped instructions and empty seeds as such, arbitrary consumer product storage, external native Tasks/plan/account memory, host compaction checkpoints, model weights, and transient subagent packets or measurement stdout not adopted into retained files.	known coverage (The scoped retained inventory is file-based, with prescribed Git retention. External provider/native checkpoints and arbitrary product storage are excluded, not classified as absent.); Live pointer and workstream parts; optional notes, PROJECT selector, inbox and archive ledger: known — files [wired] (The recorded inventory separates side files from live containers; wiring establishes file-oriented persistence, not compliance with every requested write.); Committed state, handoff arrivals and historical tag snapshots: known — repo [claimed] (Repository retention depends on the instructed commits and tag discipline; a tag existence check does not establish snapshot fidelity.); Session context telemetry and evaluator raw output: known — files [wired] (Temporary storage is included because later capture or maintainer invocations can consume it; eventual cleanup is external.)	partial coverage (The explicit human and model gates are retained separately from CLI caller ambiguity and unavailable initialization.); Exact creation wording, substantive restructuring, closure approval and confirmed handoff filing: known — manual [claimed] (These decisions are specific content/admission gates, unlike merely starting work.); Routine capture, checkpoint updates and mechanical review repairs: known — automatic [claimed] (Human initiation does not change subsequent model admission to manual.); Numerical telemetry replacement and evaluator log retention: known — automatic [wired] (Starting the driver or host session is not approval of each retained result.); Direct rewrite admission under --write: not-determinable (The generic caller and write flag authorize physical replacement but do not establish whether a human approved that content. Deterministic transformation alone does not resolve admission agency.); Initial notes and accumulated PROJECT content admission: not-determinable (User editing is available and draft confirmation is requested, but actual initial admission control in consumer projects is unavailable.)	known coverage (The inventory covers the kit routes. External native checkpoints and host compaction are excluded; their unknown direction is not inferred from the kit hook.); SessionStart pointer and derived-signal delivery: known — push [wired] (Only ACTIVE and derived metadata/signals are pushed, not full workstream prose; actual host receipt is unobserved.); Detailed work/status/review/extract, receiver inbox, Git/tag and telemetry reads: known — pull [claimed] (Requested delivery is not another push. Historical recoverability does not mean automatic delivery of removed detail.); Maintainer inspection of evaluator output: known — pull [afforded] (The driver retains inspection artifacts but does not establish an actual later read.)	known coverage (All scoped push operations are the inspected reload chain. Coarse availability and targeted identity matching coexist; pull selectors do not add push values.); SessionStart availability, roster and threshold selection: known — coarse [wired] (Availability/threshold selection is separate from active identity matching; there is no fixed total token-output budget.); SessionStart active-workstream detail selection: known — identifier [wired] (The actual identity match selects delivered active detail, not wholesale injection of the requested full workstream.); Requested full-file, history, inbox, telemetry and log reads: inapplicable (These operations are pull, so their filenames, literal searches and project matching are not push-signal findings.)	partial coverage (Positive local yes findings do not settle other retained parts. Manual gates and automatic capture coexist; no system-wide negative is asserted.); Routine capture to committed resume state, Decisions/Learnings and later work: known — yes [claimed] (This is a prescribed trace-fed pathway, not observed operation, improved capacity or parameter learning.); Experience extraction to governed standing guidance: known — yes [afforded] (A Learning marker alone does not establish this path; actual governed adoption and later use are unobserved.); Telemetry, evaluator raw-log retention and their subsequent human decisions: not-determinable (Telemetry is access metadata and raw logs afford investigation; neither record establishes an automatic trace-fed durable guidance update from these parts. No bounded absence record licenses a local no.); Initial optional notes and project-list derivation: not-determinable (No automatic trace-fed initialization route is established for these included parts. Missing control/provenance prevents either yes or no.)	partial coverage (Authority follows each retained part to its actual consumer. Positive local forces coexist with unresolved destination effects; no universal enforcement or ranking is asserted.); Retained claims, notes, inbox substance and historical outcomes at later model readers: known — knowledge [claimed] (Reading claims does not semantically accept them.); Evaluator logs at the later maintainer reader: known — knowledge [afforded] (The inspection opportunity does not establish actual investigation or automatic upstream updating.); Resume directions, retained plans and closure criteria at the foreground model: known — instruction [claimed], validation [claimed] (Criteria are model/user-consumed admissibility conditions, not a filesystem-enforced gate.); Retained pointer and status/marker state at deterministic selectors: known — routing [wired] (Wiring supports fixed parser/selector consequences, not model obedience.); PROJECT list and telemetry at model/user selection consumers: known — routing [claimed] (The scripts wire list/number availability, but their model/user-directed selection effects remain prescribed and unobserved.); Retained open/gate structure and references at rewrite admission checks: known — validation [wired], enforcement [wired] (This force is bounded to the CLI write path. It does not extend to manual edits or certify semantic preservation.); Session findings and retained experience at capture/extraction update consumers: known — learning [claimed] (Learning authority at the update consumer is separate from downstream instruction/knowledge and from demonstrated improvement.); Governed standing-guidance edit at later instruction readers: known — instruction [afforded] (Only the standing-guidance branch is included; arbitrary product effects are not generalized to instruction authority.); Other destination-specific effects of retained adopted guidance: not-determinable (No deployed destination consumer establishes additional forces. The records explicitly require destination authority to be established there; no ranking finding follows from a task order or marker count alone.)	AAS-2026-10-15-claude-workstream-kit-103204afa1ab-01
Closure-SDK	2	Included: use-accumulated Closure EA genomes (DNA, Epigenetic, Response), buffer, access/correction statistics, closure histories/events, pending feedback, caller-side Vocabulary/MusicEncoder caches, CurriculumTrace replay input, BrainState, genome-only and vocabulary checkpoint routes and their recorded later consumers. Included unresolved alternatives: accumulated helper-state parts, arbitrary caller mutation/replacement provenance and untraced example-consumer variants. Excluded: static compiled control code as memory, unconsumed returned reports/promotion packets, sibling integrity SDK/database, archived Python runtime, separate homeostasis demo, browser implementation and external compiler/filesystem responsibilities.	partial coverage (Coverage includes live numerical and symbolic objects plus all three recorded checkpoint alternatives and caller-side caches. Unenumerated helper-state alternatives prevent complete substrate coverage.); Live genomes, buffer, statistics, histories, pending prediction and replay input: known — in-memory [wired] (These named objects have inspected Rust field/vector storage; file encodings are classified separately.); Vocabulary and MusicEncoder live maps: known — in-memory [wired] (The caches are local HashMaps, not an established external KV service.); BrainState and Vocabulary JSON checkpoints: known — files [wired] (Named serializers write local files; JSON is encoding, not a separate substrate.); Genome-only checkpoint: known — files [afforded] (The distinct genome-only reuse route is afforded, not shipped cross-process use.); Included helper-state and additional example-consumer alternatives within ThreeCell: uninspected (The accepted account leaves detailed body-state, localization and other helper internals and example variants unaudited. Their accumulated-state parts and subsequent consumers are not independently enumerated; the known field routes do not classify those alternatives.)	partial coverage (Automatic admission is known on named implementation routes. Caller-controlled mutation and replacement lack control-principal evidence, and helper alternatives remain uninspected.); Buffer admission, structural/hierarchy writes, feedback and maintenance software admission: known — automatic [wired] (Calling a workflow or forcing maintenance does not make its per-content admission manual.); Fixed seeding, direct learn_response and cache admission: known — automatic [wired] (These software mechanisms append, interpolate or cache content without a human approval gate.); BrainState software installation: known — automatic [wired] (Parsing/construction installs fields; upstream replacement choice is a separate unit.); Genome-only software installation: known — automatic [afforded] (The intended resumed-training route remains afforded.); Arbitrary public mutation and caller choice/edit/replacement of checkpoints or initial content: not-determinable (Generic caller identity establishes neither an explicit human per-content decision nor an independently evidenced automatic upstream replacement mechanism. No manual finding can be derived from save/load selection alone.); Included helper-state and additional example-consumer alternatives within ThreeCell: uninspected (The accepted account leaves detailed body-state, localization and other helper internals and example variants unaudited. Their accumulated-state parts and subsequent consumers are not independently enumerated; the known field routes do not classify those alternatives.)	partial coverage (All classified delivery chains are requested computations, not unsolicited model-context injection. Incomplete helper/consumer coverage prevents a whole-boundary pull-only assertion or a system-wide exclusion of push.); Requested ingest, feedback, evaluate, generate, solve and orbit computations over retained state: known — pull [wired] (Internal selection and return fulfill caller computation requests; delivery is not another push.); Explicit higher-level and genome-only resumed-training consumers: known — pull [afforded] (These consumer roles/capabilities are afforded; automatic promoted-category delivery into primary inference is not established.); Vocabulary and role-cache later lookup/encoding calls: known — pull [wired] (Lookup, decode and encoding are requested helper computations.); Explicit full-state reconstruction: known — pull [wired] (A caller requests a chosen checkpoint; its payload is not automatically injected.); Included helper-state and additional example-consumer alternatives within ThreeCell: uninspected (The accepted account leaves detailed body-state, localization and other helper internals and example variants unaudited. Their accumulated-state parts and subsequent consumers are not independently enumerated; the known field routes do not classify those alternatives.)	uninspected coverage (No positive push-signal finding is supported. Pull-chain units are locally inapplicable, but uninspected included consumer alternatives prevent whole-boundary inapplicability.); Known requested numerical-field, checkpoint and cache delivery chains: inapplicable (The recorded chains fulfill requests and therefore have no push selector to classify. Geometric distance in a pull read and an identifier on a requested checkpoint do not establish a push signal.); Included helper-state and additional example-consumer alternatives within ThreeCell: uninspected (The accepted account leaves detailed body-state, localization and other helper internals and example variants unaudited. Their accumulated-state parts and subsequent consumers are not independently enumerated; the known field routes do not classify those alternatives.)	partial coverage (Qualifying automatic numerical update chains support yes locally. No local or system-wide no is asserted without bounded absence records; unknown caller/helper alternatives prevent complete coverage.); Perception/closure history to learned carriers and later numerical field consumers: known — yes [wired], yes [afforded] (Automatic trace-fed retained numerical artifacts qualify within the same object; file saving is not required. Higher-level later reading remains an affordance.); Committed prediction/target experience to Response state and later reads: known — yes [wired] (Trace learning here means wired durable update, not improved capacity or authenticated targets.); Query/correction aggregates to consolidation and later field consumers: known — yes [wired] (Reasons are only partially retained and higher-level primary-inference benefit is unestablished, without negating the primary-state chain.); Initial seed, arbitrary replacement, caller-input provenance and cache alternatives: not-determinable (Fixed seeding and unchanged checkpoint copying alone do not qualify. Generic caller inputs and arbitrary mutations lack recorded trace provenance/control; the records provide no bounded trace-learning absence warrant for every alternative.); Included helper-state and additional example-consumer alternatives within ThreeCell: uninspected (The accepted account leaves detailed body-state, localization and other helper internals and example variants unaudited. Their accumulated-state parts and subsequent consumers are not independently enumerated; the known field routes do not classify those alternatives.)	partial coverage (Recorded consumer effects support knowledge, ranking, routing, validation, enforcement and learning locally. No blanket instruction authority follows from executable influence; unestablished consumers prevent a complete authority union.); Carrier content at later field and recall consumers: known — knowledge [wired] (Returned values supply numerical context/prediction content, not standing natural-language directions.); Retained addresses and counts at relative-selection consumers: known — ranking [wired], ranking [wired] (Scoring fixed retained addresses is ranking even without learned ranking policy.); Retained indices, edges and hierarchy state at operation-selection consumers: known — routing [wired] (State selects subsequent inputs, transition destinations and update targets.); Layer and access statistics at correction/maintenance admission: known — validation [wired], enforcement [wired] (The force belongs to consumed layer/statistical controls, not every numerical result.); Retained observation/feedback state and statistics at automatic updates: known — learning [wired] (Update-consumer authority is separate from the output field's knowledge/ranking force.); Vocabulary and role keys at lookup/encoding consumers: known — routing [wired] (Keys choose retained associations; decoded display labels have no established semantic downstream consumer.); Installed checkpoints and explicit higher-level reads: partial — knowledge [afforded] (Installed field parts retain the forces above, but automatic promoted-category delivery to primary prediction and vocabulary display interpretation are not established.); Included helper-state and additional example-consumer alternatives within ThreeCell: uninspected (The accepted account leaves detailed body-state, localization and other helper internals and example variants unaudited. Their accumulated-state parts and subsequent consumers are not independently enumerated; the known field routes do not classify those alternatives.)	AAS-2026-10-13-closure-sdk-294e6865ca51-01
cludebot	2	Accumulated PostgreSQL memory/conversation content, JSON and SQLite memory, dream queue results, derived semantic/procedural/self-model/journal content, vectors and graph/access state, shadow/evaluation artifacts, changing prompt-sync snapshots, their admission/maintenance and later-consumer routes. Portable pack/client alternatives remain included but unresolved. Excludes transient current-call context as a retained object, static shipped/setup policy, inference-provider weights and promotional/workstation metadata; dependencies and current-call renderers matter only at retained-memory consumers.	partial coverage (Positive substrates cover inspected stores and access structures, not an exhaustive inventory of every portable/client/service alternative.); PostgreSQL content, conversations, queue/log and access-state rows: known — rdbms [wired] (These inspected relational parts have source-grounded persistence; provider weights are not their substrate.); Persistent dense-vector indexes and memory-link structures: known — vector [wired], graph [wired] (Logical vector and graph access stores coexist with their relational backing; no separate graph service is inferred.); JSON local memories, prompt-sync snapshot and benchmark result artifacts: known — files [wired] (These are accumulated content surfaces, unlike static setup instructions.); SQLite memories, vectors and dream queue: known — sqlite [wired] (The local database is not classified as PostgreSQL; queue completion is not recall adoption.); Portable pack payloads and remaining service/client retained alternatives: uninspected (Pack reading and mode dispatch are established, but retained portable parts, all client stores and external storage internals are not enumerated by these records. No additional substrate can be assigned.)	partial coverage (Automatic admission is established locally; generic callers and uninspected portable/client mechanisms prevent a complete control inventory. No explicit human-content witness establishes manual here.); Automatic chat admission, dream/journal/strategy production, graph/link feedback and tool findings: known — automatic [wired] (A user request, scheduler start or owner task assignment does not make subsequent individual admissions manual.); Upload-node admission, prompt snapshot refresh, shadow logging and queue completion: known — automatic [wired] (Known automatic transformations are separate from unresolved original input or queue admission control.); Generic SDK/MCP writes, caller patches/deletion, selected import content and local queue admission: not-determinable (Callers can be human or agent. The records establish interfaces and software checks, not an explicit human content decision for these mechanisms; manual versus automatic control cannot be completed.); Portable marketplace/client admission and remaining initial-content mechanisms: uninspected (All transfer, installation and client-side replacement controls are not traced. No manual classification follows from operator-facing UI or ownership alone.)	partial coverage (Request, automatic selection and return are separate operations. Wired automatic context supply coexists with requested pull; unresolved host/portable alternatives remain included.); Explicit SDK/MCP recall, summary/hydration, trace explanation and requested service returns: known — pull [wired] (Requested delivery is not counted again as push merely because code returns it.); Automatic memory/history supply in chat and procedural supplementation inside response generation: known — push [wired] (These automatic selectors supply retained parts to the generation model. They coexist with the outer requested response; no unsolicited background user message is asserted.); Automatic source selection for recurring maintenance and journal contexts: known — push [wired] (Later update consumers receive selected retained material, not the full foreground chat frame.); Snapshot supply to an external host: known — push [afforded] (File production is wired but the host supply route remains an affordance, not inspected loading.); Local queue-result adoption, portable/client consumers and remaining MCP host use: uninspected (Known requested APIs and available files do not establish every later consumer or delivery route. Queue results cannot be treated as pull-only or absent.)	partial coverage (Signals characterize actual automatic selection operations, not IDs printed on requested records. Coarse filtering remains separate from identity matching and semantic selection.); Chat automatic memory retrieval and conversation-history fitting: known — inferred-embedding [wired], identifier [wired], coarse [wired] (Triggers are authenticated messages; inputs are current text, owner/conversation IDs and budgets. Expansion is skipped; no universal lexical lane is inferred for this call.); Automatic procedural supplementation at memory-enabled shared responses: known — coarse [wired] (General strategy slots are coarse selection, not identifier targeting. Reranker alternatives in enhanced recall remain unresolved below.); Main/hosted maintenance, self-model/journal continuity and action refinement source windows: known — coarse [wired], identifier [wired] (The matched concept/feature actually selects retained parts. Journal divergence and contradiction-pair provider internals are not characterized by these coarse witnesses.); Snapshot export selection for afforded later host supply: known — coarse [afforded] (The signal describes the afforded host-supply chain; file replacement itself is wired.); Enhanced reranking, journal divergence, contradiction-pair selectors and opaque delivery alternatives: uninspected (Not all selector internals or downstream consumers were inspected; inferred judgment/lexical/embedding alternatives cannot be exhaustively classified.)	partial coverage (Positive wired trace-learning routes coexist with unresolved alternatives. There is no bounded absence record supporting a local or whole-system no.); Chat-experience to main/hosted semantic and procedural memory, self-model and journal continuity: known — yes [wired] (The complete raw-to-derived-to-later-consumer wiring qualifies without observed improvement; prior shortened summaries may lose reasons.); Action/outcome feedback to procedural rules and durable strategy importance: known — yes [wired] (Metric provenance and causal efficacy remain limited; qualifying trace-fed writes do not require proven behavioral benefit.); Local dream queue-result route: not-determinable (Generation and retention are wired but no later adoption/consumer is established in the completion sink. That missing consumer prevents either a positive chain or a bounded negative.); Mixed upload/import inputs, retrieval feedback origins and remaining social/client/portable writes: uninspected (Document content, complete trace origin and all later behavior-shaping consumers are not recorded for every route. Raw capture or metadata change alone is not generalized into qualifying trace learning.)	partial coverage (Authority is consumer-specific. Positive update, guidance, ranking and access effects do not resolve opaque or unconsumed alternatives.); Selected content and summaries at answering, explanation and journal consumers: known — knowledge [wired] (The unit is the contextual content path, not shipped grounding rules.); Retained procedural rules at the shared response model: known — instruction [wired] (Force is interpretation-dependent standing guidance, not an execution-time enforcement mechanism.); Recorded experience and prior self-representations at durable artifact update consumers: known — learning [wired] (Learning authority names the update consumer; it asserts neither parameter training nor improved capacity.); Text, vectors, links, importance, decay and access state at recall scoring: known — ranking [wired] (Retained ranking inputs suffice even when scoring policy is fixed code.); Queued operation/IDs at local dream dispatch and stored relation/ID inputs at trace traversal: known — routing [wired] (Routing force is on operation/input selection, independent of whether a generated queue result has a later consumer.); Retained encrypted access state and content at revoke/redelegate and read guards: known — enforcement [wired], validation [wired] (These are access/authentication effects, not truth validation; hydration and unaudited lanes prevent a universal veto claim.); Prompt-sync content at a later external host: known — knowledge [afforded], instruction [afforded] (Snapshot production is wired; host loading and compliance are only afforded.); Queue results, shadow proposals, portable revocations and remaining host consumers: uninspected (Queue adoption, production use of shadow proposals and all-importer revocation filtering are not established. Their downstream force cannot be upgraded from retention alone.)	AAS-2026-10-15-cludebot-cb70d11d215b-01
CocoIndex	2	Retained component/function memos, persistent user-state payloads and reconciliation metadata; PostgreSQL document/session text, vectors, session metadata and manual extraction products; memoized entity decisions; alternate image/filing indexes and their shipped consumers. Includes unresolved opaque payloads, upstream imported checkpoint-summary derivation, untraced producer/read-back branches and remaining included retained-output alternatives. Excludes static shipped instructions, transient retry messages/check exits/current-run maps as memory in themselves, the unacquired CocoIndex-code product and provider/model-weight implementations outside the frozen repository; their adapter-facing uncertainty remains included.	partial coverage (Inspected internal and external stores are classified separately. Alternate producers and arbitrary caller payloads prevent a complete store inventory.); Component and function memo envelopes, serialized user state and reconciliation bookkeeping: known — kv [wired] (The internal per-app LMDB store is a key-value substrate, not an external content index.); PostgreSQL document text, vectors, session rows, session metadata and manual extraction rows: known — rdbms [wired] (Dense-vector encoding does not replace the relational storage classification.); Alternate image and filing indexes: partial — vector [wired], rdbms [wired] (The record identifies Qdrant and filing_chunks consumers, but full producer/storage arrangements are untraced.); Other retained outputs and opaque caller/provider alternatives: uninspected (Generic serialized values do not enumerate every operative retained object; alternate producers and provider implementations remain untraced. No files, model-weights or service-object storage union is inferred from execution components.)	partial coverage (Software admission is established locally; initiating an app is not per-record approval. Arbitrary initialization and alternate-producer control stay unresolved.); Index row construction/reconciliation and session parsing/embedding admission: known — automatic [wired] (No per-content operator decision gates these inspected transformations.); Manual extraction proposal, schema-mediated row admission and memo resolution: known — automatic [wired] (The model proposes; code validates shape and persists successful results.); Internal memo and bookkeeping writes: known — automatic [wired] (Physical commit is not the agency test; software result resolution and reconciliation determine content admission.); Concrete counter assignment in shipped test source: known — automatic [wired] (Only the concrete code-controlled counter mechanism is classified; no test execution is claimed.); Generic state initialization/assignment and alternate-index producers: not-determinable (Caller identity alone does not decide manual versus automatic admission; alternate producer mechanisms were not traced. Source authorship does not supply a manual admission finding.)	partial coverage (Requested search/state/diagnostic returns and automatic engine reuse are distinguished within the chain. Untraced metadata/manual readers and alternate producers prevent complete direction coverage.); Document/session queries, image/filing requests and diagnostic CLI: known — pull [wired] (The consumer explicitly requests retained material; automatic ranking and printing fulfill that request rather than adding another push.); Explicit keyed StateHandle access: known — pull [afforded] (The API affords later retrieval and the counter test supplies a concrete caller; no deployed route is claimed.); Component/function automatic cached-return substitution on later computation requests: known — push [wired] (The caller requests computation, not the retained memo; engine selection automatically supplies a prior result when reusable.); Session metadata/manual row readers, generic payload consumers and other included alternatives: uninspected (Retention does not determine every later consumer request or automatic delivery path. Live source-event maintenance is not itself retained-context read-back.)	partial coverage (Signals classify push selectors only; vector/lexical ranking inside requested search is not an inferred push signal. Incomplete alternative consumers prevent complete coverage.); Engine component and function memo push selection: known — identifier [wired] (Trigger: later component/function computation. Inputs: stable component path or call fingerprint plus validity dependencies. Selected part: corresponding retained return and dependencies.); Explicit semantic search, state requests and diagnostics: inapplicable (These named operations fulfill requests for retained material; their relevance ranking and requested identifiers are not push signals.); Uninspected later metadata/manual/generic payload consumers and alternative push possibilities: uninspected (Missing consumer/selector wiring prevents deciding whether another push exists and, if so, its trigger, selector input and selected retained part.)	partial coverage (The local session trace-to-derived-search-context path supplies a positive wired route under the profile definition, not observed capacity improvement. Upstream summaries, opaque writes and other included alternatives remain unresolved.); Automatic session transcript filtering/embedding to retained text/vectors and later query snippets: known — yes [wired] (Raw conversational text is transformed into durable retrieval content and selection representations; later snippets supply context/reference to the query recipient.); Upstream imported context.md generation and generic state/memo or alternate-index trace-fed updates: not-determinable (The context summary generator is outside the inspected record; arbitrary payloads and alternative producers lack a complete trace-write-artifact-consumer account. No system-wide no is supported by these gaps.); Pair retry feedback and retained return: not-determinable (Retry messages change current-call context but are not shown retained; the final memoized decision is supported, but the records do not establish an original qualifying trace input to that durable write.)	partial coverage (Findings name actual consumers, not every force an arbitrary stored payload might acquire. Opaque payloads and untraced downstream consumers remain unresolved.); Retained document/session text and image/filing result content at search recipients; internal diagnostic summaries at operators: known — knowledge [wired] (Search/display supplies reference material, not standing instructions or factual endorsement.); Retained dense document/session vectors and filing text/features at search selectors: known — ranking [wired], ranking [wired] (Scored retained material changes relative result priority even though scoring code is fixed.); Memo dependency/identity envelopes and ownership/recovery bookkeeping at engine consumers: known — routing [wired], validation [wired] (Return payload semantics are excluded from this envelope unit.); Memoized pair-decision fields at canonical-map application: known — routing [wired] (This is the decision content consumer, not a claim about which memo schema persists it.); Arbitrary memo/state payload effects, session-metadata and manual-content later consumers; alternate opaque parts: not-determinable (Known writes do not establish uninspected downstream consumers. No instruction, enforcement or learning authority is inferred merely from retention, a static prompt, or a transient check result.)	AAS-2026-10-15-cocoindex-3514227a0537-01
Cognee	2	Accumulated raw corpus and relational/provenance records; graph content/structure, vector access representations, session QA/traces/feedback/guidance and cache alternatives; session learning documents, ranking/preferences/truth state, global summaries, skills/metadata/runs/proposals; their admission, maintenance, withdrawal and later answer/update routes. Includes recorded unresolved typed, cache, extraction/index, concurrent, custom and alternate recall/interface paths. Excludes static shipped instructions, current-turn/ContextVar frames, external model weights/provider internals and arbitrary enclosing-host behavior. The source artifact is complete but its memory-use loop is only partially traced.	partial coverage (Graph, vector, raw-file and relational witnesses coexist with partially inspected cache alternatives and untraced retained parts. No external provider weight store is included merely because inference uses it.); Retained raw ingestion files: known — files [wired] (Memory-only forget preserves raw files for rebuilding; loaders and file lifecycle are not fully traced.); Retained relational data records and conditional provenance ledger: known — rdbms [wired] (Relational data records survive memory-only forget; provenance ledger is a conditional alternative to graph-folded provenance.); Graph content text: known — graph [wired] (DataPoint text is written to configured graph storage.); Graph IDs, edges and graph-folded provenance: known — graph [wired] (Symbolic graph structure is retained; ledger-backed provenance is classified separately.); Session-learning graph representations: known — graph [wired] (Published learning documents are cognified into retained graph content.); Global feedback weights: known — graph [wired] (Numerical feedback weights are graph attributes; cache bookkeeping is a separate part.); Personal preference edges and preference text: known — graph [wired] (Both preference tracks retain state on the user/dataset graph.); Truth-alignment per-chunk state: known — graph [wired] (Per-chunk coordinates and epochs are graph properties, not provider weights.); Global summary text and hierarchy: known — graph [wired] (Summary DataPoints and summarized_in edges are graph state.); Skill procedures and metadata: known — graph [wired] (Deployed procedures and scope/discovery fields are graph Skill parts.); SkillRun summaries: known — graph [wired] (Run summaries are graph DataPoints.); Skill improvement proposals: known — graph [wired] (Old/proposed procedures, rationale and status are retained as graph DataPoints.); Retrieval vectors, payload projections and indexed text copies: known — vector [wired] (Vector collections retain embeddings and projected payloads; encoding and provenance are separate axes.); Truth centroid slots: known — vector [wired] (Learning-derived centroids are retained by the vector engine.); Session QA, traces, feedback, active guidance and update bookkeeping: cache alternatives: partial — files [afforded], rdbms [afforded], kv [afforded] (The cache factory offers FS, SQL and Redis alternatives. These are afforded substrates, not audited backend semantics; Tapes storage is unresolved.); Tapes cache backing retained session parts: not-determinable (An append-only Tapes alternative is recorded, but its backing substrate is not established by accepted records.); Alternate raw loaders, code/tool lane objects, custom pipelines and storage adapters: uninspected (Accepted route limits leave these included retained parts untraced; no complete substrate inventory follows.)	partial coverage (Automatic internal admission is supported independently of unresolved initial/caller decisions. No manual finding is licensed solely by an explicit API call, workflow start or apply request.); Cognify extraction/storage admission: known — automatic [wired] (After workflow start, software/model output enters storage without per-item human approval.); Session-turn candidate/feedback/counter admission: known — automatic [wired] (Model proposals and deterministic shape/confidence/dedup gates decide writes.); Decorated outcome and shipped SkillRun capture: known — automatic [wired] (Software records enabled host/tool outcomes; invoking a host function is not content approval.); Trace promotion and lesson publication: known — automatic [wired] (Software/model windows and writer judgment admit derived memory without per-window/per-lesson human review.); Feedback/preference/truth/summary maintenance writes: known — automatic [wired] (Arithmetic, gates, embedding projection and model summarization admit subsequent state.); Typed/schema-controlled storage and skill-ingester stamping: known — automatic [wired] (Software supplies schema/access checks and stamps fields; this finding concerns computational admission, not proof of human content selection.); Skill replacement proposal construction: known — automatic [wired] (Opt-in starts drafting, but a model constructs and software stores the proposal.); Initial raw/skill supply, explicit feedback/editing and caller-selected proposal apply: not-determinable (Caller can supply, edit, withhold or request apply, but human versus computational principal is unresolved; neither manual nor automatic caller choice is established.); Alternative direct graph/custom pipeline/interface write controls: uninspected (Uninspected write alternatives prevent complete admission-control classification.)	partial coverage (Pull request/return and push automatic context/evidence selection coexist within chains. Neither storage APIs alone nor code delivery of an already requested part creates another push.); Consumer-requested recall/search return and model-requested load_skill body: known — pull [wired] (Caller asks for recall and foreground model asks for procedure by name; fulfilling either request is pull, not an extra push.); Configured decorator selection into wrapped gateway call: known — push [wired] (At host invocation code selects cached feedback and graph context without that model requesting those retained parts.); Session history/guidance/preferences into answer model: known — push [wired] (Answer preparation automatically selects retained context for the completion consumer within a request chain.); Root/relevant summaries into graph completion: known — push [wired] (Opt-in enables automatic prelude selection for the answering model; the external answer request remains pull.); Skill catalog and graph/history guidance into agentic model: known — push [wired] (Prepared model context automatically supplies selected metadata and retained context; later body requests are separate pull operations.); Retained traces/QA/guidance/runs into subsequent maintenance consumers: known — push [wired] (Maintenance code chooses pending windows, gated entries or selected failure runs and supplies them to update consumers.); Other recall lanes, concurrent session preparation and external/remote consumers: uninspected (One known requested path does not resolve every automatic supply alternative or actual host consumption.)	partial coverage (Push witnesses retain trigger, selector inputs and delivered parts. Body-name lookup on a pull request is not counted as identifier push; the distinct writer member-resolution operation is. The decorator's graph-selector signal and alternative-route coverage remain unresolved; neither an available index nor a query-relevant result establishes route-specific embedding selection. These limits prevent inapplicability or a complete signal inventory.); Decorator invocation: recent trace feedback and query-selected graph text: partial — coarse [wired] (Host invocation triggers selection; enabled sources/last_n and a 4,000-character cap select recent feedback, while query/top_k selects graph memory. Only structured-output gateway receives it. The accepted route and its memory annotation do not establish the graph selector's operation; GRAPH_SUMMARY_COMPLETION and an available vector index do not establish embedding-based selection on this path. Coarse filtering is known, but the graph-selection signal remains not determinable.); Session answer preparation: history and active guidance: partial — coarse [wired], inferred-lexical [wired], inferred-judgment [wired] (New answer preparation selects recent/history hits and guidance by profile/section, confidence, helpfulness, priority, recency and budgets. Session-vector implementation is unresolved.); Global-summary prelude selection: known — coarse [wired], inferred-embedding [wired] (Single-query opt-in triggers root plus top_k relevant non-root summaries; the answering model receives selected text, not all child bodies.); Skill discovery catalog supplied before procedure requests: known — inferred-embedding [wired], coarse [wired] (Retrieval preparation supplies catalog metadata; vector hits are filtered by active/dataset scope and trimmed. Requested load_skill name matching is pull, not an identifier-push witness.); Maintenance evidence selection from retained activity: known — coarse [wired], inferred-judgment [wired] (Improve or revision invocation triggers software selection of pending trace/QA windows, gated context or failing SkillRuns for update consumers.); Distillation writer member-entry delivery: known — identifier [wired] (For each curator proposal, code resolves member_entry_ids in entries_by_id and supplies those retained member contents to the writer; missing IDs are omitted. This is actual identity selection, not merely a catalog identifier.); Session vector-hit selector internals and alternative recall push paths: uninspected (Recorded vector hits do not establish the complete selector internals. Code/temporal/lexical/custom lanes and concurrent preparation remain untraced.)	partial coverage (Several automatic trace-fed writes with later behavior-shaping consumers support yes. Raw recording alone is not learning; caller-selected installation and opaque input alternatives stay unresolved. This axis does not claim improved capacity.); Automatic turn-feedback to active guidance and durable preferences: known — yes [wired] (Recorded conversational turns feed automatic guidance/preference writes with later answer consumers; cache retention can span invocations despite TTL.); Automatic decorated trace extraction/promotion to later memory: known — yes [wired] (Trace output/error records feed automatic standing-guidance and graph writes, then explicit profile/distillation/graph consumers. Do not infer agent-profile injection into every agentic loop.); Automatic session lesson publication: known — yes [wired] (Recorded QA and derived context feed automatic curator/writer publication into later graph recall and truth-build consumers.); Automatic feedback/preference ranking updates: known — yes [wired] (Recorded feedback and used IDs change retained code-ranking parameters for later retrieval, not provider weights.); Learning-derived truth geometry updates: known — yes [wired] (Session-derived learning text feeds automatic numerical retained updates consumed by conditional hybrid ranking; this is trace-fed shaping, not measured improvement.); Failure-fed stored skill proposal and its installation chain: not-determinable (Automatic drafting retains failure evidence and a candidate procedure, but proposal retention alone does not establish a durable behavior-shaping update. Later skill delivery requires caller-selected apply whose control is unresolved.); Qualifying installed skill revision: caller-selected apply control: not-determinable (Proposal drafting is automatic, but human/software control of installation is unknown; no automatic replacement claim follows from generic caller identity.); Typed trace/QA/SkillRun origins and alternate maintenance routes: not-determinable (Typed data may be static, mixed or authentic activity; global-summary input origin and alternate writes are not completely established. No system-wide no or absence claim is warranted.)	partial coverage (Knowledge, instruction, learning, ranking, routing, validation and enforcement have local consumer warrants. Unconsumed metadata is not granted authority; opaque or untraced consumers prevent complete coverage.); Graph content, QA history, trace display and summaries at answer/context consumers: known — knowledge [wired] (Readable retained content supplies reference/context to answering models or callers, not enforced factual truth.); Active guidance and durable preference text at completion consumers: known — instruction [wired] (Retained preference and working guidance supply standing directions in the guidance block; delivery is not obedience.); Loaded skill procedure at subsequent AgentStep consumer: known — instruction [wired] (Requested retained procedure arrives as a capped tool result and is interpreted as operational guidance.); Run summaries, QA/trace evidence and feedback at durable update consumers: known — learning [wired] (Retained experience feeds procedure proposals, guidance/lesson creation and persistent ranking updates; no capacity improvement is implied.); Indexed retained content and numerical access representations at similarity selection: known — ranking [wired] (Retained text/index features influence relative retrieval priority even without a learned ranking policy.); Retained guidance content/counters and feedback/preference/truth state at rankers: known — ranking [wired] (Query overlap, helpfulness and scalar/geometry adjustments change relative selection when enabled.); Persist/extraction/distillation/preference watermarks and dependency identities: known — routing [wired] (Retained progress markers select input windows or retry work; graph/dependency IDs select traversal and cleanup targets.); Skill active/scope/tool declarations at availability/dispatch consumers: known — enforcement [wired] (Retained scope/active/tool fields exclude candidates and narrow permissible dispatch; they do not sandbox handler effects.); Skill scope and truth epoch metadata at admissibility guards: known — validation [wired] (Retained scope and epoch state determine whether a candidate skill or alignment may be used; equality does not prove a correct basis.); Truth epoch state at conditional ranker veto: known — enforcement [wired] (Mismatched retained epochs prevent coordinate weighting on the inspected consumer path.); Retained conflict warning/supersession metadata at later readers: uninspected (Metadata writes are inspected, but later conflict/supersession consumers and their actual force are not. Do not infer enforced withdrawal from a superseded flag.); Proposal rationale review and alternate recall/host effects: not-determinable (Review is afforded; rationale use, every lane consumer and unrelated host-provider effects are unresolved.)	AAS-2026-10-14-cognee-d8839fe03c5b-01
Company Brain	2	Included: submitted/derived hosted company and agent-self content, entity cache, extraction settings, channel spool/batches and observer carry, local tags/node/generation/cursor state, mutable skills/workspace directions, investigation checkpoints and retained approval continuations, scheduled payloads, research draft/sent history and human read routes. Optional sandbox continuity and opaque provider parts remain included unresolved alternatives. Excluded: static shipped instructions as memory objects, ordinary untransformed live turn accumulation, inference weights without an evidenced accumulated-memory update, external connected-application state, and transient reception feedback not persisted by the inspected administrative routes.	partial coverage (The inventory covers hosted documents, derived rows, extraction settings, local indexes, observer state, instructions, continuation state, schedules and research history. Their named stores are classified; opaque service representations and optional sandbox lifetime remain unresolved.); Submitted documents, derived text rows, entity cache and agent-self notes: known — service-object [wired] (The operative application-facing objects are hosted service objects; this does not identify the provider database or embedding store.); Hosted extraction settings: known — service-object [wired] (Remote container settings are the retained configuration surface, not the shipped policy file.); Channel spool/cluster batch, observer carry/cursor, tags/node mappings/generation, skills, investigation and approval snapshots, research drafts/sent history: known — rdbms [wired] (DO SQL and organization Durable Object persistence establish relational retained state. JSON encodings are not separate substrates; no SQLite implementation is asserted from SQL alone.); Workspace-prompt persistence below the organization Durable Object API: not-determinable (The record identifies a Durable Object persistence boundary but not its physical store. The prompt is not included in the SQL finding above solely because other DO objects use SQL.); Scheduled payload storage: not-determinable (agent.schedule retains the payload, but the accepted record does not identify its underlying substrate.); Hosted embeddings, extraction/versioning internals and internal access structures: not-determinable (The service is opaque. Neither vector, graph nor model-weights follows from an embedding request or graph display.); Optional reused sandbox workspace: uninspected (Scoped reuse is included as continuity; provider filesystem lifetime and persistence are not established. No files substrate is inferred for durable memory.)	partial coverage (Known automatic and manual admissions coexist. Human authorship of source messages, starting research or holding an admin role is not per-content approval. Initial settings, external reset control and debug caller control remain unresolved.); Interactive/resumed knowledge capture, channel clusters and style notes: known — automatic [wired] (Model/software admits each retained proposal subject to software guards; humans do not approve each one.); Profile provisioning, tag/node registration and partitioning, entity cache, checkpoint/compaction and scheduled payload admission: known — automatic [wired] (These admitted outputs are selected by host/model code, even when a user starts the workflow or supplies input fields.); Requester-approved skill creation and human-supplied HTTP prompt/settings edits: known — manual [wired] (The decision approves or supplies the retained content itself, unlike merely initiating a turn.); Model-mediated admin configuration replacement: known — automatic [wired] (Execution role validation does not approve the proposed text separately.); Research draft insertion: known — automatic [wired] (Starting the round is not approving each generated draft.); Reviewer pending-body editing, dismissal and explicit send decision: known — manual [afforded] (Human editing/replacement and publication selection are separately available; actual reviewer behavior is unobserved.); Approved forget apply: known — manual [wired] (Preview is automatic discovery; the apply decision is requester-gated and affects retained reliance.); Stale-generation/aspect cleanup: known — automatic [wired] (Software cleanup is distinct from human-gated forget and externally called reset.); Debug-turn capture, exposed reset caller, initial hosted/local content and provider-internal admission: not-determinable (Generic debug principal does not establish human content control; reset caller and initial content control are not established; hosted internal admission is opaque.); Optional sandbox durable-content admission: uninspected (Capability execution and reuse do not establish per-content admission control for a durable memory inventory.)	partial coverage (The inventory includes requested model/human returns and independently supplied context. Both pull and push are wired; remote extraction consumption, complete checkpoint rendering, history selection and sandbox persistence retain local limits.); Semantic/tag/tree memory tools, skill load and human recent/export/graph requests: known — pull [wired] (Returning requested material is pull, not an additional push merely because code delivers it.); Ambient and style context, workspace directions and visible skill index: known — push [wired] (Compute independently supplies these retained parts before the foreground model requests them.); Later observation supplies persisted channel batch/spool or prior style carry to update consumers: known — push [wired] (Callback supply is automatic; the foreground answering model does not request these retained inputs.); Principal/thread checkpoint restoration and approval continuation: partial — push [wired], pull [wired] (Checkpoint restoration is independent context supply; approval interaction requests a specified continuation. Full investigation rendering remains unresolved.); Scheduled instruction callback: known — push [wired] (Stored timing independently starts later execution; any memory tool calls within that execution are separate pull routes.); Research planning history and explicit stored-draft publication: partial — push [wired], pull [wired] (Stored bodies automatically guide planning; explicit send requests delivery of a selected body. Full history storage/selection horizon is not completely inspected.); Entity cache lookup: known — pull [wired] (Calling entity resolution requests the matching retained identification; the web fallback is acquisition rather than another memory read direction.); Remote extractor use of stored context/configuration and internal memory: not-determinable (Settings are submitted, but remote later selection/delivery operations are not traced in accepted records.); Optional reused sandbox durable state: uninspected (Provider lifetime and subsequent workspace-content consumption remain uninspected.)	partial coverage (Signals classify automatic supply, not requested retrieval queries. Identifier targeting and coarse eligibility/budgets coexist; provider bucket/embedding decisions and complete history/continuation/sandbox alternatives remain unresolved.); Ambient profile supply to current turn: known — identifier [wired], coarse [wired] (Trigger is compute invocation. Selector uses readable scopes, asker/mentioned-person/channel identifiers and bucket/liveness/tag eligibility; retained outputs are short derived-memory text plus a topic outline.); Style facet supply: known — identifier [wired], coarse [wired] (Trigger is invocation; selector inputs are asker tag, agent-self container, recognized bucket keys and live recent 80 rows; selected parts are grouped style-note texts.); Standing workspace text and visible skill index: known — coarse [wired] (Trigger is later context assembly. Workspace availability and active/visible scope select standing text/index metadata. Requested body load is not a push signal.); Checkpoint restoration: partial — identifier [wired], coarse [wired] (Trigger is later compute; exact thread and principal tuple match select a retained checkpoint, with six-hour TTL/schema checks. Full downstream rendering remains uninspected.); Channel batch/spool and prior observer carry supply: known — identifier [wired], coarse [wired] (Trigger is owned observation callback. Scope/cursor and oldest bounded pending batch select retained messages/clusters; the style observer receives its bounded prior-note tail.); Scheduled callback supply: known — coarse [wired] (Trigger is stored time/cron. Callback receives its retained payload; workspace/org/entitlement gates bound execution. No topical semantic selector is established.); Research planner history supply: partial — coarse [wired] (Trigger is a new round; selector supplies prior sent/pending bodies normalized to 240-character heads. Complete history-selection criteria/horizon remain unresolved.); Requested semantic/tag/tree/skill/human/entity/approval/draft returns: inapplicable (These operations fulfill an explicit consumer request. Their lexical/semantic/identifier query predicates are not push-signal findings.); Remote bucket assignment, embedding-based alternatives and extraction-context consumption: not-determinable (Local matching of returned keys is known; opaque production/selection of provider payloads is not thereby classified as inferred-embedding or inferred-judgment.); Optional sandbox state supply: uninspected (Reuse lacks an inspected durable-content selector.)	partial coverage (Automatic trace-fed retained outputs have later consumers, establishing wired capacity rather than improved behavior. No system-wide no is asserted. Unresolved provider routes, scheduler/configuration provenance and sandbox continuity prevent complete coverage.); Interactive/resumed memory capture and channel distillation: known — yes [wired] (Conversation/tool activity feeds automatic durable text admission and later answering/planning context. Humans authoring inputs or requesting work do not make admission manual.); Public-thread style observation and carried notes: known — yes [wired] (Human interaction evidence and carried notes feed automatic retained guidance delivered to later turns.); Approval conversation compaction: known — yes [wired] (Automatic reduction of conversation/tool traces is retained in resume storage and supplied to a later invocation; deterministic reshaping still qualifies.); Investigation checkpoint: partial — yes [wired] (Automatic trace-fed save and later state restoration are wired. Full final rendering remains uninspected, so coverage is partial rather than a claim of every excerpt influencing the model.); Research draft/history production: partial — yes [wired] (Automatic retention follows an isolated work turn, and prior bodies reach later planning. Full history storage/selection horizon is unresolved; send approval is separate from generation/storage.); Human-approved skill admission, human HTTP editing and explicit reviewer replacement/publication: inapplicable (These separately classified mechanisms require human content decisions, so they are not automatic trace-fed admissions. This is not an evidenced no for the system.); Fixed-policy/field-based extraction provisioning and index repartition: inapplicable (Known transformations use policy/organization fields or retained documents to configure access/extraction; the records do not establish a trace-fed durable behavioral update at these mechanisms.); Model-mediated standing configuration and scheduled instruction derivation, debug caller and hosted internal extraction: not-determinable (Automatic configuration/schedule admissions exist, but recorded provenance does not settle which inputs are traces; generic debug principal and opaque hosted transformations prevent complete classification.); Optional sandbox and other evaluator-driven durable revisions: uninspected (No accepted record traces a retained sandbox/evaluator-feedback update to a durable later behavior-shaping artifact. Live administrative feedback interpretation is not such a write.)	partial coverage (Authority is assigned to a retained part at an actual consumer. Knowledge, instructions, learning inputs, selectors and controls differ. Remote extractor effects, some checkpoint rendering and sandbox consumption remain unresolved; none establishes compliance or benefit.); Selected document/derived text and research history at later answering/planning models and human readers: known — knowledge [wired] (Facts and past outputs supply reference/context; history also guides avoidance of repeated topics, without a factual acceptance criterion.); Cached entity identification at later reference-resolution consumer: known — knowledge [wired] (Canonical identification supplies reusable reference knowledge; parse validity is not truth validation.); Style facets, workspace prompt, loaded skill body and scheduled instruction at later model invocations: known — instruction [wired] (These standing directions request voice, workflow or task behavior; they are not executable enforcement.); Retained channel spool and prior observer carry at durable-note update consumers: known — learning [wired] (Recorded activity/experience feeds durable behavior-shaping content. This update-consumer force is separate from knowledge/instruction at the output consumer.); Tag key/label/description/use counts and returned similarity inputs at recall prioritizers: known — ranking [wired] (Lexical matches and usage counts rank retained metadata; local sorting consumes service scores without establishing how the service computed them.); Node mappings, observation cursors, identity scope, skill visibility and stored schedule timing/destination at host selectors: known — routing [wired] (These retained fields select input windows, accessible parts, operations or destinations; merely unconsumed metadata has no classified force.); Stored reset epoch, memory liveness flags, skill status/scope and draft status at admission/read gates: known — validation [wired], enforcement [wired] (Retained criteria determine admissibility and implement local vetoes. Service-wide authorization and deletion do not follow.); Investigation checkpoint and approval-resume retained excerpts at restored turn state/model: partial — knowledge [wired] (Resume consumption is inspected; investigation state is restored but its full final rendering remains uninspected.); Extraction context at hosted extractor: partial — instruction [claimed] (Provisioning is wired, but actual remote interpretation, criteria and transformations are opaque.); Research draft evidence/length flags at reviewer: known — knowledge [wired] (Advisory diagnostic text is not a machine-enforced acceptance criterion.); Optional reused sandbox memory and other opaque provider-state consumers: uninspected (No accepted record inventories authority for internal embeddings or accumulated sandbox files.)	AAS-2026-10-14-company-brain-30a9e8cfaeca-01
Compound Engineering Plugin	2	Use-accumulated project solutions/patterns, writable or imported pack rules, CONCEPTS vocabulary, prior-session derivatives, continuity snapshots, optimization logs/digests, implementation receipts and recovery state, retained review/residual and evaluation evidence, and revised methodology. Includes unresolved retune, dogfood, ideation, live-polish, pending-feedback, sweep and additional glossary/solution consumer routes, plus opaque host compaction and injected auto-memory dependencies. Excludes unchanged shipped skills and distribution copies as memory, ordinary untransformed current-run context, arbitrary product code as a learning store, and external host/model implementation. Provider memory and trace provenance are unresolved, not inferred from the fact that a model uses weights.	partial coverage (Principal file and Git surfaces are supported; external destinations and named unresolved dependencies prevent complete coverage.); Solutions/patterns and project glossary; live writable pack files: known — files [claimed], repo [claimed] (These named objects have file retention and conditional Git retention; actual project contents are not supplied.); Git-sourced pack cache: known — files [wired], repo [wired] (Cache substrate is known even though third-party content and ref freshness are not.); Session skeleton/digest, managed handoff, optimization log/result markers and strategy digest: known — files [claimed] (Classification covers named local representations, not every arbitrary user-selected handoff publication surface.); External implementation manifest, authorization/receipts and Git recovery objects: known — files [wired], repo [wired] (This is operational memory, not an assertion that all product commits are learned guidance.); Saved review and evaluation evidence; revised repository methodology: partial — files [wired], files [claimed], files [afforded], repo [afforded] (PR-body/tracker destinations and arbitrary handoff destinations have no fully classified storage implementation in these records.); Opaque host memory/compaction archives and ancillary retune, dogfood, ideation, live-polish, pending-feedback and sweep state: uninspected (Reports name included dependencies and ancillary state but do not resolve all operative parts or provider retention. Complete substrate enumeration is prevented.)	partial coverage (Automatic and conditional manual admission coexist; unknown initialization and method/provider admission prevent completeness.); Ordinary learning/glossary capture, session enrichment and model-admitted safe refresh actions: known — automatic [claimed] (Starting capture or choosing a destination does not make all resulting content admission manual.); Conditional human-approved refresh changes and discoverability instruction edits: known — manual [claimed] (Human control applies only to the named gated edits, not the whole store or every pack write.); Pack-cache admission and controller authorization/receipt writes: known — automatic [wired] (Source configuration initiation remains separate from these software admission mechanisms.); Handoff content, experiment logs/digests and review residual composition: known — automatic [claimed] (No mandatory per-summary or per-measurement human content approval is described.); Saved grading results: known — automatic [wired] (Saving evidence does not automatically admit a proposed revised skill.); Initial stores/config, contributor method-change admission, host state and ancillary write mechanisms: not-determinable (Pre-existing control and the human/agent contributor or merge decision are unresolved. Ancillary and provider writers lack sufficient admission details.)	partial coverage (Pull is supported on inspected/requested chains; unresolved host and ancillary paths prevent a pull-only system conclusion.); Planning/review requests for solutions, packs and canonical terminology, including researcher return: known — pull [claimed] (Automatic relevance selection within a requested research return does not create an additional push operation.); Pack cache lookup: known — pull [wired] (A URL/ref identity on this requested lookup is not identifier-based push.); Archived-session enrichment, user-selected handoff and experiment resume/next-batch reads: known — pull [claimed] (Automatic probing and model ranking remain inside their requested workflows; no unsolicited delivery is established.); Same-plan receipt recovery, later residual/evaluation consultation and revised-method loading: partial — pull [claimed], pull [afforded] (Full recovery and publication/loading of changed methods are not fully traced. A possible host injection cannot be classified from static registration.); Host-injected memory/compaction and ancillary retune, dogfood, ideation, live-polish, feedback and sweep readers: uninspected (Available summaries and search matches do not establish request/selection/delivery ownership for every alternative. Push remains unresolved, not absent.)	uninspected coverage (No positive push signal is established. Known pull units are locally inapplicable, but unresolved alternatives make the axis uninspected rather than absent or globally inapplicable.); Resolved pull operations: learning research, cache lookup, historical capture, handoff and experiment reads: inapplicable (Signals classify push selection, not lexical/relevance selectors or identifiers used to fulfill these explicit memory requests.); Resolved requested receipt and residual/evaluation consultation: inapplicable (Exact run IDs and artifact paths belong to requested return/recovery, not targeted unsolicited delivery.); Unresolved changed-method loading, host injection/compaction and ancillary delivery alternatives: uninspected (No recorded complete push chain names trigger, selector input and delivered retained parts; stronger inapplicability or a positive signal would hide missing coverage.)	partial coverage (Three qualifying prescribed automatic trace-fed chains establish yes. No bounded absence record supports a local or system-wide no; unresolved alternative writers prevent completeness.); Automatic prior-session enrichment into durable solutions with later planning/review consumption: known — yes [claimed] (Wired extraction alone does not upgrade the end-to-end model update/consumer chain. No improved-capacity result is required or claimed.); Model-generated continuity handoff consumed by a subsequent orientation invocation: known — yes [claimed] (User initiation/source choice is not per-content admission; temporary managed retention can still span later invocations. The snapshot need not generalize.); Automatic batch-outcome digest used by subsequent hypothesis generation: known — yes [claimed] (Human scope/spend approval does not approve each digest; no observed useful activation or causal improvement is supplied.); Repository method authoring/retuning, direct current-session capture, opaque host updates and ancillary summaries: not-determinable (Feedback-responsive method edits exist as a pathway, but automatic content admission and the complete trace-to-later-consumer chain are unresolved on these alternatives. Current work is not automatically an archived trace.); Imported static pack rule acquisition alone: inapplicable (This named acquisition consumes external rule files, not activity traces; upstream pack-learning provenance is separately unresolved, not a negative about its authors.)	partial coverage (Knowledge, learning, ranking, routing, validation, enforcement and instruction have local warrants; opaque alternative consumers prevent completeness.); Solutions/rules/glossary and distilled findings at planning and review consumers: known — knowledge [claimed], ranking [claimed] (Model-interpreted evidence does not establish enforced policy or measured activation.); Prior-session evidence and experiment outcomes at their durable update consumers: known — learning [claimed] (Learning authority names the update path, not improved future capacity.); Handoff snapshot and optimization strategy digest at later orientation/hypothesis invocations: known — knowledge [claimed] (These downstream effects remain separate from learning authority at their producers.); Experiment metrics/status and result markers at comparison and recovery consumers: known — ranking [claimed], routing [claimed] (Declared comparison/recovery consumers are supported; uninspected decide code is not an enforced guarantee.); Controller authorization/receipt state at executable authorization and integration consumers: known — validation [wired], enforcement [wired], routing [wired] (These forces govern the inspected protocol, not arbitrary shell actions or a universal sandbox. Full resume alternatives remain outside this resolved unit.); Review/residual and grading artifacts at caller, human reviewer and methodology author: known — knowledge [claimed] (Current-run helper confidence partitions are not generalized standing controls; method-update admission remains separate.); Use-revised methodology at future author and host-model consumers: known — instruction [afforded] (Unchanged shipped skills are excluded from the memory inventory; revised retained instructions are included.); Opaque host memory, ancillary state, additional readers and full recovery alternatives: uninspected (Actual consumed parts and consequential effects are not fully recorded; no complete authority inventory follows.)	AAS-2026-10-13-compound-engineering-plugin-6ac79f4e7555-01
Continuity	2	Included: accumulated local facts and metadata/history/projects/links, learning observations and narratives, saved messages and research products, human profile/custom instructions, journal/canvas/database/briefing state, vault files and derived index, organization rows, Threads posts/drafts/results/activity state, documented developer memory-bank updates, and unresolved provider-retained memory alternatives. Excluded: static shipped runtime prompts, current-run model/tool protocol and research windows, plugin session registrations/capability maps, volatile sync queues/cursors as current-run transport state, general runtime configuration/credentials, historical plugin database contents, external service/client implementations and actual external vault bodies. Provider inference weights are not positively classified as accumulated memory merely because they are parametric.	partial coverage (The accepted inventory covers local memory/app databases, organization and plugin databases, vault files and developer documents. Provider-retained state remains opaque; storage of that alternative cannot be inferred from remote inference.); Local memories, access metadata, revisions, projects/links, observations, narratives and the stored vault index: known — sqlite [wired] (These retained objects reside in local SQLite memory.db; typed edges do not establish a separate graph storage engine.); Saved messages, human profile settings, journal/canvas/database blocks and briefing state: known — sqlite [wired] (These parts reside in app SQLite, including research report blocks and summary messages.); Organization memory rows: known — sqlite [wired] (Organization SQLite storage is distinct from local memory.db; the pull-target defect does not erase server storage.); Threads retained posts, scores/reasons, drafts, engagements and search logs: known — sqlite [wired] (The plugin persists these objects in threads-engagement.db.); Vault files written and later read through the shipped vault tools: known — files [wired] (The tool path persists file content separately from its SQLite overview. Actual external files and scanner internals were not inspected.); Developer-agent current-context and progress Markdown: known — files [claimed], repo [claimed] (The documented memory bank uses Git-tracked Markdown files; update/reload execution is only declared.); Possible retained provider-internal state: uninspected (Remote parameter storage and any provider-side retained state are inaccessible; no memory substrate classification follows.)	partial coverage (Software/model admission is automatic even after manual launch. Generic MCP/HTTP caller identity does not prove human per-content control; explicit human editing is classified only on its own surface.); Desktop remember/forget model dispatch: known — automatic [wired] (Code persists model-proposed content without an item-specific human confirmation gate.); Scheduled or manually launched narrative replacement: known — automatic [wired] (Model output, including fallback text, is admitted automatically. Synthesize initiation is not output approval.); Scan-to-vault-index projection and model-requested file/index updates: known — automatic [wired] (A scan initiation or approval request in prompt text is not per-output content approval in the executor.); Conversation saves and research completion report/summary writes: known — automatic [wired] (The driver and callbacks admit output without per-content review; callback success is not observed.); Human profile/custom-instruction save and journal block edit: known — manual [wired] (An explicit human decision supplies/edits this retained content; debounce/SQL does not change admission agency.); Threads scheduled post/score/draft and outcome/log admission: known — automatic [wired] (Code admits generated drafts and records activity/outcomes without individual human content approval.); Threads draft editing/selection through the intended human review surface: known — manual [afforded] (The review UI affords a human per-content edit/choice, distinct from any generic endpoint caller.); MCP memory/project/link/learning/narrative admission and generic Threads edit/select endpoint callers: not-determinable (Caller-supplied content and available endpoints do not identify whether a human selected that content or software did.); Organization server upsert and plugin pull admission: known — automatic [wired] (Code checks schema/timestamp and applies rows without per-content human approval; intended local database integration is defective.); Developer document update control and incompletely traced canvas/database/journal-auto-link producers: not-determinable (Maintenance instructions do not establish content approval control; complete alternative workspace admission routes are not recorded.); Provider-retained memory admission: uninspected (The inference adapters do not expose provider memory/weight write control.)	partial coverage (Explicit requests and independent context supply coexist. Selecting a stored identifier to fulfill a request remains pull; automatic selection inside prompt construction is push. Untraced workspace/provider alternatives prevent complete coverage.); Desktop recall/search/project and external MCP memory/history/relationship/narrative/learning readers: known — pull [wired] (These readers explicitly request retained material; returned tool/API data are not a second push.); Per-send passive facts, profile and global narrative injection: known — push [wired] (Prompt code supplies retained content without a request for those parts; failures can omit it.); Narrative synthesis input selection and automatic delivery to its model: known — push [wired] (After an initiation trigger, code selects active facts, pending observations and prior narrative for a tool-free model call.); Dedicated connected-vault index prompt supply: known — push [wired] (Prompt construction automatically selects the retained index, outside the general memory budget.); Requested vault file/path/search/folder reads: known — pull [wired] (The model requests retained file content; its chosen path is a pull argument.); Thread selection/reload and subsequent history supply to the foreground model: known — pull [wired], push [wired] (Thread reload is requested by the UI; later send construction independently supplies saved role/content as model history.); Journal date selection, memory browser/history, Threads review/approval and server org pull/search requests: known — pull [wired] (These subsequent consumers request retained work or rows. Organization frontend model discovery remains unresolved.); Threads scheduled retained-post dedup and search-budget reads: known — push [wired] (The scheduler independently selects retained activity state for later scan control, rather than returning a caller-requested memory payload.); Developer instructed fixed startup pair and specialized-file requests: known — push [claimed], pull [claimed] (Doctrine specifies automatic startup supply plus on-demand specialized context. Host execution is not established.); Complete canvas/database model context, briefing/journal-auto-link selection and org frontend prompt/tool delivery: uninspected (Known UI/API requests do not resolve all included automatic consumer chains.); Provider-internal read-back: uninspected (Request construction is inspected, but opaque provider memory retrieval is not.)	partial coverage (Signals apply to automatic selection, not to requested lookup arguments. Coarse availability/budget filtering remains distinct from identity targeting. Workspace, arbitrary metadata and provider alternatives are unresolved.); Per-send passive fact push: known — coarse [wired] (Trigger: each prompt rebuild. Inputs: active state, recency, type and budget. Selected part: recent top-50 fact content, grouped by type with a nominal character cutoff; no current-topic selector.); Profile/custom instructions, latest global narrative and dedicated vault-index push: known — identifier [wired], coarse [wired] (Trigger: prompt rebuild. Exact saved setting keys, global scope and obsidian_vault_index select delivered retained parts; nonempty/connected conditions also gate availability.); Narrative update input push: known — identifier [wired], coarse [wired] (Trigger: scheduled/manual synthesis. Scope selects pending observations and prior narrative; active-row availability, pending state and snapshot checks select input sets. All active memories across scopes also enter.); Saved thread history push to a later foreground invocation: known — identifier [wired] (Trigger: subsequent send. Upstream selected thread ID selects the retained message set mapped to role/content; current-run protocol state is not the delivered payload.); Threads later scan state push: known — identifier [wired], coarse [wired] (Trigger: scheduled scan. Post-ID matches choose retained duplicate state; a rolling time window and configured budget choose retained search counts. These are control inputs, not model relevance retrieval.); Developer startup context push: known — identifier [claimed] (Trigger: next development conversation. Doctrine fixes ACTIVE_CONTEXT/CLAUDE filenames as the selected retained startup parts; actual host delivery remains unobserved.); Untraced canvas/briefing/journal automatic selection, arbitrary metadata and org frontend delivery: uninspected (No complete trigger/input/selected-part chain is recorded for these alternatives.); Provider-internal push selectors: uninspected (Opaque provider memory cannot be classified as embedding retrieval from parametric computation alone.)	partial coverage (Automatic conversational and tool-feedback fact writes qualify locally, as does their active-fact-fed narrative continuation; documented development summaries qualify at claimed strength. Other narrative input origins, research/Threads origins, workspace alternatives and provider state remain unresolved. This is not demonstrated improved capacity or a system-wide negative.); Conversation-fed desktop remember to retained facts to later context/recall: known — yes [wired] (Conversation activity feeds an automatic model-proposed durable fact write with a wired later model consumer; usefulness and faithful extraction are unobserved.); Tool-feedback-fed desktop remember to retained facts to later context/recall: known — yes [wired] (This classifies the implemented feedback-to-fact-write branch, not raw tool-result retention or observed use.); Desktop trace-derived active facts at automatic narrative production and later briefing delivery: known — yes [wired] (The known trace-derived fact branch is separate from generic observations, other facts and prior briefing of unresolved original provenance.); Other observation/fact/prior-briefing origins at automatic narrative production: not-determinable (Automatic production and later briefing delivery are wired, but generic observation provenance and absent inspected desktop extractor do not establish original trace input for every included synthesis path.); Developer task-summary update and next-conversation context: known — yes [claimed] (Doctrine directs the coding agent to retain work-derived summaries and reread them; admission execution is not established and no improvement is inferred.); Research report/summary and Threads generated draft/reason paths: not-determinable (Acquired source/post text and generated work products do not by themselves establish an original activity-trace-fed behavior-shaping update. Retained outcome logs have no recorded feedback-to-guidance update.); Explicit human profile/journal edits and ordinary static-content vault scan acquisition: inapplicable (The named human edits are not automatic admission; the recorded scan input is file metadata/excerpts, not an established activity trace. This excludes neither model-written files nor unresolved auto-link alternatives.); Generic MCP narrative/fact/observation writes, organization imports and other canvas/journal-auto-link updates: not-determinable (Caller agency and original inputs or later model-consumer chains are unresolved; optional IDs and event wrappers alone are not trace-learning evidence.); Provider-internal trace-fed adaptation: uninspected (No provider update/retention route is exposed by inference adapters.)	partial coverage (Authority belongs to each retained part at its consumer, not to the database as a whole. Text delivery is not compliance; organization/plugin discovery and opaque workspace/provider consumers prevent complete enumeration.); Active fact text and vault overview at foreground/tool-result readers; revisions, projects and links at requested UI/MCP readers: known — knowledge [wired] (Retained content supplies context/reference or advisory relationship/history display; typed relations do not adjudicate truth.); Passive preference/constraint content and human custom instructions at the foreground model: known — instruction [wired] (The system-message path supplies behavioral directions through natural-language interpretation, without deterministic enforcement.); Narrative briefing at foreground and prior-content reader; observation text at synthesis: known — knowledge [wired] (These parts supply context and evidence-like input. Self-reported confidence is not factual validation.); Retained observations at automatic narrative update consumer: partial — learning [wired] (Conversation-derived observations can shape a durable briefing, but optional source IDs do not establish the origin of every admitted observation.); Saved conversational content at later foreground sends, including the model remember proposal consumer: known — knowledge [wired], learning [wired] (Later history supplies context and can feed automatic durable fact writes; raw storage alone would not establish the update effect.); Archive state at active recall/passive/synthesis selection: known — enforcement [wired] (The retained archive flag enforces exclusion from these active reads, not derivative forgetting.); Retained memory recency/type fields at passive top-50/type-order selection: known — ranking [wired] (Retained features influence relative delivery priority under fixed code.); Retained scope/identity fields, learning pending state and narrative snapshot at retrieval/synthesis selection: known — routing [wired] (IDs/scopes select retained inputs; pending count/hash can select or suppress subsequent synthesis.); Existing retained memory endpoints at relationship admission: known — validation [wired], enforcement [wired] (Stored endpoint existence is checked as a candidate-edge admission criterion and absence blocks the edge. This validates identifiers, not relation truth.); Workspace prose/projections and retained Threads drafts/reasons at human review or later chat history: known — knowledge [wired] (These consumers receive reference material or candidate work; no endorsement follows from availability.); Threads stored draft-selection flags and post IDs at later approval/dedup consumers: known — routing [wired], enforcement [wired] (Selection state routes the draft for publication; stored post identity prevents duplicate queue work.); Threads retained search-log counts at scan budget checks: known — enforcement [wired] (Recorded search activity blocks scans beyond the configured rolling budget.); Organization retained timestamp at server conflict admission: known — validation [wired], enforcement [wired] (Existing updated_at is the validation criterion for a candidate replacement and enforces the timestamp guard.); Organization retained rows at requested server pull/search delivery: known — knowledge [wired] (The server returns knowledge; the defective desktop pull target and missing tool discovery do not prove foreground consumption.); Developer memory-bank summaries/directions at the instructed external agent: known — knowledge [claimed], instruction [claimed] (The documented reader receives task context and maintenance directions; actual loading/compliance is unobserved.); Arbitrary metadata consumers, richer canvas/journal-auto-link paths and provider-internal memory consumers: uninspected (Missing actual consumers/effects prevent additional authority classification.)	AAS-2026-10-14-continuity-3e490e1d39f0-01
CORAL	2	Included: retained island/public notes and raw captures, review sidecars, navigation/coverage material, mutable skills/resources and roles, attempt results and feedback, shared-state checkpoint history, migration summaries, session-pointer metadata and opaque resumed conversation alternatives; their admission, maintenance and later-consumer routes. Excluded: candidate code as a product rather than a memory abstraction, unchanged shipped instruction/skill text as such, all-runtime storage/permissions, provider parameters and external harness internals as attributed CORAL implementation. Host recovery effects stay included but unresolved.	partial coverage (Named CORAL stores are resolved, including requested capture/review stores. The included host conversation alternative remains opaque. Provider weights and all-runtime storage are excluded, not classified as memory substrates.); CORAL shared knowledge files, navigation, evolved skill resources, roles, attempts, migration announcements and session-pointer metadata: known — files [wired] (These named filesystem surfaces are established; this does not inventory external host payload storage.); Requested external-source captures and synthesis-review sidecars: known — files [claimed] (The prescribed locations are known, but no actual capture or review was supplied.); Shared-state historical snapshots: known — repo [wired] (Repository history is distinct from the files it snapshots; candidate product repositories are not added as memory solely because they store code.); Host-owned recovered conversation payload, separate from session pointers: uninspected (Payload creation/replay is external and uninspected; local JSONL lookup is not a complete substrate inventory of effective resumed state.)	partial coverage (Known software/model admissions coexist with unresolved initial-content and host admission. There is no supported system-wide manual/automatic exclusivity claim.); Model-mediated knowledge, skill, role, capture, review and semantic-maintenance admission: known — automatic [afforded], automatic [claimed] (Workflow initiation does not make subsequent model admission manual; prescribed semantic steps are not upgraded to wired by file-copy capability.); Attempt/checkpoint creation, grader finalization, index generation, arrival summaries and archival metadata updates: known — automatic [wired] (A human may initiate a command; these specified mechanisms admit their generated content automatically.); Initial/custom skill and role supply or replacement decision: not-determinable (Copy behavior is inspected, but configured file selection does not establish a human decision supplying or approving that retained content; no manual value is inferred.); Host conversation persistence and any summary admission: uninspected (External host creation and compression controls were not inspected.)	partial coverage (Both pull and push exist locally. Known file requests and heartbeat injection do not resolve host recovery or alternate skill-context supply.); Later caller-requested note/skill/raw/metadata/checkpoint access: known — pull [wired], pull [afforded] (Code-backed requested delivery and prompt-mediated model use keep separate warrants; dashboard requests are not broadcasts to all agents.); Manager-selected finalized-result feedback to the committing agent: known — push [wired] (The selected result plus action prompts is not automatic embedding of full notes/skills; batching prevents per-attempt delivery guarantees.); Recovered host payload and host skill-disclosure alternatives: uninspected (CORAL can request session recovery, but payload selection/delivery and host progressive-disclosure details were not inspected; no complete direction classification follows.)	partial coverage (Heartbeat supplies positive coarse and identifier push witnesses. Pull-only units are locally inapplicable, but unresolved host alternatives keep this axis partial rather than globally inapplicable.); Heartbeat result feedback selection and recipient targeting: known — coarse [wired], identifier [wired] (Coarse scheduling and identity-targeted delivery are distinct operations; no embedding or lexical push selector is established by CLI substring search.); Explicitly requested file/CLI/API/checkpoint returns: inapplicable (These units fulfill requests, so their paths, literal query text, date/status and numeric positions are pull selectors rather than push-signal values.); Host resume payload and skill-trigger selection: uninspected (The actual automatic trigger inputs and selected retained payload are not inspected; progressive-disclosure doctrine does not establish implemented inferred-judgment selection.)	partial coverage (Positive afforded and wired routes remain separate. Raw logging, snapshotting and inference alone do not establish trace learning. No bounded negative record justifies a local or global no; host and mixed-write alternatives remain unresolved.); Experiment/log/transcript-fed durable note, skill and role revision followed by later reads: known — yes [afforded] (This is the afforded trace-fed write/retention/later-consumer chain, not a claim of learned weights or observed improved capacity.); Automatic migration arrival-summary write and later knowledge query: known — yes [wired] (A coordination summary qualifies without distilling a successful technique; actual later model dependence is unobserved.); Research-only acquisition, raw attempt/checkpoint retention and structural helper output: not-determinable (These facts alone do not establish a qualifying trace-fed behavior-shaping transformation. The mixed research/revision boundary and uninspected helpers prevent a bounded all-routes no; their known feeds are classified separately above.); Host continuation-summary or compaction write followed by resume: uninspected (No inspected implementation establishes whether host traces produce a retained continuation artifact or only replay raw turns.)	partial coverage (Each positive names a retained part and actual or prescribed consumer. Unconsumed metadata has no inferred force; model endpoint weights, static policy files and current-run permissions are not substituted for memory authority. Host payload effects remain unresolved.); Retained notes, raw evidence, attempt feedback and review reports at later readers and revisers: known — knowledge [wired], knowledge [claimed], learning [afforded] (The update path is afforded, while file delivery is wired; neither status establishes improved capacity.); Evolved skills and role descriptions at foreground/specialist agents: known — instruction [afforded], routing [afforded] (These are prompt-mediated effects, not interpreter-enforced obedience; static shipped task prohibitions are excluded.); Attempt score/status/history at manager heartbeat and migration consumers: known — ranking [wired], routing [wired] (Numeric scheduling authority is distinct from the feedback body read by the model.); Archived-attempt flags and retained note access/navigation metadata at fixed readers: known — enforcement [wired], routing [wired] (Direct-hash recovery remains possible; supersession labels do not establish universal withdrawal. Literal note search is filtering, not evidenced relative-score ranking.); Saved research-source text at the prescribed synthesis reviewer: known — validation [claimed] (This is advisory model evaluation, not an automatic admission veto. Static lint rules are not retained memory criteria.); Recovered conversation content at the host and resumed model: uninspected (The pointer selects a session, but the complete consumed payload and its force remain uninspected.)	AAS-2026-10-15-coral-25e5fdab9dc4-01
Cortex	2	Accumulated Cortex document bodies/raw Markdown, summaries, symbolic metadata, graph objects/entities/relationships, SQLite vectors, access counters/tier changes, query trail, persisted weight and temporal-version capabilities, import dedup markers and whole-store archive recovery; their admission, maintenance and later service/internal consumers. Excludes shipped static prompts/ontology/inference rules, provider weights/state, external host conversation memory, current-response diagnostics and unretained synthesis narratives.	known coverage (The scoped inventory comprises accumulated dual-store parts and their archive recovery route. External provider internals, static packaged rules/prompts and current-response dictionaries are excluded, not counted as unknown memory stores.); Document text, summaries, classification metadata, embeddings, access keys, query log, persisted weight coefficients and document-row versions: known — sqlite [wired] (The accepted part inventory identifies the SQLite location of each part, including dormant weight/history capabilities.); Graph identities, entities, assertions, mentions and descriptive properties: known — graph [wired] (Graph state is a separate retained part from content-authoritative SQLite.); Operator-created whole-store backup and restoration artifact: known — files [wired] (This is whole-store recovery, not an additional graph/vector service.)	partial coverage (Automatic transformations and an explicit operator replacement coexist with unresolved generic-caller content decisions and weight initialization. Neither entry channel nor physical I/O establishes manual agency.); Per-content decision supplying original capture/edit/classification fields: not-determinable (Generic caller identity does not establish an explicit human decision or autonomous caller admission; upstream content control remains unresolved.); Importer mapping, dedup and per-item admission: known — automatic [wired] (An operator starting import is not manual admission of each mapped object; upstream file authorship remains separate.); Internal normalization, embedding replacement, link/entity admission, enrichment and inferred-edge writes: known — automatic [wired] (Includes optional model branches and deterministic rules; source-level wiring is not observed execution.); Read/feedback-derived counters and tier writes; query logging; before-update row snapshots: known — automatic [wired] (Caller action triggers the workflow but does not manually approve generated counts, logs or snapshots.); Explicit whole-store replacement by restore: known — manual [wired] (This separately evidenced replacement decision does not make ordinary pipeline admissions manual.); Nondefault weight-content admission: not-determinable (Setter/reset APIs and absent shipped adaptation callers leave the decision supplying nondefault values unresolved.)	known coverage (Accepted records inventory the shipped consumers as request-driven and explicitly distinguish dormant APIs from routes. External host injection and hypothetical library wiring are excluded. No independent push operation is established in this assessed service boundary.); Requested search/context/read/list/graph/export, dossier and synthesis input selection: known — pull [wired] (The ranking/projection operations are inside a requested chain, not additional unsolicited deliveries.); Requested status/reason alerts and diagnostic query trail: known — pull [wired] (Proactive wording does not establish an unsolicited notification route.); Internal reloads for capture processing, inference and access/tier maintenance: known — pull [wired] (No independent context-injection selector is established for these requested processing reads.); Temporal-version and persisted-weight getter capabilities: inapplicable (Read methods exist but no shipped later consumer is found. They are inventory capabilities, not alternate established read-back directions.)	inapplicable coverage (Inapplicable to this completely assessed direction inventory: all established read-back chains are pull, and the remaining inventoried getters have no shipped consumer. IDs and lexical/embedding selectors on a request are not push signals.); All established requested delivery and processing-read chains: inapplicable (The assessed chains are pull: query/ID/budget inputs select requested material, not independent push deliveries. Therefore no push selector signal applies.); Dormant weight/history getter APIs: inapplicable (No established later read-back route exists for these capabilities, so no push selection operation exists to classify.)	partial coverage (Access-derived adaptation supplies a positive witness. Generic content provenance leaves other possible trace-fed transformations unresolved; raw query retention and unwired weight helpers do not upgrade that witness to measured learning.); Successful read/positive-feedback activity to access counters, tier promotion and later briefing/enrichment: known — yes [wired] (Raw event reasons are not preserved; count/tier adaptation is wired, not observed benefit.); Automatic content-to-summary/vector/link routes with generic captured/imported source material: not-determinable (Inputs can be ordinary documents; no accepted record establishes whether particular retained inputs are session/tool/environment traces. A qualifying trace-fed route cannot be asserted or excluded for all these inputs.); Query-log retention and weight/history capabilities without an automatic update-consumer bridge: inapplicable (Raw diagnostic logging and unchanged snapshots alone do not qualify; searched absence of the shipped adaptation bridge prevents treating helper names as automatic training. No system-wide negative follows.)	partial coverage (Actual consumer effects are classified separately. Dormant persisted weights/history prevent complete consumer-authority coverage; no instruction priority is inferred from prose and no learning benefit from usage.); Body, summary and descriptive graph content delivered to requesting humans/agents and internal synthesis input: known — knowledge [wired] (Tool-result content is informational, not standing host instruction.); Document lexical content, vectors, graph connectivity and created-at metadata in hybrid retrieval: known — ranking [wired] (Tier/access_count and persisted weight setters are not silently added to this ranking consumer.); Type/project/entity/relationship metadata at filters, dossiers, diagnostics and inference selectors: known — routing [wired] (This is scope/input selection; the static rule bodies themselves are not accumulated memory.); Retained object identities and importer dedup markers at admission checks: known — validation [wired], enforcement [wired] (These retained criteria have actual reject/skip consumers, unlike unconsumed metadata. Static schemas/transport policy are excluded.); Access metadata at promotion, and promoted tier at subsequent enrichment: known — learning [wired], routing [wired] (The update consumer and later tier-preservation consumer have distinct effects. Tier presentation is informational, not a ranking coefficient.); Query-log entries at diagnostic trail consumers: known — knowledge [wired] (No automatic policy update is attributed to raw log display.); Persisted weight coefficients and historical row versions without shipped later consumers: not-determinable (Library get/read APIs do not establish an actual ranking, instruction, validation or recovery consumer. Their behavioral authority cannot be classified from these records.)	AAS-2026-10-17-cortex-3da14ce55a2b-01
CQ	2	Retained knowledge-unit insight, metadata, flags and arbitrary extensions; domain/FTS access structures and optional SQLite vectors; built-in and injected-store admission, maintenance and query/review consumers; Cursor failure-context records and discovery NodeInfo caches accumulated through use. Includes unresolved initialization, application-specific extension consumption and custom-store alternatives. Excludes static shipped/installed prompts and manifests, credentials, external host/model/provider weights and context management, transient query embeddings, and reflection intermediates not shown to be durably retained.	partial coverage (Inventory includes unit payloads and access structures, optional vectors, auxiliary failure state and discovery caches. Custom-store substrate and full alternative-store persistence are unresolved; a wired SQLite witness does not cover them.); Knowledge-unit insight, symbolic metadata, flags, extensions and domain/FTS access structures in inspected SQLite stores and remote SQL repository: known — sqlite [wired] (SQLite unit JSON, index rows and server SQL data are operative retained storage.); SDK PostgreSQL knowledge records and domain-query alternative: partial — rdbms [wired] (PostgreSQL storage is identified, but the accepted records do not exhaustively trace its durability and mutation paths.); Python in-memory store knowledge-unit alternative: partial — in-memory [wired] (Dictionary-value selection is inspected; all mutations and persistence parity are not exhaustively covered.); Injected custom Store implementations: uninspected (Interface injection is included; implementation substrate is not supplied or inspected.); Optional retained knowledge-unit embeddings: known — sqlite [wired] (The retained numerical output is SQLite storage, not the external provider weights.); Cursor per-conversation failure record: known — files [wired] (The bounded event state survives to a later stop hook.); Discovery NodeInfo disk and resolver memo caches: known — files [wired], in-memory [wired] (Both cache levels are described; ephemeral memo lifetime differs from disk TTL.)	partial coverage (Interactive per-content human choice is an affordance, not a universal software prerequisite. Known automatic maintenance is kept separate from unresolved generic caller control and initialized state.); Interactive candidate approval/editing before proposal: known — manual [afforded] (Human decisions supply or approve specific content under the shipped procedure; invocation alone is not that decision.); Headless agent proposal admission: known — automatic [afforded] (No human per-content decision is required in the headless path; model admission remains instructed, not exercised.); Generic direct proposal and Store content replacement caller: not-determinable (Generic caller authorship or identity does not establish human versus computational per-content choice.); Generated IDs/default evidence, supported feedback, drain, embeddings, indexes and format migration: known — automatic [wired] (Software determines these retained changes after requests; starting an operation does not make its transformations manual.); Browser review decision and automated seed review alternative: known — manual [afforded], automatic [wired] (Browser is a per-unit human affordance; seed approval is a separately wired programmatic mechanism.); Generic JWT-bearing review API caller: not-determinable (Authentication does not establish human presence or decision agency.); Cursor failure write and cache admission/cleanup: known — automatic [wired] (Software selects and retains event snippets and decoded endpoint metadata.); Pre-existing stores, vectors, caches and arbitrary extensions: uninspected (Initialization and actual extension admission principals are not established by later transformations.)	partial coverage (Explicit advice/review requests are pull. Cursor stop delivery and resolver-selected cache metadata are push operations within different chains. Additional custom-store/application consumer routes remain unresolved.); Knowledge query and requested review queue/get/list: known — pull [wired] (Selection/ranking inside a requested return is not an additional push.); Named-unit feedback reload and requested status views: known — pull [wired] (Retained unit/evidence reads serve the invoking operation; they do not automatically inject lessons into the host.); Cursor stop reminder: known — push [wired] (Stop is a lifecycle trigger, not a request for retained failure material.); Discovery cache selection within a requested remote operation: known — push [wired] (The outer remote operation is requested. Its resolver automatically supplies retained endpoint metadata rather than fulfilling a foreground request to read NodeInfo.); Extension application consumers and additional injected-store read-back: uninspected (Transport capability does not establish application-specific interpretation or additional later consumers.)	partial coverage (Both established push selectors use an actual identity match. Availability, latest-record retention and expiry remain separate filters, not extra semantic or embedding signals. Unresolved consumer alternatives prevent complete coverage.); Cursor stop selects retained failure by conversation identity: known — identifier [wired] (Trigger is stop; input is sanitized conversation key; delivered part is the latest bounded failure/error/input record.); Remote-operation resolver selects cached NodeInfo by configured address: known — identifier [wired] (Trigger is remote operation; input is normalized node address; selected retained part is matching NodeInfo. Memo precedence/disk TTL affect availability.); Explicit knowledge/review/feedback/status pull operations: inapplicable (These operations fulfill requests for retained material; their lexical/embedding ranking is not a push signal.); Opaque extension and injected-store additional delivery selectors: uninspected (No additional consumer selector is inspected; known pull routes cannot make these alternatives inapplicable.)	partial coverage (A clipped retained failure context qualifies under the profile definition without diagnosis or benefit. Interactive reflection lacks automatic per-content admission; headless/opaque experience routes and custom alternatives do not establish a complete trace-fed-write inventory.); Automatic Cursor failure-event write to later stop context: known — yes [wired] (Formatted/clipped event material is retained and supplies later host context. Event-to-record-to-stop wiring is sufficient; a generalized lesson or observed improvement is not required.); Interactive reflection session-to-proposal chain: not-determinable (Session extraction and later persistence are afforded, but the specified interactive path requires human content approval and actual host reasoning is unobserved. It establishes no qualifying automatic write or bounded system-wide no.); Headless candidate generation and asserted-experience feedback: not-determinable (Headless admission and automatic score changes are supported, but no accepted record identifies an original trace input actually consumed by these updates. Caller assertions are not themselves a recorded session/tool trace.); Embedding, transport-cache and imported/initialized custom-store alternatives: not-determinable (Known vector input is insight text and known cache input is endpoint discovery metadata; these facts do not establish qualifying trace-fed learning. Opaque initial/custom inputs prevent a complete negative.)	partial coverage (Authority is classified at actual consumers, not from field names or model compliance. Opaque extensions/custom stores and missing criticism part declarations prevent complete coverage.); Knowledge insight at querying host and browser reviewer: known — knowledge [wired], knowledge [wired] (Imperative action text is advice for the task, not standing host instructions or executable enforcement.); Insight text, domains/context/confidence and retained vectors at query rankers: known — ranking [wired] (Text and metadata drive lexical/relevance scoring; vectors drive distance preselection and final scoring. Benefit and truth are not implied.); Remote review status at query gate and pending-status admission check: known — enforcement [wired], validation [wired], enforcement [wired] (Query eligibility and eligibility for another review decision are separate consumers.); Retained tier and identifiers at feedback/drain operation selectors: known — routing [wired] (Only actually consumed metadata is assigned authority; timestamps and superseded_by receive none merely from storage.); Optional Go-local explanatory criticism at host reader: partial — knowledge [wired] (Prose can reach the host, but no independent criticism-part declaration resolves a complete part-addressed authority inventory.); Arbitrary extensions at application-specific consumers: not-determinable (JSON transport is established, but no semantic consumer/effect licenses an authority value.); Cursor failure strings at later host stop: known — knowledge [wired] (Automatic delivery supplies event context; actual corrective action remains unobserved.); NodeInfo at HTTP destination construction: known — routing [wired] (Endpoint reuse is transport behavior, not host instruction.); Recorded session experience at requested reflection-to-candidate consumer: known — learning [afforded] (This is an afforded experience-to-durable-advice update path; downstream lesson knowledge is separately classified.); Injected Store-specific consumers/effects: uninspected (Custom-store internals are not supplied; fixed built-in witnesses cannot cover their additional effects.)	AAS-2026-10-14-cq-8fb3dc41c334-01
CrewAI	2	Retained unified-memory text, metadata and embeddings; acquired knowledge chunks and opaque backend access state; training feedback/suggestions; Flow feedback lessons; overflow summaries; task replay logs, Flow snapshots/pending feedback and runtime checkpoints, including replaceable providers and unresolved file/tool-cache/retrieval-adapter consumers inventoried by the accepted account. Includes writes, maintenance and later-consumer routes. Excludes static shipped skills/instructions, ordinary untransformed turn/task-context accumulation, generic runtime permissions, model weights not written by these routes, remote hosted internals and unrelated application storage. Restored use-derived state and compacted summaries remain included despite those exclusions.	partial coverage (Built-in retained objects have positive substrate findings. Replaceable providers, knowledge persistence and additional adapter/cache surfaces prevent complete coverage; external model weights are not use-written memory in this boundary.); Built-in unified-memory content, metadata and embedding fields: LanceDB and Qdrant Edge: known — vector [wired], files [wired] (Built-in backing objects are identified; merge/deletion and cross-process durability guarantees are not implied.); Supplied unified-memory backend and process-wide factory replacement: uninspected (Backend replacement is explicit; the operative storage behind an arbitrary supplied implementation is not inspected.); Trained-agent suggestions and raw training files: known — files [wired] (The training file substrate is separate from provider weights.); Overflow summary in executor messages: known — in-memory [wired] (Within-run retention is included even without a file write.); Default Flow snapshots, pending-feedback rows and task replay outputs: known — sqlite [wired] (Encoding as JSON does not replace the SQLite substrate.); Built-in runtime checkpoint providers: known — files [wired], sqlite [wired] (Provider substrate identity is recorded despite incomplete implementation coverage of maintenance.); Knowledge RAG client and custom knowledge/checkpoint/Flow persistence: uninspected (Client collections and replaceable providers are recorded, but their underlying retained substrate is unresolved. A client call is not enough to infer service-object or vector persistence.); Additional file-input, tool-cache and retrieval-adapter retained parts and their later consumers: uninspected (The accepted account inventories but does not trace these memory-detail alternatives. Their storage substrate cannot be inferred from the general tool/cache route or from unified-memory witnesses.)	partial coverage (Automatic derived and recovery-state admissions are wired. No explicit-human per-content decision is established for the generic caller routes, so manual is not asserted; those included mechanisms remain unresolved.); Runtime extraction/encoding, consolidation and generated access fields: known — automatic [wired] (Starting an execution does not make downstream admission manual.); Feedback lesson distillation, training evaluator suggestions and overflow summary replacement: known — automatic [wired] (The inspected admissions are automatic even when human feedback triggers them.); Runtime checkpoints, Flow snapshots and replay-log production: known — automatic [wired] (A caller choosing a checkpoint/replay target is a read choice, not evidence of manual replacement admission.); Explicit remember/tool writes, update/forget/reset, configured knowledge sources and initial state: not-determinable (Caller supplies or decides content but its human/software identity and per-content approval are unresolved. Do not infer manual from API invocation or automatic from physical software I/O.); Custom-provider admissions, pending-feedback continuation and TUI editing alternatives: uninspected (These admission branches are only inventoried or incompletely traced; no per-content control classification is available.); Additional file-input, tool-cache and retrieval-adapter retained parts and their later consumers: uninspected (The accepted account inventories but does not trace these memory-detail alternatives. Their write agency cannot be inferred from the general tool/cache route or from unified-memory witnesses.)	partial coverage (Both request fulfillment and automatic supply exist, with operation boundaries stated. Opaque downstream consumers and additional cache/adapter paths prevent complete direction coverage.); Unified-memory runtime caller, Search memory and feedback pre-review requests: known — pull [wired], push [wired] (Separate operations in each chain: requested recall return is pull, automatic context selection for the later model is push. Public API capability alone adds no unnamed consumer.); Knowledge query and task-prompt delivery: known — pull [wired], push [wired] (Generated query and prompt injection are distinguished from the storage query response.); Trained suggestions and overflow summary delivery: known — push [wired] (Neither later model explicitly requests this retained content.); Explicit checkpoint/replay and Flow UUID restore operations: known — pull [wired], push [wired] (Pull names requested restoration, not an extra push merely because bytes are returned. Replay downstream context selection is a separate operation.); Custom storage/callers, provider-specific compaction and pending-feedback application consumers: uninspected (The inspected returns do not establish all later consumers or whether opaque content is requested or automatically supplied.); Additional file-input, tool-cache and retrieval-adapter retained parts and their later consumers: uninspected (The accepted account inventories but does not trace these memory-detail alternatives. Their read-back direction cannot be inferred from the general tool/cache route or from unified-memory witnesses.)	partial coverage (Positive signals belong to named automatic supplies. No inferred-lexical witness is recorded. Pull recovery identity matching is not counted as push, and opaque alternatives are retained rather than made inapplicable.); Automatic task/standalone memory and Flow pre-review push: known — inferred-embedding [wired], inferred-judgment [wired], coarse [wired] (Short-query/shallow paths do not inherit the deep judgment basis. Source/private is a visibility predicate, not authenticated lesson identity.); Knowledge snippets automatically appended at task preparation: partial — inferred-judgment [wired], coarse [wired] (Opaque RAG/custom selector internals and uninspected parsers prevent deciding every signal; no embedding or lexical mechanism is inferred solely from the RAG label.); Role-keyed trained suggestions automatically appended to task prompt: known — identifier [wired] (This is an actual retained-part identity match, not merely an identifier on a requested file.); Overflow summary supplied to next model invocation: known — coarse [wired] (The producer uses a model, but model production is not automatically inferred-judgment selection among already-retained summaries.); Stored replay outputs supplied as later task context: known — identifier [wired], coarse [wired] (This covers downstream scheduler push. Requested checkpoint filenames and UUID reloads alone establish no identifier-based push.); Opaque restored payload delivery, custom knowledge/provider selectors and pending-feedback continuation: uninspected (Requested recovery reads do not reveal all downstream push selectors; mixed direction coverage forbids treating these alternatives as pull-only inapplicable.); Additional file-input, tool-cache and retrieval-adapter retained parts and their later consumers: uninspected (The accepted account inventories but does not trace these memory-detail alternatives. Their push selection signals cannot be inferred from the general tool/cache route or from unified-memory witnesses.)	partial coverage (Four automatic trace-to-retained-artifact-to-later-consumer chains support yes. No bounded absence record establishes a local no, and uninspected writes prevent a system-complete trace-learning inventory.); Task/standalone trace extraction and full-input fallback through encoding to later recall: known — yes [wired] (The route is wired even when extraction fails into a full-content memory; improvement and observed activation are not required or claimed.); Crew training feedback to durable suggestions and later role-selected instructions: known — yes [wired] (Human feedback does not make derived admission manual.); Flow feedback distillation to retained lessons and later pre-review: known — yes [wired] (This is the traced learn-enabled branch, not complete pending-feedback continuation coverage.); Overflow history to retained continuation summary: known — yes [wired] (Within-run retention meets the continuation-summary rule; provider-weight updates are not asserted.); Explicit API/tool writes, recovery-state application transformations and custom-provider routes: not-determinable (Generic caller provenance and arbitrary application consumers do not establish whether these mechanisms produce qualifying trace-fed artifacts. Raw logs/checkpoint copies alone are insufficient.); Pending-feedback async continuation and provider-specific summaries: uninspected (Inspected positive branches do not settle alternate write and later-consumer chains.); Additional file-input, tool-cache and retrieval-adapter retained parts and their later consumers: uninspected (The accepted account inventories but does not trace these memory-detail alternatives. Their trace learning cannot be inferred from the general tool/cache route or from unified-memory witnesses.)	partial coverage (Authority is classified at each actual update, ranking, filter, reconstruction and model consumer. Unknown application/provider payloads prevent a system-complete union; configured static skills and generic guardrails are excluded from memory authority.); Recalled memory/knowledge text at task and tool-model consumers: known — knowledge [wired] (Delivery establishes context authority, not observed semantic influence.); Stored text/vectors and metadata at retrieval ranker and selectors: known — ranking [wired], routing [wired], enforcement [wired] (Ranking and selection consumers differ from downstream text interpretation. Consolidation can lose source/private fields.); Retained old memory content at encoding review/update consumer: known — learning [wired] (Learning authority is local to the durable artifact update, not a claim of improved capacity or factual acceptance.); Recorded training outputs/feedback at evaluator and trained suggestions at later task model: known — learning [wired], instruction [wired] (The raw update input and derived instruction output have distinct consumers and forces.); Flow feedback and retained lesson content at distillation/pre-review: known — learning [wired], knowledge [wired] (The reviewer may receive public non-lessons too. No exclusive lesson lineage or mandatory compliance is inferred.); History at summary producer and retained summary at next model invocation: known — learning [wired], knowledge [wired] (These are separate update and downstream paths, not provider-weight learning.); Replay identities, restored task outputs and runtime/Flow reconstruction fields: partial — validation [wired], enforcement [wired], routing [wired], knowledge [wired] (Arbitrary application fields and opaque checkpoint payloads lack uniform downstream authority. Schema code itself is not counted as retained memory.); Additional file-input, tool-cache and retrieval-adapter retained parts and their later consumers: uninspected (The accepted account inventories but does not trace these memory-detail alternatives. Their behavioral authority cannot be inferred from the general tool/cache route or from unified-memory witnesses.)	AAS-2026-10-15-crewai-62e53db969a7-01
Decapod	2	Accumulated knowledge and promotion events; federation decisions and access edges; archive originals/replacement summaries; LCM originals/derived summaries; aptitude preferences/prompts/observations; trajectory/Jev governance history; feedback; persisted context capsules/bundles; internalization payload/manifests/leases; authored living specs; retained session/task/workspace custody and acceptance/evaluation state. Included untraced decide, worker lesson, health/reflex, scheduled/evaluator, RPC trace, full spec/governance and remote-driver alternatives remain unresolved. Excludes shipped static constitution/host instructions, ordinary transient controller state, installed executable updates and external host/model/provider internals.	partial coverage (Files, repository versions, database tables and federation graph access have positive witnesses. Untraced alternative stores prevent complete inventory; external host/provider parameters are excluded, not classified as this memory substrate.); Knowledge rows, promotion events, feedback rows and LCM summary table: known — sqlite [wired] (Canonical objects retain the SQLite information supplied by reconciled epistemic declarations; summary text is distinct from original events.); LCM original stream/index, aptitude preferences/prompts/observations, federation node and edge tables: known — rdbms [wired] (Relational tables and transactional streams are recorded; do not infer every backend implementation from the shared database name.); Federation node/source/supersedes access structure: known — graph [wired] (The inspected edge structure is known; complete traversal, search and rebuild remain in the alternative-surface gap.); Archive originals and replacement file; internalization payload/manifest/mount/log; bundle/capsule; live trajectory/Jev sections; sessions and eval JSON: known — files [wired] (These operative files are known even where custom payload semantics are opaque.); Committed governance history, authored living specs and accepted-base executor authority: known — repo [wired] (Live files and immutable repository versions are distinct retained surfaces; transient controller memory is not accumulated memory.); Included decide producer and federation graph maintenance; worker lessons; health/reflex remediation; scheduled maintenance/evaluation; RPC trace consumers; spec/override initialization; full plans/claims/checkpoints and remote driver evidence: uninspected (The accepted inventory names these available but untraced alternatives. Their retained parts, admission, transformations and later consumers lack complete recorded facts; no complete system value inventory follows. References identify the neighboring established mechanisms, not invented records for missing parts.)	partial coverage (Several derived admissions are automatic. No manual value is supported solely by caller identity, human intent or approved_by text; explicit content admission and custom producer control remain unresolved.); LCM derived summary admission after requested summarize: known — automatic [wired] (Starting the operation is not per-content approval of the generated prefix.); Successful handoff knowledge/observation admission: known — automatic [wired] (Caller authors summary but software independently admits the derived content; writes are best effort.); Jev ledger append, governance normalization, session reacquisition, generated readiness healing, bundle export and internalization manifest/noop writes: known — automatic [wired] (Known software-produced structures are separated from caller semantic admissions and opaque custom output.); Eval statistics aggregate admission: known — automatic [wired] (The caller selects settings, not each generated aggregate value.); Explicit knowledge/promotion/decision/preference/prompt/feedback/trajectory content and archive summary/replacement; LCM raw ingestion; external eval verdict and gate requirement: not-determinable (Generic callers, approver strings and workflow starts do not establish an explicit human decision. Software field validation alone does not resolve content control.); Configured internalizer custom payload admission: not-determinable (Executable production is known, but its semantic admission mechanism and possible approval step are opaque.); Included decide producer and federation graph maintenance; worker lessons; health/reflex remediation; scheduled maintenance/evaluation; RPC trace consumers; spec/override initialization; full plans/claims/checkpoints and remote driver evidence: uninspected (The accepted inventory names these available but untraced alternatives. Their retained parts, admission, transformations and later consumers lack complete recorded facts; no complete system value inventory follows. References identify the neighboring established mechanisms, not invented records for missing parts.)	partial coverage (All classified read-back chains are pull. No positive push witness is recorded, but included alternative consumers remain uninspected, preventing a complete pull-only or absent-push assertion.); Requested knowledge/decision search, archive restore, LCM summary/show, aptitude reminder/pending, feedback proposal, bundle import and current spec resolution: known — pull [wired] (Request, internal selection and returned delivery form one pull chain, not an extra push.); Requested assurance/trajectory and structural session, artifact, ledger, evaluation and readiness consumers: known — pull [wired] (Request-owned operations reload retained inputs as part of fulfillment; unsolicited model supply is not established.); Custom inference loader, RPC trace consumers and scheduler/reflex/worker-lesson delivery alternatives: uninspected (Known metadata reload and automatic writes do not decide opaque inference supply or untraced later-consumer direction.); Included decide producer and federation graph maintenance; worker lessons; health/reflex remediation; scheduled maintenance/evaluation; RPC trace consumers; spec/override initialization; full plans/claims/checkpoints and remote driver evidence: uninspected (The accepted inventory names these available but untraced alternatives. Their retained parts, admission, transformations and later consumers lack complete recorded facts; no complete system value inventory follows. References identify the neighboring established mechanisms, not invented records for missing parts.)	uninspected coverage (No controlled push-signal finding is established. Partial direction coverage prevents axis inapplicability; unresolved alternatives remain uninspected even though known pull units are inapplicable.); Fully inspected requested text and structural reload chains: inapplicable (Signal classifies push selection only. Identifiers/lexical matching on requested pull returns are not push signals.); Opaque inference and untraced automatic/scheduled later-consumer alternatives: uninspected (No recorded push selector trigger/input/selected payload is available for these included alternatives; identifiers on manifests or requested files cannot fill the gap.)	partial coverage (Two local automatic trace-fed context routes support yes. This comparison classification neither claims improved capacity nor contradicts unobserved learning/adoption in the reports. Other included routes cannot be assigned a blanket no.); LCM interaction/tool originals to automatic durable prefix to later requested summary consumer: partial — yes [wired] (This qualifies for trace-fed durable context writing despite deterministic formatting and no demonstrated improvement. Artifact/nontrace input alternatives remain unresolved.); Successful task handoff activity to persistent formatted knowledge to later search/pending: known — yes [wired] (Automatic best-effort admission and later consumer route establish wiring, not execution success or automatic recipient injection.); Trace origin and behavioral update in caller trajectory, Jev history, feedback/pattern proposals, eval, custom adapters and worker/scheduled remediation: not-determinable (Curated declarations, raw retention and fixed statistics do not establish a complete qualifying automatic experience-to-behavior update. Custom/worker/remediation writes and future consumers are unresolved; no bounded absence record licenses no.); Included decide producer and federation graph maintenance; worker lessons; health/reflex remediation; scheduled maintenance/evaluation; RPC trace consumers; spec/override initialization; full plans/claims/checkpoints and remote driver evidence: uninspected (The accepted inventory names these available but untraced alternatives. Their retained parts, admission, transformations and later consumers lack complete recorded facts; no complete system value inventory follows. References identify the neighboring established mechanisms, not invented records for missing parts.)	partial coverage (Knowledge, instruction, ranking, routing, validation, enforcement and learning each name retained parts and actual consumers. Excluded static constitution does not establish retained instruction; partial custom/alternative consumers prevent a system-complete authority account.); Knowledge/decision/archive/LCM/feedback and observation content at requesting host; trajectory at assurance; bundle snippets and living specs at host: known — knowledge [wired] (Context/advice delivery establishes the implemented channel, not observed reliance or truth.); Retained aptitude prompt and preference directions at host reminder consumer: known — instruction [wired] (Standing directions are distinct from their selection metadata; host compliance is unobserved.); Prompt priority/usage; knowledge optional recency fields; federation update order: known — ranking [wired], ranking [afforded], ranking [wired] (Each ranking finding names the recorded ordering input; no outcome-driven learned ranker is inferred.); Knowledge status, prompt active/context, session identity/expiry and workspace assignments; same-run governance ledger association: known — routing [wired], routing [wired] (Retained state chooses delivered scope or continuation. Consumers/effects are not generalized to all commands.); Promotion event at procedural add; archive index/budget inputs; internalization manifest/lease at attach and validation; bundle bindings at import; eval requirement/aggregate at publisher; accepted executor and retained proof at readiness: known — validation [wired], enforcement [wired] (These retained parts supply actual fixed criteria and executable vetoes. A hash read alone does not make adapter semantic content a criterion.); LCM recorded interaction/tool originals at automatic summary update; handoff activity at durable knowledge derivation: known — learning [wired] (Learning authority is local to experience-fed durable artifact production, independent of improvement or host adoption.); Custom adapter at hypothetical inference consumer; full policy/spec authority and untraced coordination consumers: uninspected (Integrity/mount consumers are known above; parameter loading, full policy provenance and other effects remain uninspected.); Included decide producer and federation graph maintenance; worker lessons; health/reflex remediation; scheduled maintenance/evaluation; RPC trace consumers; spec/override initialization; full plans/claims/checkpoints and remote driver evidence: uninspected (The accepted inventory names these available but untraced alternatives. Their retained parts, admission, transformations and later consumers lack complete recorded facts; no complete system value inventory follows. References identify the neighboring established mechanisms, not invented records for missing parts.)	AAS-2026-10-14-decapod-c44705119841-01
deja-vu	2	Included: explicit/promoted notes and lifecycle rows; normalized historical text and access metadata; mined fix, token, command/failure and session-outcome sidecars; dense retrieval outputs and unresolved associated provider retention; cached/served digests and delivery/failure/usage state; continuation content and identity; adopted standing-rule source/copies; changed activation/suppression controls; import and recovery writes; unresolved specialized queue/checkpoint, parser, maintenance and consumer alternatives. Excluded: shipped static instructions, unchanged binary/generated wiring, external host inference/tool scheduling and original transcript production, and temporary offline-evaluation state without a later production consumer. Provider weights are not assumed to change through use.	partial coverage (Local filesystem witnesses and lifecycle memoization are established. Opaque provider retention and specialized state prevent a complete substrate inventory; no SQLite, graph or vector-database substrate is inferred from serialization or dense encoding.); Plain/promoted notes and local lifecycle rows: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Normalized historical text records: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Session and imported lifecycle metadata: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Pre-judgment fix evidence: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Judged fixes and held candidates: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Token neighbor map: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Recurring command table: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Failure warning table and its update evidence: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Session commands/outcomes: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Continuation content: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Continuation identity/freshness metadata: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Standing rules source and installed copies: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Dense numerical sidecar: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Rendered digest caches and served-text snapshots: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Persistent delivery/failure/usage metadata: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Changed activation policy and suppression state: known — files [wired] (The named local retained part is stored on the filesystem; JSONL, gob and binary are encodings, not database substrates.); Process memoization of local lifecycle lookup: known — in-memory [wired] (The keyed lifecycle lookup also has a process-resident memoized representation.); Opaque embedding-provider state associated with retained numerical outputs: uninspected (Endpoint substrate and any provider-retained input/output state are uninspected; the local sidecar does not resolve them.); Included specialized stop/deferred/checkpoint hooks, other host integrations and parser/recovery variants: uninspected (Specialized retained queue/checkpoint payloads and usage rotation/layout are not classified by the common filesystem witnesses.)	partial coverage (Automatic admissions are wired; manual adoption is prescribed for standing rules and remains claimed. CLI caller identity, direct initialization/edit paths and specialized writes prevent complete agency coverage.); MCP remember admission: known — automatic [wired] (The live software API admits content without per-content human approval.); CLI remember/promote caller-supplied content and lifecycle choices: not-determinable (A generic CLI caller can be a human or software; human content approval is not established for every invocation.); Default promoted-note distillation after invocation: known — automatic [wired] (Starting promotion is not approval of each extracted excerpt.); Index, sidecar, embedding, digest-cache and continuation admission: known — automatic [wired] (Software/model derivation and structural filtering own these admissions.); Rules source adoption in the prescribed candidate-review workflow: known — manual [claimed] (The skill requires selection of exact lines before append; the API does not enforce this procedure.); Pre-existing rules source and direct source edits outside the prescribed workflow: not-determinable (Who supplied/approved each source line is not established by file possession or sync invocation.); Rules copy distribution: known — automatic [wired] (The source-adoption decision and subsequent file-copy admission are separate.); Imported-batch local admission: known — automatic [wired] (Choosing a peer or workflow is not approval of each admitted row.); Recovery write-back and span destination writes: known — automatic [wired] (Invocation/destination control does not require human per-content approval of reconstructed text.); Privacy/lifecycle and activation-policy source decisions: not-determinable (Generic caller identity leaves whether a human explicitly approved retained suppression/configuration content unresolved.); Included specialized stop/deferred/checkpoint hooks, other host integrations and parser/recovery variants: uninspected (Specialized queued/checkpoint writes, host parser alternatives and usage-history admission are not fully inspected.)	partial coverage (Pull and push both exist. Request fulfillment, automatic event delivery, standing-loader affordance and internal update reads remain separate; specialized alternatives prevent complete route coverage.); Explicit CLI/MCP recall, context and resource requests: known — pull [wired] (The caller requests history; ranking/delivery fulfilling it are not extra push operations.); Explicit handoff, native resume and span recovery: known — pull [wired] (Explicit/newest-project choice still fulfills a recovery request.); Independent startup/prompt/tool/error/plan and spawned-input supplies: known — push [wired] (Host events/present task signals are triggers, not requests for retained memory.); Session/workspace continuation recovery on eligible hooks: known — push [wired] (Automatic hook selection supplies the matching retained capture.); Rule candidate extraction request: known — pull [wired] (Candidate output is requested; it does not itself update the rules file.); Standing rule copies at later host context assembly: known — push [afforded] (The shipped files supply an intended automatic instruction-loader path; actual loading is not observed.); Retained raw fix sightings and previous carry requested by automatic update code: known — pull [wired] (This is update-only read-back; ensuing model delivery has its own direction.); Included specialized stop/deferred/checkpoint hooks, other host integrations and parser/recovery variants: uninspected (Deferred/specialized hook chains and opaque provider retention lack fully assessed request/selection/delivery operations.)	partial coverage (Coarse, identifier and inferred-lexical push signals are established at distinct selectors. Dense rerank is recorded on requested recall, not sufficient evidence for inferred-embedding push. No model-judgment push selector is established; this is not bounded absence.); Startup digest coarse recency/availability and budget filtering: known — coarse [wired] (Trigger: startup event. Inputs: recent project history, policy and budget. Selected part: bounded cached/new digest.); Startup project convention identity selection: known — identifier [wired] (Trigger: startup. Input: current allowed project names and latest accepted note state. Selected part: matching local/imported decision lines.); Startup task-file, prompt, file/error and spawn prompt relevance: known — inferred-lexical [wired] (Triggers: startup/prompt/tool/failure/spawn events. Inputs: changed task files, current terms/path/error or child prompt. Selected part: corresponding prior history.); Automatic plan warning and remedy selection: known — inferred-lexical [wired] (Trigger: plan submission. Inputs: extracted plan-step terms/error walls and retained fix flags. Selected part: recurring-wall receipts and eligible remedies.); Continuation native session/workspace match: known — identifier [wired] (Trigger: eligible context/prompt/tool hook. Inputs: native session/workspace and stored capture metadata. Selected part: the matched continuation packet.); Standing whole-rule block in intended host loading: known — coarse [afforded] (Trigger: later host instruction assembly. Input: available marked global rules block. Selected part: that standing block, not topical history.); Explicit recall and recovery request fulfillment: inapplicable (Read-back signal classifies push; query/prefix/URI identifiers on these pull operations do not establish identifier push.); Included specialized stop/deferred/checkpoint hooks, other host integrations and parser/recovery variants: uninspected (Deferred/checkpoint selectors and other hook implementations are uninspected; semantic pull witnesses cannot classify them.)	partial coverage (Wired yes findings describe automatic trace-fed durable behavior-shaping artifacts, not demonstrated improved capacity. No bounded absence record licenses a local or system-wide no; unresolved routes are retained.); Automatic error/remedy extraction, judging and later consumers: known — yes [wired] (Recorded command/error/edit experience feeds durable selector/advisory state used by later requests/events.); Automatic continuation extraction/recapture and later host delivery: known — yes [wired] (Within-session later invocation is sufficient; faithful semantics, new knowledge and demonstrated benefit are not required.); Recurring command/failure derivative routes with stated later consumers: known — yes [wired] (Consumers include fixed lookup and per-action warning; exact all-caller behavior remains limited.); Token-neighbor rescue and session-fact derivative routes: uninspected (Automatic trace derivation is recorded, but full later selector/rendering paths remain uninspected; this prevents a complete qualifying-route assessment.); Rules source formation and out-of-workflow updates: not-determinable (The prescribed human approval route does not establish qualifying automatic source formation; direct source provenance/control is unresolved. Copying approved text does not by itself close an automatic trace-fed write chain.); Opaque embedding-provider learning/retention: uninspected (The client requests embeddings, not training; provider parameters and retained experience are uninspected. No local no is licensed for provider internals.); Included specialized stop/deferred/checkpoint hooks, other host integrations and parser/recovery variants: uninspected (Specialized hook/checkpoint writes, variant host summaries and adaptive usage/evaluator consumers prevent a bounded complete negative or full qualifying-route enumeration.)	partial coverage (Knowledge, instruction, learning, ranking, routing, validation and enforcement have distinct retained-part/consumer warrants. Model obedience, improved capacity and all alternative consumer paths are not established.); Historical text, receipts, fixes, commands and outcomes at recall/model readers: known — knowledge [wired] (The force is contextual evidence or advice, not executed actions or warranted truth.); Continuation objectives, conclusions and recorded outcomes at recovery reader: known — knowledge [wired] (Framed continuation is advisory; successful host receipt/use is unobserved.); Accepted convention prose at the requested standing-decision consumer: known — instruction [wired] (The imperative request coexists with untrusted-history framing; compliance and effective priority remain unresolved.); Continuation standing constraints and carry directions at later model reader: known — instruction [wired] (Direction-bearing content is separated from factual packet evidence.); Rules prose at the intended later global host instruction loader/model: known — instruction [afforded] (Wired file distribution establishes the intended channel, not actual host loading or obedience.); Retained corpus and raw sightings at automatic fix/sidecar update consumer: known — learning [wired] (Update force and downstream model evidence are distinct.); Previous continuation content at subsequent capture: known — learning [wired] (Recapture reads surviving carry/rules/goal before producing the next durable packet.); Historical text/features and dense vectors at retrieval/rerank consumer: known — ranking [wired] (Fixed scoring can give retained content ranking authority without a learned scoring algorithm.); Judged fixes and lifecycle state at priority/selection consumers: known — ranking [wired] (Confirmed-first fallback is consumer-specific; rejected history may be demoted rather than erased.); Session/project IDs, capture fingerprints and delivery ledgers at selectors: known — routing [wired] (These retained parts choose reader scope, packet revision or repeated-delivery eligibility.); Activation grants, tombstones, failure/delivery tokens at withholding consumers: known — enforcement [wired] (The enforced boundary is local admission/output eligibility, not host action permission or universal erasure.); Suppression and capture metadata at write/admission checks: known — validation [wired], enforcement [wired] (The checks concern admissibility and identity/freshness, not proposition truth.); Stored judged candidate flags at plan-remedy admission: known — validation [wired], enforcement [wired] (Requested lookup can expose candidates while plan remedy refuses them.); Token-neighbor rescue and incompletely traced command/session-fact/update/usage consumers: uninspected (Objects and some consumer roles are known, but all consequential force paths are not traced; existence of a table does not settle every consumer.); Included specialized stop/deferred/checkpoint hooks, other host integrations and parser/recovery variants: uninspected (Specialized hooks, other host loaders and opaque provider state leave further authority paths uninspected.)	AAS-2026-10-14-deja-vu-22137d4d39a5-01
dense-mem	2	Included: accumulated PostgreSQL evidence/relationship/support/lifecycle state; search documents/vectors; raw session events and extraction/linking checkpoints; ontology records/receipts, community projections and conflict audit/recovery; terminal/recovery outcomes, optional feedback/diagnostics, historical Dream payload/display and evaluation mapping/checkpoint/trace artifacts. Included uncertainty: initial vocabulary and historical provenance, opaque payloads, detailed admission/selection and downstream consumer alternatives. Excluded: external model weights/training, static shipped prompts/configuration, transient provider working history, external host prompt assembly and unestablished persistence/installation of exported packs or drafted skills.	partial coverage (The inventory includes primary memory, access structures, recovery/provenance, optional feedback, evaluation and historical state. Named storage gaps prevent a complete substrate union; external model parameters and ordinary process-local working conversations are excluded.); Persistent evidence, semantic/support/lifecycle state, search text, raw sessions, checkpoints, ontology/receipts and conflict audit parts: known — rdbms [wired] (The records identify PostgreSQL for these operative stores.); Retained dense search vectors and their contract metadata: known — rdbms [wired], vector [wired] (Vector access and relational persistence coexist; this does not classify external provider weights as use-accumulated memory.); Evaluation mappings, checkpoints and saved traces/scores in run artifacts: known — files [wired] (The shipped runner establishes file-artifact retention/reload, not observed artifact creation.); Optional feedback, assessment diagnostics, other recovery ledgers, community state and historical Dream payloads: partial — rdbms [wired] (Feedback persistence is identified, but the accepted records do not locate every included historical, diagnostic and recovery part precisely enough for complete substrate enumeration. A service/repository interface alone does not locate storage.)	partial coverage (Automatic admission has several inspected witnesses. Manual requires explicit human content control, absent from the available positive evidence; that is not a bounded claim that manual control never occurs.); Remember semantic admission, session staging/extraction/checkpoint/semantic commits, automatic ontology and conflict review: known — automatic [wired] (This classifies inspected software admission, not humanity of whoever supplied initial content.); Search/community projection maintenance, optional snapshot recording and policy-triggered retention/erasure: known — automatic [wired] (Caller-triggered erasure and volunteered feedback have distinct unresolved operator-control paths below.); Explicit initial evidence supply, lifecycle patches/confirmations, manager edits/publication, voluntary feedback and caller erasure decisions: not-determinable (Generic caller/owner/manager identity does not establish a human decision. Per-content control at these caller decisions cannot be classified manual or automatic merely from role names.); Historical Dream initialization, initial ontology vocabulary, full evaluation import/admission and diagnostic persistence controls: uninspected (The reports leave these admission mechanisms or initial control uninspected. Known later transformations do not resolve them.)	partial coverage (Pull requests and automatic push-to-internal-consumer operations coexist. A request for recall is not counted as push merely because the service selects results; upstream intake initiation does not erase independent prior-context selection.); Caller-requested recall, handle trace/export, administrative/feedback readers and explicit replay/scoring loads: known — pull [wired] (Concrete request/return and later-consumer routes are recorded; external host answer use is not required for this route classification.); Automatic prior-session selection into a new extraction frame: known — push [wired] (The caller initiates intake, but software independently supplies prior context to the model under count/token bounds.); Automatic organization source/vocabulary context and projection/conflict worker inputs: known — push [wired] (Automatic worker/model supply is within the memory boundary even though no unsolicited host conversation injection is established.); Opaque historical payload reuse, full diagnostic readers and all alternative recovery consumers: not-determinable (Known displays/replay do not establish every later opaque payload or diagnostic consumer; complete alternative-route coverage is unavailable.)	partial coverage (Identifier, coarse and lexical signals attach to named internal push selections. Complete pull-only scope is not established; signal is therefore not globally inapplicable.); Prior-session events automatically supplied at new extraction: known — identifier [wired], coarse [wired] (Identity targeting and coarse recency/budget selection are separate operations in one push chain.); Ontology context supplied to later organization: known — identifier [wired], inferred-lexical [wired], coarse [wired] (These inputs select delivered retained parts, not merely catalog IDs. No inferred-embedding or inferred-judgment push value is inferred from requested recall ranking.); Search/community maintenance and conflict-review automatic inputs: partial — coarse [wired] (The recorded selectors establish coarse policy/work availability. Detailed repository selection and all dossier composition predicates remain uninspected.); Requested recall, trace/export, administrative readers and explicit replay/scoring loads: inapplicable (These are pull operations. Query vectors or requested relationship/file IDs do not establish a push signal.); Other historical/diagnostic/recovery alternatives: not-determinable (Insufficient consumer/selection coverage to decide whether additional push selectors operate. Known pull routes cannot make this unresolved branch inapplicable.)	partial coverage (A conditional positive survives incomplete system coverage. No local or system-wide no is asserted from missing run evidence, absent parameter training or ordinary event retention alone.); Automatic session extraction/checkpoint-to-semantic-memory chain when event content records interaction experience: partial — yes [wired] (The conditional route is implemented; opaque/mixed caller origin prevents treating every session submission as a qualifying trace.); Feedback/evaluation records, conflict/organization inputs, historical Dream and other opaque write provenance: not-determinable (Feedback and evaluation readers do not establish an automatic behavior-shaping update; their broader consumers and other historical/input provenance remain uninspected. Without bounded complete negatives, these branches receive neither no nor inapplicable.)	partial coverage (Knowledge, ranking, enforcement, validation, routing and conditional learning have local consumer warrants. No instruction value is inferred merely from preferences in recalled text or the static export prompt; unresolved consumers prevent a complete authority classification.); Eligible evidence, relationship text and historical/feedback/trace displays at interpreting readers: known — knowledge [wired] (This classifies the delivered reference parts, not uninspected opaque payload effects or full fragment delivery.); Search text, vectors and ontology name/lexical inputs at retrieval: known — ranking [wired] (Ranking authority needs no retained scoring policy and does not imply learned retrieval parameters.); Retained IDs, scope/generation, lifecycle/support state and current ontology controls at service admission/selection: known — enforcement [wired], validation [wired] (Only the recorded criteria actually consumed by software are classified. Ordinary evidence need not have a qualifying relationship; no universal final support gate is asserted.); Session events/checkpoints, durable Remember outcomes, ontology dependencies and evaluation mappings at later progression: known — routing [wired] (Fixed control effects are distinct from semantic interpretation of the same retained objects.); Trace-bearing raw session content at automatic extraction and linking: partial — learning [wired] (This conditional update-consumer force is distinct from downstream knowledge authority and does not establish improved capacity. Actual input origin is unresolved.); Ontology definition prose at the later organizer: known — knowledge [wired] (Manager overrides have separate fixed control force; definition prose alone is not instruction authority.); Opaque historical payload consumers, full diagnostic/review reasons and possible later exported-skill adoption: not-determinable (Display/reference and receipt-reuse paths do not establish all payload or rationale consumers. Skill drafting has no established retained installation. Shipped prompts are excluded from accumulated-memory authority.)	AAS-2026-10-15-dense-mem-1a314bed9a5c-01
DocMason	2	Local retained source evidence and derived semantics/overlays/indexes; captured and pending/promoted interaction content/policy; linked later-turn briefs/control and hook reminders; retained answers/traces; runtime review/candidates and subsequent evaluator cases/baselines/results. Includes unresolved parser/attachment, initial-content control and answer-history consumer alternatives. Excludes static shipped instructions, adapter/core replacement, external host model weights/provider caches and native host context compaction.	known coverage (The accepted inventory identifies local file storage across the included memory surfaces. External host/provider state, static instruction projection and core replacement are excluded, not silently classified as file memory.); Document evidence, semantics, overlays and reader/provenance indexes: known — files [wired] (Staged/current directories and target-local JSON/text/assets are the retained substrate. Graph edges are file contents, not a separate graph database.); Native ledgers/cursors, interaction entries and pending/promoted artifacts: known — files [wired] (Runtime JSON and per-memory staging/current directories persist these parts; provider parser variants do not establish a different retained substrate.); Linked task control, publication selection and canonical answer files: known — files [wired] (Filesystem control and separate answer Markdown remain available to linked later invocations.); Review projections, candidates, regression drafts/cases, baselines and evaluation results: known — files [wired] (Runtime review/evaluation JSON and scorecard files feed subsequent review and replay.)	partial coverage (Automatic transformations and explicit human decisions coexist with unresolved initialization and request-control mechanisms.); Initial live corpus supply: partial — manual [afforded] (Operator placement is afforded, but generic caller identity and unknown private installation control do not resolve every initial content admission.); Build, conservative semantics, indexes and publication: known — automatic [wired] (Software proposes/transforms/adopts after checks; initiating sync is not per-assertion human approval.); Host semantic text/overlay proposal: known — automatic [claimed] (Keep requested model admission separate from the wired software publication branch.); Native reconciliation, derived pending/promoted content and maintenance: known — automatic [wired] (Python admits entries, grouping and metadata; user messages provide content but do not approve each derived memory.); Linked control normalization and hook continuation claim: known — automatic [wired] (Python normalizes/persists bindings and one-shot claims; user gate decisions are a separate route.); Explicit user gate/scope decisions: known — manual [wired] (The recorded decision/confirmation route requires a user decision for accepted task scope, not for every generated answer claim.); Review projection and candidate extraction: known — automatic [wired] (Software filters/groups traces and writes candidates independently of later promotion selection.); Candidate promotion and baseline replacement control: not-determinable (An explicit request selects cases/run, but generic request-file identity does not establish human per-content approval versus automated admission control.); Initial suites/references and host answer/work-brief admission: not-determinable (Supplied evaluator inputs and host-authored content are recorded, but all initial/control decisions are not specified; authorship alone cannot settle agency.)	partial coverage (Both requested retrieval and automatic advisory supply are wired; incomplete alternative consumer coverage remains local.); Requested published/pending retrieve and provenance trace: known — pull [wired] (The consumer requests evidence or a specified trace; delivery fulfilling that request is not an additional push.); Explicit linked ask continuation and gate resolution: known — pull [wired] (Explicit revision/gate anchors request continued task state, distinct from optional automatic prompt reminders.); Optional prompt-hook advisory continuation supply: known — push [wired] (Prompt event and bound state automatically select linked task/gate/scope reminders without a separate memory request.); Requested runtime review and regression replay/promotion: known — pull [wired] (Explicit review/action requests load retained summaries, candidates, cases and baselines; background projection writes are not themselves host delivery.); Untraced answer-history and provider/attachment consumer alternatives: uninspected (Records do not establish every later consumer route for these included retained parts.)	partial coverage (A positive targeted hook route cannot make untraced alternatives inapplicable; pull lexical selection is not inferred-lexical push.); Prompt-hook linked task targeting: known — identifier [wired], coarse [wired] (On prompt submission the bound session/task linkage selects matching canonical state and scope/gate IDs; latest-linked-turn selection is coarse recency within that identity-bound set. No token budget or embedding selector is recorded.); Requested evidence, continuation, review and evaluation routes: inapplicable (These units fulfill explicit requests; their lexical ranking and requested IDs are not push signals.); Untraced later-consumer alternatives: uninspected (Whether additional push operations and selectors exist in these alternatives is not established by accepted records.)	partial coverage (Positive automatic trace-fed context/candidate paths coexist with unresolved update/input routes. No bounded absence record establishes a local or system-wide no.); Automatic pending interaction projection and published semantics: partial — yes [wired] (Recorded conversational turns feed automatic retained context/semantic writes with later retrieve/host consumers. Individual provider/attachment inputs are not fully inspected; this is wired trace learning, not demonstrated improved capacity.); Automatic runtime review/candidate production: known — yes [wired] (Captured session/tool state automatically produces retained review/candidate artifacts read by later review and case-generation consumers. Case promotion control remains separately unresolved.); Linked work-brief revision and native capture recovery alternatives: not-determinable (Explicit host-authored brief deltas and parser tail checkpointing do not themselves establish an automatic trace-fed behavior-shaping summary; complete qualifying input/update paths are not recorded.); Source-document semantic compilation: not-determinable (Imported document content is not necessarily activity trace content. Private payload origins are unknown, so neither a positive nor bounded no is supported across this unit.); Evaluator case/baseline adoption: not-determinable (Trace-derived cases have a later consumer, but explicit request control is unresolved; automatic candidate production does not settle downstream adoption agency.)	partial coverage (Each force is attached to its actual consumer. Untraced history/provider/asset alternatives prevent complete authority enumeration.); Evidence and semantics delivered to answering/authoring hosts: known — knowledge [wired] (Retained evidence and semantic text supply context/reference; delivery does not prove compliance or factual truth.); Reader search content and interaction policy at retrieval: known — ranking [wired], routing [wired], enforcement [wired] (These are distinct fixed-reader effects; contextual-only is not a universal support veto.); Linked gate/scope/closure fields and publication pointer: known — routing [wired], validation [wired], enforcement [wired] (Control force is bounded to cooperating invoked paths, not arbitrary host filesystem actions.); Work briefs and advisory continuation text at later host invocation: known — knowledge [wired], instruction [wired] (Brief content supplies task context; selected continuation guidance directs exact revision/gate use. Optional delivery is wired, host adherence unobserved.); Overlay fingerprints and slots at coverage evaluation: known — validation [wired], routing [wired] (Persisted fingerprints/slot claims determine coverage credit and remaining follow-up work, not universal stale-text quarantine.); Native cursor and entry bookkeeping at reconciliation/promotion: known — routing [wired] (Retained cursor/entry status selects recovery, capture and eligible promotion inputs.); Retained sessions/entries at automatic semantic and projection updates: known — learning [wired] (Retained trace content feeds durable context/semantic/candidate artifacts with later consumers; this update authority does not establish improved capacity.); Review/candidate content at operator and case-generation consumers: known — knowledge [wired], ranking [wired], learning [wired] (Questions/reasons advise review, retained priorities order candidates, and candidates feed durable regression cases consumed later.); Regression expectations, reference facts and frozen baseline at later replay: known — knowledge [wired], validation [wired] (References supply judging context; expected-state/baseline fields determine bounded replay conformity and comparisons, not corpus truth.); Persisted answer and trace outcomes at later tracing/closure: known — knowledge [wired], validation [wired] (Answer text is the support-check target; retained trace/support results license or refuse a declared closure label.); Untraced answer-history and alternate attachment/provider consumers: uninspected (Records retain incomplete consumer/capture coverage; no complete authority union for these alternatives can be asserted.)	AAS-2026-10-13-docmason-deba9ae199d5-01
Dynamic Cheatsheet	2	Use-accumulated final reference, full responses, within-call retained answers, checkpoint envelopes and diagnostic/selection fields, retained evaluation targets, exported client conversation history, local native dispatch handle and associated opaque provider execution payload; includes initialization/replacement, maintenance, continuation, evaluation and later model consumers. Excludes shipped static prompt templates, precomputed CSV embeddings and model weights as accumulated memory, initial dataset storage, transient unretained search snippets and deleted temporary scripts. Search-conditioned retained response contributions and native payload alternatives remain included with uncertainty.	partial coverage (Local checkpoint and client storage are covered; provider payload storage remains unresolved. Static prompt files and CSV vectors are excluded.); Benchmark reference, response corpus, diagnostic/selection fields, answer/target and checkpoint envelope: known — in-memory [wired], files [wired] (The recorded operative checkpoint parts have local strings/containers and JSONL persistence.); Exported client history and native dispatch handle: known — in-memory [wired] (Both local state surfaces reside in client/wrapper instances.); Associated native provider payload: not-determinable (Provider-managed execution state is opaque; a service handle does not establish the storage substrate of accumulated payload.)	partial coverage (Automatic downstream admission is known despite unresolved caller initialization/replacement control and native writes.); Curator extraction/install, driver corpus/snapshot admission and client chat append: known — automatic [wired] (Software/model admission has no per-content human approval.); Seed-file, checkpoint replacement and set/reset-history content control: not-determinable (Caller APIs afford replacement, but caller identity and explicit human content decision are unresolved; manual is not established.); Native payload writes/admission: not-determinable (Remote writers and admission controls are opaque.)	partial coverage (Both explicit requested reads and unsolicited model-context supply occur in local chains; native payload direction remains unresolved.); Automatic question/round-conditioned reference and example supply: known — push [wired] (Incoming task triggers context selection without a generator request for retained examples.); Automatic exported-client prior-turn supply: known — push [wired] (Subsequent generate/chat includes the retained list, not a requested retrieval return.); Explicit checkpoint reload and retained-output evaluation reads: known — pull [wired] (Caller-selected resume and notebook reads fulfill requests for saved material; downstream context selection remains separately push.); Native payload reuse/read-back: not-determinable (Repeated handle delivery does not establish a payload read request, selector or later consumer.)	partial coverage (Known push selectors are embedding similarity and coarse inclusion; explicit file requests add no identifier-based push. Model semantic content choices during curation are not independently evidenced selectors of preserved delivered items.); Cosine-selected prior examples on incoming question: known — inferred-embedding [wired] (Current question vector ranks prior question vectors; matching indices select delivered prior outputs.); Whole-reference, full-history and client-history context supply: known — coarse [wired] (These push operations include the available whole reference/list rather than targeted item matches.); Native payload selection: not-determinable (Handle is dispatch state, not evidence of identifier matching to delivered memory parts.)	partial coverage (Two wired automatic trace-fed reference routes qualify. Raw logging does not establish another route; unresolved supplied/native writes prevent complete system coverage or a global negative.); Cumulative/hybrid automatic successor reference: known — yes [wired] (Current generated response, optional execution feedback and old reference produce retained text consumed by next round/question or resumed run.); Retrieval-synthesis automatic successor reference: known — yes [wired] (Prior response traces and derived reference feed query-conditioned replacement, carried forward by driver.); Raw corpus/history append, diagnostic persistence and scoring without derived update: inapplicable (Raw logging/replay and evaluator counts alone do not constitute a qualifying transformed behavior-shaping artifact write.); Caller replacement and native payload update routes: not-determinable (Opaque origins/control and remote writes prevent complete classification of possible qualifying updates.)	partial coverage (Recorded advisory, update, restart and scoring effects are included. Static instructions and embedding ranking inputs are excluded; no privileged instruction or correctness-enforcement force is inferred for reference text.); Live reference and raw history/examples at solving consumers: known — knowledge [wired] (Reference and prior conversational material supply advisory context; delivery is not demonstrated reliance.); Prior reference and response traces at curator update consumer: known — learning [wired] (Update-consumer force is separate from successor reference use.); Checkpoint envelope at continuation controller: known — routing [wired] (Saved row count selects restart position, and last reference/all outputs select reconstructed model inputs.); Retained targets at offline task evaluator: known — validation [wired] (Targets are measurement criteria, not curator admission feedback.); Native payload at later provider tool consumers: not-determinable (Actual payload, consumer and force are not established by handle reuse.)	AAS-2026-10-12-dynamic-cheatsheet-ca0a0b29d273-01
eal-bench	2	Retained authorization profiles and frozen payloads; audit/access/checkpoint metadata used by later consumers; event logs and reduced state; saved replay, annotation assessment artifacts, controlled variants and closed-loop write-back. Includes unresolved pressure/challenge, hybrid, domain maintenance and analysis alternatives. Excludes shipped static policy/corpus as memory by itself, ordinary current-invocation messages, and external hosted weights/provider internals not established as retained state changed through this workflow.	partial coverage (File-backed and process-resident witnesses do not enumerate every included alternative. Hosted model internals are not asserted to be memory stores accumulated by this workflow.); Inspected authorization payloads, event log/current state, attempts, executor checkpoint and saved annotation child runs: known — files [wired], in-memory [wired] (These are object storage locations, not JSON encoding or the repository containing program source.); Generic public annotation persistence and evaluation exports: not-determinable (The shipped annotation child run is file-backed; persistence selected by arbitrary public callers and export retention is unresolved.); Included pressure/challenge reuse, hybrid formation, other domain repair/maintenance and diagnostic-analysis routes: uninspected (Accepted reports retain these included alternatives without complete selector, producer or consumer inspection. Their classification cannot be inferred from the ordinary writer or Procurement annotation path.)	partial coverage (Inspected mechanisms establish automatic admission; manual admission cannot be inferred for unresolved public or maintenance paths.); LangMem content admission including repair/fallback and seeded feedback: known — automatic [wired] (Workflow annotations retain empty fallback on terminal failure without prior state.); Event delta admission, gate transformation and whole-candidate selection: known — automatic [wired] (Separate mechanisms share this agency value, not a universal validation policy.); Procurement annotation extraction/admission and controlled widened-memory production: known — automatic [wired] (Physical JSONL I/O is not the reason for this classification.); Generic public annotations and caller-provided initial content: not-determinable (Caller identity alone does not establish human versus computational content admission.); Included pressure/challenge reuse, hybrid formation, other domain repair/maintenance and diagnostic-analysis routes: uninspected (Accepted reports retain these included alternatives without complete selector, producer or consumer inspection. Their classification cannot be inferred from the ordinary writer or Procurement annotation path.)	partial coverage (Both requested reuse and automatic model-context supply exist. Their coexistence does not resolve other shipped study selectors.); Requested saved replay, shared checkpoint continuation and event restoration: known — pull [wired] (Delivery fulfilling these requests is not separately counted as unsolicited push.); Host-built writer/executor/reviewer context and source-authority paired jobs: known — push [wired] (An upstream operator study choice can feed selection; current raw request/history alone is not retained memory.); Accepted annotation loaded for later requested assessment: known — pull [wired] (This classifies the established assessment chain, not arbitrary external storage behavior.); Included pressure/challenge reuse, hybrid formation, other domain repair/maintenance and diagnostic-analysis routes: uninspected (Accepted reports retain these included alternatives without complete selector, producer or consumer inspection. Their classification cannot be inferred from the ordinary writer or Procurement annotation path.)	partial coverage (Identifier and judgment witnesses cover actual automatic selection, not every operation in their surrounding chains. Requested file loads do not independently establish identifier push.); Current-chain profile and case/job evidence pushes: known — identifier [wired] (Trigger is the scheduled update, executor job or review; selection inputs are chain/case/evidence or pool identities, selected products are current payload/state or the matching candidate pool. IDs here select delivery, not merely label files.); Best-of-k choice feeding selected payload to executor: known — inferred-judgment [wired] (Trigger is review-mode preparation; the selected part is one retained candidate. Oracle-only comparison is a distinct symbolic selector, not this finding.); Oracle-only fidelity selection and all alternate context targeting: not-determinable (Gold field-error comparison is inspected but no more specific controlled signal follows without conflating it with lexical or model judgment. Event/replay scheduling outside inspected selectors and pressure/challenge alternatives remain unresolved.); Included pressure/challenge reuse, hybrid formation, other domain repair/maintenance and diagnostic-analysis routes: uninspected (Accepted reports retain these included alternatives without complete selector, producer or consumer inspection. Their classification cannot be inferred from the ordinary writer or Procurement annotation path.)	partial coverage (One qualifying feedback route supports yes; other included writes remain unresolved. No system-wide no is asserted and raw log retention alone is not counted.); Closed-loop action/outcome write-back to subsequent request memory: known — yes [wired] (The route meets the trace-fed artifact definition without requiring new knowledge, improved capacity or weight updates. Neutral and rejected-update branches do not erase this positive route.); Baseline history compression, event extraction and annotation derivation: not-determinable (Durable products and later consumers are established, but records do not fully classify original histories as recorded activity rather than static benchmark content. Annotation input is derived prose, not itself established as an original trace.); Included pressure/challenge reuse, hybrid formation, other domain repair/maintenance and diagnostic-analysis routes: uninspected (Accepted reports retain these included alternatives without complete selector, producer or consumer inspection. Their classification cannot be inferred from the ordinary writer or Procurement annotation path.)	partial coverage (Authority is consumer-specific. Static policy/corpus criteria are outside the accumulated-memory inventory, so their enforcing use is not automatically attributed to remembered grants.); Memory payload at executor and prior profile at writer: known — knowledge [wired] (This is not an enforced organizational permission.); Recorded executor request/action/outcome at closed-loop update consumer: known — learning [wired] (Learning authority here does not assert improved capacity or parameter training.); Retained candidates at reviewer and oracle-fidelity selector: known — ranking [wired] (A model reviewer and symbolic gold comparison are distinct selectors; neither proves zero error.); Saved execution plan and current chain/state identities at continuation consumers: known — routing [wired], validation [wired], enforcement [wired] (Routing and checkpoint veto concern execution/recovery, not authorization truth.); Persisted annotation status, hash and extracted-state agreement at scoring resolver: known — validation [wired], enforcement [wired] (The typed extraction supplies assessment content; eligibility is not semantic endorsement.); Annotation and typed authorization fields at preservation/formation assessment: known — knowledge [wired] (This consumer computes diagnostic results, not a live execution grant.); Retained event log at reducer recovery: known — validation [wired] (Log events are symbolic replay inputs; they are not pushed as standing model instructions.); Included pressure/challenge reuse, hybrid formation, other domain repair/maintenance and diagnostic-analysis routes: uninspected (Accepted reports retain these included alternatives without complete selector, producer or consumer inspection. Their classification cannot be inferred from the ordinary writer or Procurement annotation path.)	AAS-2026-10-15-eal-bench-f9a76e2fa858-01
Ecdysis	2	Shipped use-derived memory: extracted trajectories and failure display; accumulated review transcript, normalized specification and checkpoint content/control/accounting; score-retained harness reference and round audit; skill content/artifact envelope and explicit reload/merge; inference outcomes and explicit persistence. Opaque generic harness, task/tool/output payloads, metadata and caller admission alternatives remain included where they limit classification. Excludes static role prompts/source code, provider weights and provider-internal state, external task/editor/scorer implementations and external harness-factory stores; their responsibilities are not attributed to Ecdysis.	known coverage (The recorded shipped inventory is covered. Provider parameters, external adapter stores and static source files are not use-derived stores of this memory boundary; files and in-memory do not imply automatic harness persistence.); Extracted trajectories and formatted failure display: known — in-memory [wired] (Input files are acquisition sources; the extracted objects are held in process memory.); Review transcript, normalized specification and checkpoint parts: known — in-memory [wired], files [wired] (Content and control parts share these substrates, not necessarily the same consumers.); Retained harness reference and round audit tuple: known — in-memory [wired] (The opaque harness is included as the live retained value. No external factory store is attributed to the shipped plane.); Skills and artifact envelope, including caller-defined metadata: known — in-memory [wired], files [wired] (Opaque metadata encoding does not make the known file/dataclass substrate unknown.); Inference outcome records and checkpoint usage accounting: known — in-memory [wired], files [wired] (Output payloads can be opaque, but the shipped enclosing stores are inspected.)	partial coverage (Automatic subsequent production/admission is supported. Initial caller content, real callback control and separately supplied replacement content remain unresolved; no manual positive is established.); Extraction, review turn/spec admission and generated checkpoint writes: known — automatic [wired] (Caller launch/path configuration is not manual curation of every resulting item.); Trainer candidate reference admission and round decisions: known — automatic [wired] (Callback-internal human contributions remain unknown, but do not erase the known software reference gate.); Inference records and usage accounting production: known — automatic [wired] (Explicit output requests describe I/O, not content-admission authorship.); Initial harness/results and callback-internal candidate authoring: not-determinable (Generic caller identity and unavailable real editor internals do not establish explicit human content decisions or fully automatic initial admission.); Skill/artifact construction, caller replacements and standalone transcript payloads: not-determinable (Utility callers can supply manual or generated content and replace files; no explicit human decision over that content is established. Physical JSON writing alone does not classify admission.)	partial coverage (Both push and pull are established on named chains. Opaque consumer alternatives remain unresolved. Static role prompts and mere immediate returns are not accumulated-memory read-back.); Subsequent role context receives retained review turns: known — push [wired] (Trigger is the next scheduled role invocation; code builds its history context.); Checkpoint reentry and cached completed return: known — pull [wired] (Status/index selection within a requested return is not an additional unsolicited delivery.); Review/specification and next-round retained-harness delivery: known — push [wired] (Automatic workflow delivery is distinct from a consumer request for retained material.); Retained harness at explicit infer/infer_many: known — pull [wired] (This is an inspected request route, not evidence about an external factory store.); Skill artifact reload/merge caller: known — pull [wired] (IDs suppress duplicates inside a requested collection; they do not make it targeted push.); Opaque harness/metadata later-consumption alternatives: not-determinable (External callback/factory consumers and arbitrary metadata routes are not established. Known request routes do not resolve these opaque alternatives.); Round/outcome/usage audit and standalone transcript files: inapplicable (Immediate return/output storage alone is not a later consumer invocation. Checkpoint uses are separately classified above; no standalone transcript reader or outcome-to-training reload is wired.)	partial coverage (Coarse automatic supply is wired. There is no recorded lexical, embedding, judgment or identifier-based push selector; opaque alternatives prevent a complete selector inventory.); Automatic transcript history selection: known — coarse [wired] (Trigger: scheduled role call; input: full transcript; selected part: tail of retained role/content turns; no token budget.); Trainer specification and next-round harness supply: known — coarse [wired] (Trigger: editing/next round; selector input: current review result and retained reference; selected part: spec/harness. No semantic targeting is implemented.); Checkpoint reentry, explicit inference and artifact load requests: inapplicable (These are requested retained-material returns. Explicit paths, matching context/model and dedup IDs do not establish independent identifier-based push.); Opaque real harness and metadata consumer alternatives: not-determinable (Their automatic selection operations, if any, are unavailable; known pull branches cannot make unresolved alternatives inapplicable.)	partial coverage (A wired yes exists for retained trace-fed review guidance and the adapter-mediated update chain, with unresolved alternatives kept local. This axis uses the write-route definition, not the independently unestablished improved-capacity learning property in the reports.); Automatically retained review transcript/specification with later role/editor consumer: known — yes [wired] (This is wired trace-fed durable guidance, not demonstrated improved capacity or semantic acceptance.); Adapter-mediated retained harness update: partial — yes [wired] (Wired update/consumer chain is supported; real content responsiveness, initial lineage, external durable persistence and callback semantics remain unresolved. Quickstart ignores review text.); Separately supplied skill artifacts and caller replacements: not-determinable (Artifacts accept manual/model/arbitrary content without verified trace origin. Checkpoint replacement admission is caller-controlled. No qualifying alternative update can be inferred from a source label.); Raw extraction, outcome/audit and accounting retention alone: inapplicable (These are evidence/control/output objects rather than standalone learned guidance updates. Their use as inputs to qualifying review/harness updates is classified above; raw retention alone is insufficient.)	partial coverage (Authority is local to consumed parts. Learning inputs, downstream patch instruction, control validation/enforcement and ranking are distinct. No task-agent skill injection or parametric learning authority is inferred.); Trajectory evidence at review/editor update consumers: known — knowledge [wired], learning [wired], ranking [wired] (Ranking is influence of retained evidence under fixed code, not a learned ranking policy.); Review transcript at subsequent review calls: known — knowledge [wired], learning [wired] (Wired delivery does not establish actual criticism or improved capacity.); Specification at candidate editor: known — instruction [wired], knowledge [wired], learning [wired] (These forces are on the editor path. The synthetic editor ignores the spec; semantic compliance is not enforced.); Checkpoint control fields at reentry: known — routing [wired], validation [wired], enforcement [wired] (The equality gate has force without establishing comprehensive freshness or trust.); Artifact scope and identifiers at load/merge: known — validation [wired], enforcement [wired] (Scope checks are optional on some reads; first-ID suppression is not contextual relevance ranking.); Skill text/envelope returned to explicit structural test/API caller: known — knowledge [wired] (No standing instruction force at a task agent is established by this inspected consumer.); Synthetic harness revision at collector/executor: not-determinable (The revision determines a synthetic reward predicate and rendered output, but those effects do not establish a controlled knowledge/instruction/routing force. Program-level behavioral dependence is known; its authority category is not determined.); Generic retained harness at real collector/executor: not-determinable (Live argument delivery is known; opaque content and external consumer interpretation prevent assigning a particular controlled force.); Audit/outcome/usage parts and opaque metadata: partial — knowledge [wired] (Round audit and unused accounting/metadata have no evidenced consequential update consumer; opaque task outputs and arbitrary metadata prevent a complete authority assessment.)	AAS-2026-10-13-ecdysis-bb4c64dbdf60-01
echel	2	Echel-retained legacy project, knowledge, work, findings, run commands/summaries and verification evidence; canonical record families and access metadata; disposable index; transaction journals, staged records and migration backups; their admission, maintenance and later-consumer routes. Canonical specialized task, release, artifact-payload and learning-adoption routes remain unresolved. External referenced artifact bytes, Hermes/provider memory and model weights, static shipped skills, contributor source history and synthetic benchmark datasets/reports are excluded from this project-memory comparison. Transient compiled context is considered only as a consumption/derivation path from retained files and as text copied into retained run commands.	known coverage (Records cover the scoped retained inventory. External payload and provider substrates are excluded, not classified from metadata.); Legacy retained project, knowledge, work, findings, runs and evidence: known — files [wired] (The producer/read routes establish the legacy file inventory.); Canonical record envelopes and artifact access metadata, not external payload bytes: known — files [wired] (Generic storage covers even families whose specialized consumers are unresolved.); Disposable canonical search/relationship projection: known — sqlite [wired] (The retained database and its canonical source are distinguished.); Transaction and migration journals, staging and backups: known — files [wired] (Operational recovery objects are included, not generalized to all runtime storage.)	partial coverage (Automatic admissions and explicit human-decision protocols coexist; generic caller and bypass paths prevent complete agency coverage.); Automatic run/evidence retention, index rebuild admission and structural portable import: known — automatic [wired] (Classification concerns admission, not whether a human started the command.); Canonical protected decisions, finding dispositions, staleness and profile/maturity transitions under the human-decision protocol: known — manual [wired] (Manual protocol coverage is local; fabricated identity and direct writes are not authenticated or forced through it.); Initial/caller-authored records, direct store writes, finding/link creation and migration resolutions: not-determinable (Caller identity and content-level human control are unresolved. Physical software writes and explicit apply alone cannot decide manual versus automatic admission.)	partial coverage (Pull is implemented on assessed routes; unresolved included consumers prevent a complete pull-only conclusion.); Explicit work context/run request and canonical query/clarification/lifecycle/recovery requests: known — pull [wired] (Relationship selection remains part of requested context delivery, not a second independent push route.); Specialized canonical task/release/learning/artifact consumers and run/evidence later consumption: not-determinable (Storage affordances do not establish these consumer routes or their direction; no push absence claim covers them.)	not-determinable coverage (No positive push signal is established. This is not evidence that every included alternative lacks push.); Inspected explicit-request routes: inapplicable (These are pull operations. Their identifier membership and lexical ranking are not push signals.); Unestablished specialized canonical and retained trace consumer alternatives: not-determinable (No recorded push trigger, selector input and delivered retained part establish a controlled signal; unresolved directions prevent axis-wide inapplicability.)	not-determinable coverage (No qualifying automatic trace-to-durable-behavior-shaping-output-to-later-consumer chain is established; lack of such a record is not a system-wide no. External parameter learning is outside this memory scope.); Legacy execution-output and verification-result automatic writes: not-determinable (Trace-fed retention is wired, but durable behavior-shaping derivation and later consumer are unestablished; neither yes nor a bounded absence warrant for no is supplied.); Canonical learning, evidence-linked content and other caller-produced updates: not-determinable (Learning labels, evidence IDs and imported metadata do not establish automatic trace-fed production or later adoption. Caller-dependent provenance prevents a complete negative.)	partial coverage (Known consumer effects are retained separately from unresolved specialized consumers. No learning authority or universal approval enforcement follows.); Legacy objective/acceptance/verification directions delivered to Hermes: known — instruction [wired] (Static implementation skills are not counted as retained project instructions.); Legacy knowledge/findings in prompt and canonical content/reasons returned to query caller: known — knowledge [wired] (Delivery establishes the contextual path, not beneficial reliance.); Canonical searchable strings in FTS/BM25 ordering: known — ranking [wired] (No learned ranking policy is required or inferred.); Retained relationship, kind, stage, profile and recovery-state selectors: known — routing [wired] (These fixed-rule effects do not depend on semantic understanding of retained rationale.); Canonical eligibility and index/record freshness controls; legacy stage gate: known — validation [wired], enforcement [wired] (Retained controls have local enforcing force, not a system-wide product-truth security guarantee.); Specialized canonical task, release, artifact and learning adoption; legacy run/evidence later behavior-shaping uses: not-determinable (Generic storage/query does not establish specialized adoption or an update consumer; learning authority cannot be asserted from names or evidence IDs.)	AAS-2026-10-15-echel-08d27848b99a-01
echoes-vault-opencode	2	Included: accumulated page bodies/frontmatter, daily entries, generated index, persisted session/access bookkeeping, raw/assets, and their proposal, admission, maintenance and later-consumer routes. Excluded: shipped static guides/protocol/command templates, installed executable machinery, external host weights/provider state, host compaction and arbitrary future compatible runtimes. Initial origins and independent raw/asset routes remain unresolved, not excluded.	known coverage (The memory inventory is files throughout, including unresolved raw/assets. Host weights, installed executables and static templates are outside this accumulated-memory scope; repository locality alone does not add a repo substrate.); Page bodies/frontmatter, generated index and daily entries: known — files [wired] (The records inventory all managed knowledge files, including legacy daily files.); Persisted session/access bookkeeping: known — files [wired] (JSON is encoding, not a separate substrate.); Raw source and asset containers/payloads: known — files [wired] (Unknown payload encodings and producers do not obscure their recorded filesystem substrate.)	partial coverage (Known software admission is not erased by unresolved initial control. No manual positive is supported merely because users initiate lifecycle commands.); Managed append/upsert/end admission, including page replacements: known — automatic [wired] (Tool and direct CLI alternatives share software admission; a caller-provided confirmation is not evidence of human content approval.); Index/state regeneration and legacy summary migration: known — automatic [wired] (Starting maintenance does not make each generated-content admission manual.); Requested host session-distillation admission to proposed artifacts: known — automatic [claimed] (Classification retains the weaker evidence layer of semantic selection.); Initial/preexisting file admission and independent raw/asset writes: not-determinable (Generic caller identity and available containers do not establish explicit human decisions or automatic admission on these paths.)	partial coverage (Both request-driven project reload and periodic bookkeeping display are included. Opaque independent raw/asset consumption prevents complete direction coverage.); Start/search requests and their selected delivery of managed knowledge: known — pull [wired] (Automatic selection inside a requested return is still pull; result and count caps are not token budgets.); Host-directed complete-page reads: known — pull [afforded] (No actual complete-page reads are supplied.); Persisted session state on periodic TUI display: known — push [wired] (This is access/status memory delivery to the operator, not unsolicited project-knowledge injection into a model.); Raw/asset referenced payload delivery: not-determinable (References do not establish a dedicated retained-payload read-back operation or resolve independent host-tool routes.)	partial coverage (The coarse positive is local to TUI bookkeeping. Lexical matching on a pull request does not establish inferred-lexical push; page identifiers on index rows do not establish identifier-targeted push.); Periodic TUI push of persisted session/access bookkeeping: known — coarse [wired] (This coarse fixed-field display has no task query, embedding or identity-match targeting.); Requested start/search and afforded complete-page delivery: inapplicable (These are pull operations; recent filename selection, literal queries and requested paths are not push signals.); Independent raw/asset delivery alternatives: not-determinable (Whether a push route exists and which payload selector it uses is unresolved; known managed pull routes cannot make this branch inapplicable.)	partial coverage (A claimed qualifying route supports a local positive. Unknown input lineage prevents complete coverage; lack of background conversation capture does not negate foreground-agent distillation.); Requested session distillation to retained final notes/pages and later host context: known — yes [claimed] (This meets the trace-to-artifact-to-later-consumer shape at the prescribed layer, not observed learning or measured improved capacity. Wired persistence does not upgrade the semantic producer.); Arbitrary append/upsert inputs, legacy content and raw/asset-derived alternatives: not-determinable (The records do not establish whether these inputs are traces or derived behavior-shaping artifacts. No bounded trace-learning negative covers them.)	partial coverage (Project knowledge, concurrency criteria/veto and prescribed experience-to-artifact learning have distinct consumers. Static guides and executable machinery are excluded; neither supplies an instruction/enforcement finding for accumulated content merely by being installed. Literal inclusion and stable file order do not establish relative-priority ranking.); Page/daily/index content at later host and status bookkeeping at operator: known — knowledge [wired] (Delivery and display establish the local force, not compliance, factual endorsement or beneficial decisions.); Current retained page bytes at replacement preflight: known — validation [wired], enforcement [wired] (This is concurrency authority, not project-truth validation or authority of a retained schema.); Session experience at the requested distillation consumer: known — learning [claimed] (The update-consumer force remains prescribed; downstream output knowledge authority is independently wired.); Raw/asset payloads at later host-tool consumers: not-determinable (Reference syntax affords access but does not identify actual payload consumption or its authority.)	AAS-2026-10-15-echoes-vault-opencode-239c6738bb9e-01
EchoWiki	2	Retained raw notes/transcripts/audio; summary, concept and entity pages; descriptions, lineage fields, links, index and operation log; stored operator-edited AGENTS.md; persisted pending queue and operator configuration affecting admission/consumption; externally populated source/page/image/utility payload alternatives. Provider retention is unresolved, not presumed absent. Excludes static shipped instructions and fallback strings, current-call drafts/plans/task arrays, release artifacts, host rendering and model weights as computation components rather than established accumulated memory.	partial coverage (Files are supported throughout the local content inventory; host plugin-data substrate and possible provider retention prevent complete coverage.); Raw acquisition files, wiki bodies/frontmatter/navigation/log, stored AGENTS.md and external utility payload files: known — files [wired] (The recorded local object inventory resolves substrate independently of actual content and host durability.); Persisted plugin queue and edited configuration: not-determinable (Saved plugin data and reload are established, but these records do not identify the host's physical storage substrate; do not substitute files or service-object by assumption.); Possible provider-retained state from chat/STT use: uninspected (Remote infrastructure and parameter mutation are opaque; neither an accumulated provider-memory object nor its substrate is established.)	partial coverage (Both automatic and manual mechanisms are supported, with initialization, generic callers and providers unresolved.); Page/description/frontmatter/link/index/log admission, watcher queue insertion/removal and returned transcript admission: known — automatic [wired] (Human queue rejection is a separate upstream decision; physical persistence is not the admission classification.); Explicit note supply/copy, audio capture choice, queued input approval and operator guidance/configuration edits: known — manual [wired] (These human decisions coexist with automatic subsequent transcript and wiki admission.); Initial external wiki/raw/utility content and generic API/helper caller-supplied changes: not-determinable (Caller identity and file presence do not establish an operator per-content decision or autonomous caller admission.); Possible provider retention or parameter updates: uninspected (Remote admission/update control is uninspected.)	partial coverage (Request fulfillment and automatic context supply are different operations within the triggered workflow. Both are supported; unresolved consumers prevent a complete inventory.); Automatic context selection/delivery inside subsequent compilation and plugin reload: known — push [wired] (The finite chain is workflow-triggered, not a continuously running unsolicited-memory service; automatic context supply still constitutes push at the consumer boundary.); Current raw-file reads on compile/preview request: known — pull [wired] (Returning the requested file is pull, not a second push merely because software delivers it.); Documented human browsing of retained summary/index/page/link material: known — pull [afforded] (Documented consumer role establishes affordance, not observed operation.); External utility/media/log consumers and possible provider-memory consumption: not-determinable (Utility read capabilities do not establish a later invocation consuming accumulated content; log-to-compiler feedback is absent locally; provider alternatives remain opaque.)	partial coverage (Known automatic supply has coarse, judgment and identifier selectors at distinct stages. Unresolved alternatives prevent making the axis wholly inapplicable or known.); Compile-triggered brief enumeration and stored instruction/configuration/plugin-queue reload: known — coarse [wired] (Stored AGENTS.md availability selects guidance, not an identity match selecting among knowledge entries.); Planner selection of retained update targets and code resolution of those targets into delivered old bodies: known — inferred-judgment [wired], identifier [wired] (Judgment selects targets and exact path resolution selects delivered bodies; an identifier merely appearing in a catalog would not establish this signal.); Requested raw reads and documented human browsing: inapplicable (These operations fulfill retained-material requests; a requested path/link is not a push signal.); Potential external utility/media/provider push alternatives: not-determinable (No established later consumer/selector resolves whether push exists here or which controlled signal would apply.)	not-determinable coverage (A local log-feedback no is supported by bounded absence, while source origin and external/provider routes remain unresolved. This is not a system-wide no or an assessment of improved capacity.); Operation-log/quality-result feedback into subsequent compiler guidance: known — no [wired] (Complete classification of this bounded feedback route is no, not a system-wide absence of trace learning.); Pending event path/time queue to later approval UI and routing: inapplicable (Event metadata is retained access state for pending work, not a learned-parameter or derived behavior-shaping content update. This boundary reason establishes inapplicability, not a separate evidenced absence.); Automatic source-to-page/description updates with later planning/update consumers: not-determinable (Durable generation and later contextual use are wired, but notes/transcripts are not established as activity/experience traces by content and origin. Neither yes nor a universal no follows.); Explicit operator instruction/configuration edits: inapplicable (These retained updates are manual, not automatic trace-fed writes; the compiler does not revise its own instructions.); External utility writers and possible provider updates: not-determinable (Caller control, input origin, later consumer and provider retention/update internals are unresolved; no qualifying trace-fed chain is established.)	partial coverage (Authority is classified at actual local consumers; unspecified utility/media/provider consumption prevents a complete union.); Prior page descriptions/previews and entity type/source-count context at the planning model: known — knowledge [wired], routing [wired] (This records the implemented selection channel, not observed dependence or a relative-priority scoring mechanism.); Selected prior bodies at update generation; raw text at subsequent compilation: known — knowledge [wired] (Known contextual authority does not imply preservation, truth or learned capacity.); Summary/index/page prose and navigation at documented human browsing: known — knowledge [afforded] (Human reference use retains affordance strength rather than inheriting the compiler's wired strength.); Stored edited AGENTS.md at compilation models: known — instruction [wired] (Consumer/channel/force are explicit; compliance and benefit remain unobserved.); Retained target identities at generated-link checking and cleanup: known — validation [wired], enforcement [wired] (Retained identities contribute criteria with an enforced formatting effect; no eventual-existence guarantee follows.); Pending queue paths at approval UI/compile handoff and edited configuration at subsequent dispatch: known — routing [wired], enforcement [wired] (Routing/gating do not establish queue removal as a universal completion witness or failure-retention guarantee.); Log, opaque audio/images, external utility payloads and possible provider-retained state at unknown later consumers: not-determinable (Log-to-guidance is bounded absent; media and utility capabilities do not establish later memory consumers, and provider consumers are opaque. No ranking or learning authority is established for these alternatives.)	AAS-2026-10-14-echowiki-8d428f79523e-01
Eidetic	2	Included: accumulated personal cards and topic content; mixed text, metadata, Evidence and dense retrieval projections; unindexed handoffs and generated rules; session history/hints; judged-candidate, test/filing and usage/lifecycle control traces; managed maintenance regions; exported pages and export control state; unresolved dark child/report payloads and possible provider-side accumulated state. Excluded: pretrained model weights as static inference components, shipped static instructions/HINTS as independent memory, ordinary unretained current-run state, host conversation compaction, product-update executable replacements and unfrozen SDK application stores. Raw host transcripts are trace inputs to included derivations, not an asserted complete Eidetic archive.	partial coverage (Files and SQLite are supported for named local objects. Provider-side accumulation remains uninspected; no substrate value is inferred for it.); Personal cards, topic corpus, unindexed handoffs, generated rules and exported vault pages: known — files [wired] (The inspected durable text objects are filesystem artifacts; handoff producer identity does not obscure its file substrate.); Card text/metadata/event projections, dense indexes, Core logical-index state and session counters: known — sqlite [wired] (SQLite covers the inspected projections and session_log; vector encoding is not another substrate.); Judged-candidate cache, filing/test correlation, lifecycle/usage/passive logs and export manifest/cache: known — files [wired] (Named local JSONL/log/manifest parts retain later-selection and re-export state.); Provider-side retained state, if any, on extraction, embedding, reranking, translation and export routes: uninspected (Adapters do not inspect provider retention or updates through use. Pretrained weights are not classified as accumulated memory merely because inference uses them.)	partial coverage (Automatic local mechanisms and afforded/claimed manual controls are retained separately. Generic direct callers and initial/handoff/provider admission prevent complete coverage.); Stop signal compounding and M3 support-gated filing/outcome-cache admission: known — automatic [wired] (Code/model decides admission with no per-content operator approval; configuration and starting hooks do not make it manual.); Index/vector/code projection, generated rules and session-counter/hint admission: known — automatic [wired] (Code derives and installs these objects without a per-entry human content decision.); M1/M2/event/confidence updates, test promotions and decay: known — automatic [wired] (Activation switches differ from per-content approval; the test route is conditional on matching identities.); Vault export/polish/topic, delta control state and cleanup candidate/log deletion: known — automatic [wired] (Invocation chooses a mode, but software selects and writes each resulting retained transformation.); Direct personal writer and base-add caller content: not-determinable (Caller identity is unspecified; body supply alone does not establish explicit human content approval. Base MCP describes but does not enforce a user-only policy.); Base MCP explicit-user-only policy: known — manual [claimed] (This is a claimed per-content human-request policy, separate from the ungated implemented caller route.); Human source-card/Evidence/declaration editing and reversal: known — manual [afforded] (The reports establish available human edit/reversal surfaces, not observed edits or a universal human gate.); Initial existing files and fresh handoff admission: not-determinable (Scanner/read selection does not determine who supplied or approved the retained file.); Opaque callback/child dark payload and provider-side admission: not-determinable (Known local commits remain automatic; initial opaque payload control and provider admission remain unresolved.)	partial coverage (Both requested pull and independent push chains are supported; opaque later consumers prevent complete direction coverage.); Requested personal/base recall/detail, serendipity and Core client search: known — pull [wired] (A caller requests retained material; returning it is not a second unsolicited push. External SDK agent consumption is not wired by this classification.); SessionStart card context, fresh handoff and count-selected hint: known — push [wired] (Lifecycle code independently supplies selected retained material before foreground work, under the host loading contract.); Automatic maintenance neighbors and M3 cache/source-support selection: known — push [wired] (Changed-card/session triggers cause selectors to supply retained neighbors, outcomes or source spans to update/check consumers.); Requested export/curation and their manifest/usage/page consumers: known — pull [wired], push [wired] (CLI/MCP requests start these reads; automatic note/source selection within the workflow is distinguished from the requested output.); Human vault browsing: known — pull [afforded] (The declared readable vault supports requested browsing without supplied actual use.); Dark report/child payload consumers and possible provider-state read-back: not-determinable (Known cache retry consumers are covered above; detailed report and provider selection/delivery operations are not recorded.)	partial coverage (Coarse, identifier, inferred-lexical, inferred-embedding and inferred-judgment selectors are local findings; opaque alternatives prevent complete signal coverage.); SessionStart feedback availability and budget/recency/status selection: known — coarse [wired] (Trigger is SessionStart; type, age, weights and budget select feedback/project/recent representations independently of a query.); SessionStart project matching and fresh handoff location: known — identifier [wired], coarse [wired] (Trigger is SessionStart; normalized CWD/project match selects project chunks and CWD/project-pattern state files. Age/newest/budget filtering remains a separate coarse signal.); SessionStart adaptive guidance from prior session_log: known — identifier [wired], coarse [wired] (Trigger is SessionStart; project identity selects retained rows, count thresholds select the emitted hint, not semantic similarity.); Changed-card M1/M2 neighbor supply and gate-selected retained content: known — identifier [wired], inferred-embedding [wired], inferred-judgment [wired] (Trigger is changed ingest; same-project filters and vector/relevance/conflict judgments select neighboring card material for later maintenance.); M3 same-session cache and re-retrieved support/dedup material: known — identifier [wired], inferred-lexical [wired], inferred-embedding [wired] (Trigger is later Stop/mined candidate; session/candidate identity suppresses processed entries, while query retrieval/dedup selects retained source spans and neighbors.); Post-filing test promotion producer: known — identifier [wired] (Trigger is enabled later producer; exact session/project matches select filed pages and corresponding retained test signals. The ordinary hooks do not establish matching project identities.); Enabled vault topic-model selection: known — coarse [wired] (Trigger is requested export with topic generation enabled; weighted note ordering, top-20 selection and text cap supply retained pages to the model.); Requested recall/detail/discovery and human browsing delivery operations: inapplicable (These delivery operations fulfill specific consumer requests. Their query/file IDs alone do not establish identifier-based push.); Dark child/report and provider-state selectors: not-determinable (Known local cache selection does not resolve the remaining automatic consumers or provider payload.)	partial coverage (Several local yes chains are wired. Unresolved admissions and provider behavior preclude a complete trace-learning inventory; no unsupported no is authored.); Stop session extraction to durable signals/History and later injection/recall: known — yes [wired] (Automatic session-fed writes produce later context; neither complete transcript retention nor demonstrated improvement is required.); M3 recall filing/M2 handoff and judged-outcome retry cache: known — yes [wired] (Session candidates yield durable answer/managed-region material or definitive outcome controls consumed on later invocations. Dark would_file is not a card write.); Session history to durable count-selected behavioral hint: known — yes [wired] (Session events feed retained history and generated guidance; selecting fixed hint text does not exclude a trace-fed behavior-shaping update.); Test/filing correlation to typed Evidence and later confidence/ranking: known — yes [wired] (Automatic promotion capacity requires matching supplied session/project records and flags. Ordinary hook promotion is not established because slugs differ.); Static topic imports, optional source-code entity projection and vector rebuild from supplied records: inapplicable (The named routes acquire/index content, not activity/experience. Source-code bodies are not tool/session traces simply because indexing runs in a hook.); Caller-supplied personal bodies, initial/handoff admission and opaque callback/dark report routes: not-determinable (Origin and automatic control are not jointly established for all included alternatives; no local no or system-wide negative follows.); Provider-side trace-fed durable updates: uninspected (Call-site inference without training requests does not establish absence of external provider updates.)	partial coverage (Knowledge, instruction, ranking, learning, routing, validation and enforcement have local consumers. No force or completeness is inferred for opaque payloads or downstream host compliance.); Non-feedback card/snippet/handoff/answer text at foreground recall or context reader: known — knowledge [wired] (Delivery is wired context/evidence supply, not observed reliance or compliance.); Feedback text at foreground host rules loader: known — instruction [wired] (Standing directions are delivered under ALWAYS APPLY; name-only/cluster delivery is weaker than full-rule delivery.); Session-count-selected hint at foreground host: known — instruction [wired] (Fixed hint language is selected by accumulated history; the static HINTS table is not itself learned memory.); Card text, dense vectors, source/status/confidence metadata and Evidence at retrieval/assembly/maintenance selectors: known — ranking [wired] (Content scored by fixed lexical/vector code still supplies ranking input.); Retained transcript experience at signal/M3 durable-card update consumers: known — learning [wired] (Automatic session-derived writes shape later available context; this is learning authority at the update consumer, not improved capacity.); Retained session history and test/filing traces at adaptive-hint and Evidence update consumers: known — learning [wired] (Count-selected guidance and gated correlated-test promotions are separate durable behavior-shaping updates. Ordinary hook promotion remains limited by project-identity mismatch.); M3 judged cache at candidate retry/selection: known — routing [wired], enforcement [wired] (Definitive cached outcomes suppress same-session candidates; transient outcomes retry.); Retained metadata/source spans at maintenance and filing gates: known — validation [wired], enforcement [wired] (Event types/authority criteria and retrieved spans are retained inputs to admissibility decisions, not authenticated truth.); Retained project/path/count/hash/status fields at scope, hint and delta-export routing: known — routing [wired] (These are actual operation/window selectors, not inert catalog identifiers.); Exported pages at human vault reader: known — knowledge [afforded] (Human-readable vault browsing is supported as an affordance, not an observed reader invocation.); Dark acquisition/child report payloads, optional callback effects and provider retained state: not-determinable (Known retry outcomes are classified above; exact other report consumers/effects and provider behavior are not resolved.)	AAS-2026-10-14-eidetic-0e54f3c114b3-01
eigenius	2	Included: use-accumulated document decision records, task result artifacts, layer-resident knowledge/program/proof/statistical resources, published notebook source, numerical access segments and rebuilt institution access state, with their admission, maintenance and later-consumer routes. Included unresolved alternatives: initial/bootstrap/import control, sense/proposal rationale use, full text-index derivation, local/offline mutations, reflection, autosave/merge persistence and durable task recovery. Excluded: shipped static ontologies/instructions/prompts as such, unretained current-document working context, external model weights/services and external dataset bytes beyond their retained descriptors and inspected consumer seams. The artifact remains a partial operational loop.	partial coverage (The accepted objects resolve graph/layer, metadata, vector and rebuilt dispatch storage; physical backend internals and the remaining retained-state inventory prevent complete coverage.); Layer corpus, ProposalDraw envelopes, published cells, retained programs, proof inputs, verification traces and statistical results: known — graph [wired] (These are retained typed resource graphs on parent-linked layers, not a repository merely because the implementation is in Git.); Formalization task result metadata on the persistent backend: known — kv [wired] (The task-UUID result key addresses a retained metadata blob. CBOR is an encoding, not a substrate.); Retained vector segments and optional HNSW access structure: known — vector [wired], graph [wired] (The vector store and its optional graph are separately meaningful access structures; neither is provider weights.); Rebuilt institution index/runtime during the serving process: known — in-memory [wired] (This is derived access state retained for subsequent dispatch, distinct from durable declarations.); Other backend, notebook, task and import storage alternatives: uninspected (Bootstrap/import admission, wider institution and reflection routes, local/offline mutations, full text-index maintenance, autosave/merge persistence and durable task recovery remain included but uninspected. The recorded inventory is not a whole-system enumeration.)	partial coverage (Software-controlled admission is established on concrete routes. Generic caller identity leaves human control of other writes unresolved; browser publication alone does not certify per-claim human approval.); Live draw recording, generated result retention and optional notebook land: known — automatic [wired] (These writes do not require an explicit human decision on each retained output.); Program/query output, domain-result disposition, index refresh and maintenance transformations: known — automatic [wired] (Starting a query or maintenance job is not manual admission of each computed result; full maintenance CAS/recovery is outside this control finding.); Explicit browser editing/publishing of notebook source: known — manual [afforded] (A human can explicitly supply/edit/publish retained notebook content; generic SDK publication has a separate unresolved principal.); Generic Load, inline records and SDK publication initial content decisions: not-determinable (Callers may be human or automated. Code performs I/O, but that alone cannot classify the upstream per-content decision.); Bootstrap/import, offline mutation, autosave and durable recovery admission: uninspected (Bootstrap/import admission, wider institution and reflection routes, local/offline mutations, full text-index maintenance, autosave/merge persistence and durable task recovery remain included but uninspected. The recorded inventory is not a whole-system enumeration.)	partial coverage (Requested returns coexist with automatic internal selection. Push is assigned only to the independent retained-part selectors inside the computation, not a second delivery of requested query/notebook results.); Explicit query/inspect, notebook reopen, result fetch and program-by-IRI requests: known — pull [wired] (These consumers ask for retained material or a specifically named retained program; code fulfilling the request is not an additional push.); Document-branch draw loading followed by keyed decision replay: known — pull [wired], push [wired] (The document request selects the branch/recording source; within it, a replay implementation automatically matches each decision question and supplies the saved answer to the parser.); Post-commit institution index/runtime refresh: known — push [wired] (A commit automatically refreshes retained declaration-derived dispatch context for later operations; it is not an unsolicited chat-context route.); Chain-attached witness-dependent checking: known — push [wired] (A later checking request causes automatic evidence lookup; the checking consumer receives selected retained credit rather than a user-requested trace document.); Alternative CLI, reflection, autosave and durable task consumer directions: uninspected (Bootstrap/import admission, wider institution and reflection routes, local/offline mutations, full text-index maintenance, autosave/merge persistence and durable task recovery remain included but uninspected. The recorded inventory is not a whole-system enumeration.)	partial coverage (Signals are classified on actual push selectors, not borrowed from pull-query similarity or an identifier attached to a requested notebook. Untraced delivery alternatives prevent complete coverage.); Keyed replay to parser/resolver/lander on a formalization decision: known — identifier [wired] (Trigger: decision seam invocation. Input: question key. Selected part: matching saved choice/rank/proposal/kind. Selection/proposal use context-rich keys; kind uses sentence/gloss only.); Post-commit declaration-derived dispatch-context installation: known — coarse [wired] (Trigger: commit hook. Input: new committed chain. Selected retained material: available institution declarations used to rebuild the runtime; no relevance-ranked or per-task declaration selection is established.); Witness credit supplied during chain-attached checking: known — identifier [wired] (Trigger: witness-dependent check. Inputs: category, target IRI, normalized proposition and dependencies. Selected retained part: matching local/ancestor proof judgement or trace credit.); Push selectors on untraced reflection, task recovery and alternative consumers: uninspected (Bootstrap/import admission, wider institution and reflection routes, local/offline mutations, full text-index maintenance, autosave/merge persistence and durable task recovery remain included but uninspected. The recorded inventory is not a whole-system enumeration.)	partial coverage (Automatic recording of decision experience into durable replay artifacts supports a local yes. Raw program traces, index derivation and uninspected recovery alternatives do not establish a system-wide no or complete inventory.); Live formalization decision exchanges to retained draws to later replay: known — yes [wired] (The automatic recorder extracts executed decision questions/answers into persistent structured artifacts that change later parser/resolver/lander behavior. Branch reset limits horizon but does not negate the wired next-invocation route.); Retained program traces and durable task recovery alternatives: not-determinable (Trace retention is known, but a later trace-derived policy/behavior artifact and full recovery consumers are not established. This is not bounded absence of every qualifying write.); Vector refresh over resource material: not-determinable (Automatic durable numerical derivation is wired, but original indexed content is not established as trace experience. No trace-learning yes or no is inferred from asynchronous indexing alone.); Other imports, reflection, maintenance and recovery learning alternatives: uninspected (Bootstrap/import admission, wider institution and reflection routes, local/offline mutations, full text-index maintenance, autosave/merge persistence and durable task recovery remain included but uninspected. The recorded inventory is not a whole-system enumeration.)	partial coverage (Authority is attached to each actual consumer, not assigned uniformly to persisted content. Supported ranking, routing, instruction, knowledge, validation, enforcement and learning paths do not resolve untraced consumers.); Recorded selection, ranking, proposal and classification fields at replay consumers: known — routing [wired], ranking [wired], enforcement [wired] (Replay supplies operational decisions rather than optional background text. The classification concerns returned fields, not uninspected rationale interpretation.); Resource and vector material at retrieval and notebook inspection consumers: known — knowledge [wired], ranking [wired] (Returned reference/source content and similarity ranking are distinct effects, neither a truth certificate.); Retained completion text at the CompleteText model: known — instruction [wired] (Prompt guidance has instruction force, not guaranteed compliance or a grant to register capabilities.); Retained program symbolic body at IO evaluation and institution declarations at dispatch: known — routing [wired] (Execution/dispatch paths choose supported operations under the active registry; symbolic execution is not evidence that all contained text is instruction.); Retained proof judgements and dependency state at witness-dependent type checking: known — validation [wired], enforcement [wired] (Formal proof credit and dependency checks have bounded conformance and veto effects. Declared/Observed attribution and ProgramTrace do not inherit Verified authority.); Live recorded decision experience at durable replay-artifact production: known — learning [wired] (The recorder converts prior decision experience into durable behavior-shaping replay implementations, independently of downstream ranking/routing force.); Untraced rationale, imported policy, institution and recovery consumers: uninspected (Bootstrap/import admission, wider institution and reflection routes, local/offline mutations, full text-index maintenance, autosave/merge persistence and durable task recovery remain included but uninspected. The recorded inventory is not a whole-system enumeration.)	AAS-2026-10-15-eigenius-fe68ed89519b-01
Engraph	2	Retained vault Markdown; SQLite file/chunk/FTS/vector/graph/tag/access structures, identity facts, folder centroids, correction/CLI telemetry, query cache, unresolved links and migration history; use-admitted identity configuration, detected vault profile and saved migration previews; their admission, maintenance and later client/internal consumers. Includes unresolved incoming-content control/provenance, telemetry consumers, graph-output details and withdrawal derivatives. Excludes shipped static prompts/skills, model weight caches and opaque model internals, transient search/context/health outputs as retained objects, HTTP permission configuration, product source/releases and external conversational-agent state.	known coverage (The memory inventory separates source files, use-admitted configuration and previews from SQLite derivatives. Loaded static model weights, shipped prompts, product source/releases and transient outputs are excluded; numerical encoding alone is not a model-weights substrate.); Source vault Markdown bodies and frontmatter: known — files [wired] (Source files remain separately retained and reread from disk.); SQLite file/chunk/FTS/vector/edge/tag structures, identity facts, centroids, query cache, correction and migration logs, unresolved links and CLI telemetry: known — sqlite [wired] (The records identify the persisted SQLite structures, including logs whose operative consumers remain unresolved.); Use-admitted config identity, detected vault profile and saved migration previews: known — files [wired] (TOML configuration/profile and JSON/Markdown previews are filesystem artifacts.)	partial coverage (Automatic transformation/admission is positive; caller-origin control and telemetry admission remain unresolved. Authorship and physical I/O are not substituted for agency.); Index admission, chunk/edge/tag/vector transformations, detected profile, identity extraction and cache admission: known — automatic [wired] (Software/model admission occurs without a separate per-derived-content human decision; starting index/setup/search does not make it manual.); Writer tag/link/placement transformations and filesystem-feedback updates: known — automatic [wired] (Caller proposal is separate from automatic transformation. Feedback code accepts qualifying moves without authenticated per-correction approval.); Original note text, edit/replacement/withdrawal requests, supplied identity and migration apply/undo decision: not-determinable (Generic CLI/MCP/HTTP or external editor identity does not determine whether an explicit human supplied or approved each retained content change; no manual value is established.); CLI-event admission call sites: uninspected (Append/prune APIs alone do not identify the active write/admission mechanism.)	partial coverage (Requested content return and internal automatic supply of retained selector/update state are different operations. Watcher persistence alone is not context push. Untraced included consumers prevent a completely assessed pull-only or complete direction inventory.); Requested search/topic/read/section/list/person/project/map/identity returns: known — pull [wired] (Consumer requests retained material; internal delivery is not counted again as push.); Automatic cache, profile, tag-registry and centroid supply to later operations: known — push [wired] (Consumers request an operation, not these specific retained selector objects; code automatically chooses/supplies retained policy/access state.); Automatic selection of indexed inputs for identity extraction and move-fed adaptation: known — push [wired] (Refresh/move triggers cause code to supply selected retained inputs to update consumers without a request for those individual parts.); Migration preview apply/history undo and requested health diagnostics: known — pull [wired] (Apply/undo explicitly requests retained proposed/history state; health requests diagnostic evidence over the retained index.); Correction-log later consumers, CLI telemetry and independently uninspected graph-output alternatives: uninspected (Storage APIs are not a consumer route. The memory coverage table retains incomplete producer/consumer and standalone graph-output inspection.)	partial coverage (Signals characterize the named automatic supplies only. IDs on requested reads are not identifier push, and requested search delivery is not another embedding push. Opaque/untraced alternatives remain unresolved; no inferred-judgment finding is established for retained-part push selection.); Same-query cache push on search with an orchestrator present: known — identifier [wired] (Trigger: later query search. Selector input: original query SHA256. Selected part: cached orchestration JSON. No model/age match is imposed.); Automatic semantic folder selection during later note writes: known — inferred-embedding [wired] (Trigger: write without explicit folder or successful type rule. Input: new-content embedding versus saved folder vectors. Selected part: best above-threshold centroid/folder.); Registered tag resolution during writer transformations: known — identifier [wired], inferred-lexical [wired] (Trigger: proposed tags. Input: supplied tag spelling. Selected part: persisted registry names through exact/fuzzy/prefix matching.); Loaded profile supply to placement/context/identity and fixed identity-input extraction: known — coarse [wired], inferred-lexical [wired] (Profile load supplies available configured roles. Extraction selects project/date/path/edge-ranked and keyword-matching retained parts. These mechanisms are not model judgment.); Move-triggered vector supply to centroid update: known — identifier [wired] (Trigger: detected move. Input: matched file/content identity plus provenance and suggested-folder conditions. Selected part: that indexed note's chunk vectors.); Untraced CLI/correction consumers and remaining graph selection detail: uninspected (Their direction/selection wiring is not established, so selector signals cannot be completed or declared inapplicable.)	partial coverage (A wired yes is supported for move-fed centroids. Other included origins/consumers remain unresolved, so neither a system-wide no nor complete enumeration of qualifying routes is warranted.); Qualifying move-fed placement adaptation: known — yes [wired] (Detected activity feeds an automatic persistent selector update subsequently consumed by semantic placement; no parameter training, authenticated human correction or benefit is required for this route classification.); Identity derived from arbitrary vault notes and query-policy cache origin: not-determinable (Derived content has later consumers, but accepted records do not establish that original inputs are recorded sessions/tool experience rather than arbitrary content/current requests. This prevents deciding additional qualifying trace-fed routes.); CLI telemetry producer and later behavior-shaping consumer: uninspected (Raw-log storage/pruning does not establish or rule out a qualifying automatic update. Full producer/consumer call sites remain uninspected.); Other initial content, profile and ordinary centroid derivations: not-determinable (The inspected derivations are identified, but arbitrary source experience and original write control are unresolved; there is no bounded absence record warranting a local no.)	partial coverage (The known consumers establish knowledge, ranking, routing, learning and bounded append validation/enforcement. Untraced log consumers prevent a complete authority inventory. Read-only/HTTP grants and shipped instructions are runtime controls, not use-derived memory authority.); Note bodies/snippets, identity values and L0 identity at read/context/identity clients: known — knowledge [wired] (These parts supply reference/context. External obedience or reinterpretation as standing instructions is unobserved and outside the service boundary.); Chunk text/vectors, graph/date metadata and query cache at retrieval selectors: known — ranking [wired], routing [wired] (Retained candidate content and relationships influence relevance/order even with fixed ranking code.); Folder centroids at semantic placement and old chunk vectors at feedback update: known — learning [wired], ranking [wired], routing [wired] (The update consumer and downstream placement consumer have different effects; explicit folders and type rules can bypass semantic placement.); Profile roles, file metadata, tag registry and migration destinations/history at machine consumers: known — routing [wired], validation [wired], enforcement [wired] (Profile roles select folders; identifiers resolve reads; registered tags influence admitted names; saved migration paths select forward/reverse moves.); Stored snippets/edges/metadata and unresolved-link rows at health diagnostic consumer: known — knowledge [wired] (These are evidence inputs for an advisory structural report, not retained admission criteria. Fixed diagnostic code is excluded static machinery.); Placement correction rows, CLI telemetry and migration preview display at unresolved or optional consumers: partial — knowledge [afforded] (Placement consumes centroids rather than correction rows; CLI producers/consumers are incompletely traced. Preview Markdown is available for review without established review operation.)	AAS-2026-10-14-engraph-0cd4f00695b0-01
Enoch	2	Retained use-accumulated facts and metadata; installed private identity and its unresolved initialization; session access mappings and opaque native retained payloads; conversation/action/task journals and retry/publication state; retained task/backlog/cron briefs; evidence signals/checkpoints, candidates, curations, work artifacts and lineage inbox material, with their write, maintenance and later-consumer paths. Includes unresolved provider compaction, ancillary maintenance and deployment-specific plugin/profile state. Excludes static shipped prompts/body identity/skills, inference weights, generic permissions and repository software as storage merely because tasks can change it. Transient completion/diagnostic views are classified through their retained inputs/results and later-consumer routes, not as independent durable stores.	partial coverage (Local file objects are identified individually. External native payloads, scheduled/workflow storage details and plugin objects prevent complete inventory classification. Transient completion payloads are views of journals, not extra retained stores; model weights and static shipped code are excluded.); Long-term facts and their access/provenance metadata: known — files [wired] (The record identifies the file-backed object and its local persistence.); Installed private personal identity: known — files [wired] (The record identifies the file-backed object and its local persistence.); Codex and Claude native-session access mappings: known — files [wired] (The record identifies the file-backed object and its local persistence.); Conversation turn archives: known — files [wired] (The record identifies the file-backed object and its local persistence.); Task lifecycle journal: known — files [wired] (The record identifies the file-backed object and its local persistence.); Extracted evidence signals: known — files [wired] (The record identifies the file-backed object and its local persistence.); Evidence scan checkpoints: known — files [wired] (The record identifies the file-backed object and its local persistence.); Durable evolution candidates: known — files [wired] (The record identifies the file-backed object and its local persistence.); Semantic curation history: known — files [wired] (The record identifies the file-backed object and its local persistence.); Learning-artifact index and Markdown work records: known — files [wired] (The record identifies the file-backed object and its local persistence.); Lineage inbox source, assessment and adoption state: known — files [wired] (The record identifies the file-backed object and its local persistence.); Request-keyed action-recovery journal: known — files [wired] (The record identifies the file-backed object and its local persistence.); Retained task/backlog/cron briefs and retry/publication diagnostics/state: not-determinable (Retention and later consumers are recorded, but these records do not specify all underlying storage parts; no substrate inferred from workflow terminology.); Provider-native session payload: not-determinable (External harness/service retention is recorded; its actual storage implementation is opaque. A local ID map does not classify that payload.); Deployment-specific extension/profile memory: uninspected (Installed implementations and their retained objects are uninspected.)	partial coverage (Automatic derived admission and an explicitly human decision branch coexist. Initialization, generic APIs, provider maintenance and plugins remain unresolved; no global manual-control guarantee follows from approval labels.); Final-reply fact admission: known — automatic [wired] (Explicit/high defaults do not establish manual admission.); Snapshot, activity injection, evidence scan/synthesis, brainstorming and curation writes: known — automatic [wired] (Operator workflow initiation is not manual admission of the resulting content.); Session mappings, raw journals, checkpoints, learning artifacts and workflow stages: known — automatic [wired] (Physical file I/O is not the basis; recorded host admission is.); Published skill and lineage assessment admission: known — automatic [wired] (Source selection and later adaptation approval are separate decisions.); Automatic candidate-source retirement and model-issued removal/approval controls: known — automatic [wired] (A human actor string alone does not classify all invocations as manual.); Explicit human-requested removal/adaptation selection: known — manual [wired] (Manual is local to the human-decision branch; generic handler invocation is not authenticated as that branch.); Initial content, explicit library fact replacement, identity install/clear and generic command callers: not-determinable (Caller identity/control is unspecified for these capability surfaces; no blanket human or automatic classification.); Provider-native compaction/checkpoint admission: not-determinable (Provider producer and any operator replacement controls are unresolved.); Deployment-specific plugin admissions: uninspected (Installed admission mechanisms are uninspected.)	partial coverage (Pull requests and independent automatic supplies coexist and are separated within chains. Opaque provider/plugin alternatives prevent a complete direction inventory.); Startup fact/private-identity rendering: known — push [wired] (Both application and adapter injection layers are included; double delivery is possible.); Snapshot creation and later task delivery: known — pull [wired], push [wired] (Requested summary return is pull, not an additional push; later worker supply is a separate operation.); Activity and scheduled evidence/curation inputs: known — push [wired] (An upstream scan trigger is not a model request to recall the selected payload.); Named candidate/lineage approval context and administrative fact display: known — pull [wired], push [wired] (The requested display is pull; subsequent unsolicited worker-context supply is push.); Native resume access and recovery/publication lookup: known — push [wired] (External native content selection remains unresolved separately.); Declared descendant learning-artifact inspection: known — pull [claimed] (This is declared guidance only, not wired deployed read-back.); Provider-internal retained parts and reload alternatives: not-determinable (Resume wiring does not establish every internal request/delivery operation.); Plugin read-back: uninspected (Installed read/selection paths are uninspected.)	partial coverage (Push selectors include coarse filters, real identity matches and model relevance judgment. No inferred lexical/embedding selector is established by the accepted records. That is not a bounded absence of such alternatives.); Startup injection trigger and confidence/date budget selection: known — coarse [wired] (No query-conditioned semantic or lexical targeting is recorded; displayed IDs do not establish identifier selection.); Snapshot selection for retained worker brief: known — inferred-judgment [wired] (Trigger is task/backlog/cron creation; selector input is prior chat and summary objective; selected retained part is the bounded brief.); Activity summary selection at note flush: known — coarse [wired] (Trigger is note flush; operation result and nonempty queued-note availability are selector inputs; delivered part is the clipped operation note, not the worker transcript. The conversation key routes that note to a consumer session; it does not identity-match retained note parts for delivery and establishes no identifier signal in this unit.); Evidence scan input/checkpoint selection: known — identifier [wired], coarse [wired] (Manual/due scan triggers selection from logged turns/task events; checkpoint identities do real filtering, unlike merely displayed refs.); Candidate/completion selection and downstream recommendation: known — coarse [wired], inferred-judgment [wired], identifier [wired] (Curator trigger is proposal construction; candidate/completion fields are inputs; selected retained parts are candidate snapshots and reasons. Semantic removal suggestions do not themselves delete.); Session and task/publication recovery matching: known — identifier [wired] (Continuation/retry is trigger; matching identity selects operative retained state. Native internal payload selection is not inferred from this local match.); Provider-internal selection among retained transcript/checkpoints: not-determinable (Opaque payload and compaction prevent complete signal classification.); Plugin selection signals: uninspected (Installed selectors and retained parts are uninspected.)	partial coverage (Supported yes routes coexist with unresolved writes and consumers. No bounded absence records justify a local or global no; this axis is independent of demonstrated learning or self-improvement.); Automatic feedback/experience to signals/candidates to curator/worker: known — yes [wired] (Wired retained guidance qualifies without installed software, improved capacity or observed benefit.); Conversation snapshot to retained task/backlog/cron brief to worker: known — yes [wired] (Continuation/brief summaries qualify; fidelity and scheduled refresh are separate limits.); Operation-result activity note to chat session: partial — yes [wired] (Host chain is wired, but exact external retention/compaction and later use are uninspected.); Brainstorm candidates using completed-work summaries: partial — yes [wired] (Theme-only brainstorming is not converted to a trace-fed path; full original provenance of all completion inputs is unresolved.); Final-reply remembered facts and generic initial/library writes: not-determinable (Automatic fact retention is known but complete original input/provenance is not; no blanket trace yes or no is inferred from a memory marker.); Learning-artifact descendant update: not-determinable (Trace-derived work record is retained, but concrete later behavior-shaping consumption is not wired in accepted records; declared inspection advice does not close that route.); Foreign-skill and lineage input history: not-determinable (Imported static skill content alone is not trace origin; full parent/source input content is not established enough for a complete trace-fed-route classification.); Native provider compaction/checkpoint alternatives: not-determinable (Automatic compaction and alternative retained outputs are uninspected; plain raw transcript accumulation is not itself a qualifying update.); Plugin trace-fed updates: uninspected (Installed transformations/later consumers are uninspected.)	partial coverage (Consumer-specific authority includes update learning, downstream context/guidance and executable selection controls. Provider/plugin effects and incomplete maintenance producers prevent complete enumeration. Static capability policy and activated source are not substituted for retained memory.); Fact prose at startup model: known — knowledge [wired] (Model compliance and actual reliance remain unobserved.); Fact confidence/time metadata at startup selector: known — ranking [wired] (Ranking input need not be a learned ranking policy.); Private identity text at startup model: known — instruction [wired] (This concerns installed private identity, not static body identity.); Conversation/task trace content at automatic extraction and brief generation: known — learning [wired] (Learning authority at update consumer does not establish improved capacity.); Derived signals and candidate text at synthesis, curator and later worker: known — knowledge [wired], learning [wired], instruction [wired] (These are separate update, reference and worker-guidance consumers; proposal retention is not deployment.); Task snapshot at later worker: known — knowledge [wired], instruction [wired] (Clipping and summary correctness are not independently verified.); Retained diagnostic prose at later repair-model invocation: known — knowledge [wired] (This unit covers retained diagnostic prose consumed through head/tail excerpts on the bounded repair path, separately from doctor predicates and publication receipts. Complete diagnostic preservation and successful repair remain unestablished.); Activity summary injected into later chat-native context: known — knowledge [wired] (Exact provider retention and reliable recall remain unresolved.); Session mapping/version, scan checkpoints and action/publication state at host: known — routing [wired], enforcement [wired], validation [wired] (Validation finding is local to publication expected-state checks, not every listed field.); Candidate scores/status, evidence linkage and curation results: known — ranking [wired], routing [wired], validation [wired], enforcement [wired], knowledge [wired] (Suggestions do not remove entries by themselves; actor labels do not authenticate human approval.); Lineage source/assessment and linked state: known — knowledge [wired], instruction [wired], routing [wired], enforcement [wired] (Stored adoption is not evidence of restarted use.); Learning-artifact advice at declared descendant reader: known — knowledge [claimed] (Only declared reader guidance, not a wired descendant invocation.); Native context internal part/consumer authority: not-determinable (Known injected text does not determine complete stored payload or provider authority construction.); Plugin retained state consumers: uninspected (Installed consumers and forces are uninspected.)	AAS-2026-10-13-enoch-a73695a57e3b-01
equipa	2	Included: accumulated episode/lesson prose and access metadata, embeddings and graph edges, hard/soft checkpoints, reconstructed sessions and cross-invocation summaries, project records, initiative appends, inter-agent messages, optional output offload, and through-use prompt/configuration variants and their maintenance metadata. Included unresolved alternatives: initial-content admission/provenance, provider-owned compaction/state, arbitrary hooks/plugins, complete optimizers/scorers/rollback, training metadata consumers, housekeeping/setup/migration/sync and alternate callers. Excluded: unchanged shipped instructions, ordinary current-call conversation/REPL accumulation, general runtime storage/permissions, external model weights merely used for inference, and product Git/source changes except the particular retained initiative/prompt artifacts. No live deployment or execution evidence is asserted.	partial coverage (The named retained-object inventory establishes filesystem, repository, SQLite and process-memory witnesses. Opaque provider memory and remaining shipped maintenance alternatives prevent complete substrate coverage; no graph/vector/model-weight substrate follows from an encoding or model name.); SQLite episode prose, lessons, project records, messages, session state, Q/effectiveness metadata, vectors and graph-edge rows: known — sqlite [wired] (These named table parts have recorded storage; provider weights are not included as learned memory merely because inference uses them.); Hard/soft checkpoint files and optional output-offload files: known — files [wired] (Storage classification does not assert that optional offload is enabled.); Initiative memory and accumulated prompt/configuration variants and backups: known — files [wired], repo [wired] (SQLite evolution/version metadata is covered with its own metadata unit below; static shipped instructions are excluded.); GEPA version and change metadata: known — sqlite [wired] (Metadata storage is separate from the natural-language instruction payload.); Exposure tracker and retained cross-invocation summary strings: known — in-memory [wired] (Process memory qualifies for these accumulated/reshaped objects; ordinary current-call Ollama/RLM conversation accumulation is excluded.); Training-marker metadata: known — files [wired], sqlite [wired] (Storage is inspected even though downstream consumers are not.); Remaining included maintenance, plugin, training and provider-owned memory alternatives: uninspected (Arbitrary hook/plugin writes, complete optimizer/scoring/evaluation and training consumers, and opaque host compaction remain included but incompletely inspected. Setup/migration/sync and housekeeping routes lack complete recorded admission/consumer coverage. These alternatives cannot inherit the classifications of the inspected routes.)	partial coverage (Known automatic admissions coexist with afforded content-specific human editing. Initial-content and alternate admission control stays unresolved; neither generic caller identity nor workflow launch establishes manual approval.); Episode admission/reflection replacement and Q updates: known — automatic [wired] (Starting a task is not approval of its subsequent reflection or Q changes.); Template/SIMBA lesson admission, score updates and pruning: known — automatic [wired] (Model authorship and physical database I/O are not the admission classification.); Hard/soft checkpoint, session capture, deterministic summary and agent-message writes: known — automatic [wired] (Unknown authorship of an input state file does not erase known automatic capture.); Initiative outcome appends: known — automatic [wired] (Initial plan and separate editing control are not settled by this append mechanism.); GEPA/autoresearch prompt admission and ForgeSmith automatic configuration or enabled auto-patch paths: known — automatic [wired] (ForgeSmith manual-default patch proposals are not silently counted as installed automatic patches.); Interactive human supply/edit of project context and declared plan editing: known — manual [afforded] (Skill requests and declared editing surface establish a capability, not an observed approved entry.); Initial project records, state-file content, trusted role overlays and manual-gated patch installation: not-determinable (Generic writers/trusted commits and recorded proposals do not establish who made a content-specific decision or how a later manual installation occurs.); Optional output-file persistence: known — automatic [afforded] (The persistence capability is inspected but enabling callers and reload remain unestablished.); Imported training-marker row admission: known — automatic [wired] (External training and marker production control remain unresolved.); Remaining included maintenance, plugin, training and provider-owned memory alternatives: uninspected (Arbitrary hook/plugin writes, complete optimizer/scoring/evaluation and training consumers, and opaque host compaction remain included but incompletely inspected. Setup/migration/sync and housekeeping routes lack complete recorded admission/consumer coverage. These alternatives cannot inherit the classifications of the inspected routes.)	partial coverage (Wired automatic supply and afforded request-driven retrieval are supported separately. Unknown reload/consumer/provider alternatives prevent complete direction coverage; the pull witness does not make push selection inapplicable.); Automatic experience, project, initiative, message and recovery delivery: known — push [wired] (Each route is conditional on flags, available state and history/budget gates; these conditions do not remove the implemented push path.); Accumulated role/version and configuration reuse: known — push [wired] (Static shipped prompt bytes are excluded; retained edits and version selection are included.); Explicit MCP memory queries: known — pull [afforded] (Requested return is pull, not a second push. The handler-status disagreement is retained rather than selecting the stronger status.); Output-offload files, training registry, rubric/audit artifacts at later consumers: not-determinable (File pointers and persisted metadata alone do not establish a reload or later-consumer route.); Remaining included maintenance, plugin, training and provider-owned memory alternatives: uninspected (Arbitrary hook/plugin writes, complete optimizer/scoring/evaluation and training consumers, and opaque host compaction remain included but incompletely inspected. Setup/migration/sync and housekeeping routes lack complete recorded admission/consumer coverage. These alternatives cannot inherit the classifications of the inspected routes.)	partial coverage (Identifier matching, coarse availability/ordering/budget filters, lexical inference and optional embedding ranking are positive witnesses. No inferred-judgment value is inferred from the mere presence of a model; missing alternative selectors keep the axis partial.); Task prompt experience selection: role/project/error matching and availability/budget gates: known — identifier [wired], coarse [wired], inferred-lexical [wired] (These are separate selectors in one chain, not three descriptions of an identifier on a catalog entry.); Optional dense and graph episode selectors: partial — inferred-embedding [wired], coarse [wired] (Ordinary telemetry does not forward optional embedding config; alternative vectors/callers remain possible. Mixed integer identity spaces and graph-error fallback prevent a complete coherent-semantic-network claim.); Project history push: known — identifier [wired], coarse [wired] (Resolved questions and older decisions can leave automatic visibility without deletion.); Initiative plan push: known — identifier [wired] (Actual identity matching, not merely a filename returned on explicit request, selects the delivered plan.); Recovery and inter-agent message push: known — identifier [wired], coarse [wired] (Session and hard checkpoint can both enter history; delivery is not guaranteed merely because storage was loaded.); Changed prompt/version supply: known — identifier [wired], coarse [wired] (Performance thresholds/random assignment are coarse selectors, not inferred model judgment of current task semantics.); Request-driven MCP response selection: inapplicable (This unit is pull: the caller requests records, so a push-selection signal is inapplicable here. It does not make the whole axis inapplicable.); Unknown output reload, metadata consumers and other automatic selectors: uninspected (No actual selector/delivery fact classifies these included alternatives; opaque host compaction and plugin selection remain uninspected.)	partial coverage (Several automatic trace-fed durable-artifact paths support yes at wired level. No observed learning benefit or weight training follows. Unresolved writes prevent a system-wide no or a complete all-routes assessment.); Post-task episode/reflection and outcome-priority writes with later retrieval: known — yes [wired] (Qualifying write/consumer wiring, not observed use or improved capacity.); Failure-template and SIMBA lesson/eligibility writes: known — yes [wired] (Unknown semantic novelty and causal usefulness do not negate the qualifying trace-fed write.); Retained continuation snapshots/session reconstruction and summaries: known — yes [wired] (A generated continuation summary qualifies even within a run. Flags, expiry and history gating limit activation, not the existence of the wired route.); Sibling-task initiative outcome memory: known — yes [wired] (Initial operator plan content is a separate unresolved admission branch.); Trace-fed prompt/configuration evolution: known — yes [wired] (Manual-default ForgeSmith patch proposals alone are not an automatic installed update; config and separately enabled automatic paths supply the positive.); Interactive project entries, raw-message-only retention, embedding initialization and offload/metadata routes: not-determinable (Records do not completely establish automatic trace derivation into a durable behavior-shaping result on every one of these alternatives. Raw retention alone and imported training claims do not prove a qualifying write; insufficient bounded absence evidence prevents a local no.); Remaining included maintenance, plugin, training and provider-owned memory alternatives: uninspected (Arbitrary hook/plugin writes, complete optimizer/scoring/evaluation and training consumers, and opaque host compaction remain included but incompletely inspected. Setup/migration/sync and housekeeping routes lack complete recorded admission/consumer coverage. These alternatives cannot inherit the classifications of the inspected routes.)	partial coverage (The inventory distinguishes model-interpreted context/instructions from learning, ranking, routing and executable validation/veto consumers. Missing downstream consumers and remaining plugin/optimizer routes keep system coverage partial.); Historical experience, recovery, project, initiative and message text at later task inference: known — knowledge [wired] (Raw reasons reach this consumer only to the extent extraction and truncation preserve them.); Changed standing role text at later inference: known — instruction [wired] (This is distinct from advisory historical-memory text and from unchanged shipped prompts.); Episodes, task traces and retained failures at update consumers: known — learning [wired] (Authority at the update consumer does not establish truth or improved future capacity.); Episode/lesson prose and access metadata at retrieval ranking consumers: known — ranking [wired] (Fixed scoring code still gives retained text ranking force; graph identity/fallback caveats limit quality, not the existence of the ranking path.); Active/Q/expiry/read-state fields at selection and withdrawal consumers: known — enforcement [wired], routing [wired] (Withdrawal is consumer-specific: inactive lessons remain available to explicit MCP queries.); Existing rules and retained performance metadata at candidate/maintenance admission: known — validation [wired], enforcement [wired] (These criteria concern structural/operational admission rather than factual warrant.); Explicit MCP returned records at requesting client: partial — knowledge [afforded] (Use the capability-level warrant; the retained handler-status conflict does not license a stronger downstream-consumption assertion.); Offloaded output, training marker, rubric details and audit-written LITM weights at later consumers: not-determinable (No complete later-consumer/effect path is established for these retained outputs; do not infer authority from availability or a numeric name.); Remaining included maintenance, plugin, training and provider-owned memory alternatives: uninspected (Arbitrary hook/plugin writes, complete optimizer/scoring/evaluation and training consumers, and opaque host compaction remain included but incompletely inspected. Setup/migration/sync and housekeeping routes lack complete recorded admission/consumer coverage. These alternatives cannot inherit the classifications of the inspected routes.)	AAS-2026-10-15-equipa-32016e8dd9ae-01
EvoOntology	2	Included: accumulated semantic version families and consumer-specific views; project/workload/task/trajectory records; evolution run/hypothesis/evaluation/source-reference parts; checkpoint/reminder metadata; annotations and explorer snapshots; unresolved InsightBench analytical-history/checkpoint and fallback-associated consumption alternatives. Excluded: shipped static skills/prompts/tool schemas, ordinary unretained per-call history and transient caches as storage, external host compaction/provider state, user databases and evaluator-service internals. Runtime views and caches are discussed only where they select retained content.	partial coverage (Known declared persistence is file-backed. External host/provider retention and user data are excluded, not classified as this memory substrate. The included InsightBench recovery/history inventory remains unresolved; positives do not establish complete enumeration.); Five-family semantic versions, persistent task/trajectory files, workload catalog, lifecycle run/round parts, source pointers, checkpoint metadata and version reports/HTML: known — files [wired] (Declared durable objects use workspace JSON, JSONL or HTML files. Semantic links encode a graph but do not establish a graph storage engine; numeric metrics do not establish model-weight storage.); Included InsightBench analytical-history/checkpoint payloads and fallback-associated later-consumer routes: uninspected (The accepted account explicitly retains an untraced recovery/history inventory beyond normalized trajectories. No canonical part/route declarations classify its storage_substrate; available evidence was not independently traced. This prevents complete boundary coverage, not the positive findings on inspected routes.)	partial coverage (Prescribed model admission and implemented derivations are positive automatic routes. Generic administrative and initial supplied-content control remains distinct and unresolved; no explicit manual witness is recorded. The included InsightBench recovery/history inventory remains unresolved; positives do not establish complete enumeration.); Host-model Build/Evolve semantic proposal and patch admission under workflow authorization: known — automatic [claimed] (The prescribed model computation has no mandatory per-content human approval. Starting or budgeting the workflow is not manual admission.); Trace projection, history-to-workload derivation, checkpoint recomputation and generated explorer snapshot: known — automatic [wired] (Code selects/projects/adopts derived content without per-content operator decisions; flags triggering a workflow do not make its internal admission manual.); Generic direct save/selector, configure/prepare supplied-content, annotation, hypothesis/round/evaluation/source-reference calls and original thresholds: not-determinable (Caller identity and per-content approval are unresolved. Physical software I/O alone cannot decide manual versus automatic admission; neither value is asserted for this alternative.); Included InsightBench analytical-history/checkpoint payloads and fallback-associated later-consumer routes: uninspected (The accepted account explicitly retains an untraced recovery/history inventory beyond normalized trajectories. No canonical part/route declarations classify its write_agency; available evidence was not independently traced. This prevents complete boundary coverage, not the positive findings on inspected routes.)	partial coverage (Both requested later consumption and a checkpoint-derived unsolicited reminder are established; actual host injection and meaningful model use remain unobserved. The included InsightBench recovery/history inventory remains unresolved; positives do not establish complete enumeration.); Requested root/plugin semantic resources/tools, shipped benchmark task setup/tools, workload/task reuse, status/resume/source requests and explorer rendering: known — pull [wired] (Automatic selection inside a requested return does not make it unsolicited push. DDR task initialization is recorded as requested setup, not a separate event.); Host reads of persisted prior evidence, explanations and evolution reports during instructed diagnosis: known — pull [afforded] (The skill directs reads and reuse; actual access to referenced bodies is not established by a pointer-return API.); Claude SessionStart checkpoint reminder: known — push [wired] (No consumer request for retained material precedes the hook selector; the delivered material is a short state-derived reason/instruction, not whole trajectories.); Included InsightBench analytical-history/checkpoint payloads and fallback-associated later-consumer routes: uninspected (The accepted account explicitly retains an untraced recovery/history inventory beyond normalized trajectories. No canonical part/route declarations classify its read_back_direction; available evidence was not independently traced. This prevents complete boundary coverage, not the positive findings on inspected routes.)	partial coverage (The established push uses coarse eligibility. Requested semantic lexical/ID selectors do not warrant inferred-lexical or identifier push classifications. The included InsightBench recovery/history inventory remains unresolved; positives do not establish complete enumeration.); Claude SessionStart reminder selection from workspace availability and checkpoint/count/time metadata: known — coarse [wired] (Trigger is SessionStart; selector inputs are active.json presence, trajectory count since checkpoint and elapsed time; selected retained part is workspace/checkpoint eligibility metadata reduced to reason/action context.); Requested root/local semantic tools, task/workload/status/resume/reference and explorer routes: inapplicable (These are pull operations, including requested benchmark setup. Their ID/token/alias matching is not a push signal, and DDR optional task filtering is not invoked by shipped manifest callers.); Included InsightBench analytical-history/checkpoint payloads and fallback-associated later-consumer routes: uninspected (The accepted account explicitly retains an untraced recovery/history inventory beyond normalized trajectories. No canonical part/route declarations classify its read_back_signal; available evidence was not independently traced. This prevents complete boundary coverage, not the positive findings on inspected routes.)	partial coverage (Quote yes as a string. This axis denotes automatic trace-fed durable updates, independently of the runtime learning conclusion about improved capacity. The included InsightBench recovery/history inventory remains unresolved; positives do not establish complete enumeration.); Automatic history-to-workload question/provenance write followed by Builder/start_task consumption: known — yes [wired] (Qualifying durable behavior-shaping material is the catalog, not raw log retention alone. Its later task/build-input use is wired; improved capacity is unobserved.); Model-directed trace diagnosis to candidate semantic content and later semantic guidance: known — yes [claimed] (The prescribed automatic host-model update is distinct from Python storage/gating. Candidate production from experience is instructed, and downstream selected semantic delivery is wired; no actual patch history is supplied.); Generic supplied-content updates and opaque evaluation/artifact inputs: not-determinable (No recorded source-to-automatic-update-to-later-consumer chain resolves these alternatives. Raw retention, thresholds and gate scores alone do not establish a further qualifying trace-learning write, but no bounded absence record licenses no.); Included InsightBench analytical-history/checkpoint payloads and fallback-associated later-consumer routes: uninspected (The accepted account explicitly retains an untraced recovery/history inventory beyond normalized trajectories. No canonical part/route declarations classify its trace_learning; available evidence was not independently traced. This prevents complete boundary coverage, not the positive findings on inspected routes.)	partial coverage (Authority is classified at actual consumers, separating learning at update inputs, knowledge at outputs, ranking and scoped vetoes. The included InsightBench recovery/history inventory remains unresolved; positives do not establish complete enumeration.); Semantic content at host/benchmark analysis models and report/task content at later readers: known — knowledge [wired] (The effect is interpretive context/reference, not an enforced semantic rule or demonstrated compliance.); Retained semantic text and workload features at deterministic relevance/priority selectors: known — ranking [wired], ranking [afforded] (Ranking inputs need not themselves be a retained ranking policy. DDR optional task filtering is only a capability.); Active/lifecycle, task/version/source and checkpoint/run metadata at deterministic selection/continuation consumers: known — routing [wired] (These retained controls choose state, input windows and continuation routes; BIRD lacks the root lifecycle filter.); Stored gate protocol, round/candidate/status and supplied score/verdict/regression inputs at session acceptance: known — validation [wired], enforcement [wired] (Admissibility and enforced refusal coexist at this API. The inputs are not independently authenticated and alternate selector writes bypass the veto.); Retained task status/events and source/purpose/provenance fields at workflow admission guards: known — validation [wired], enforcement [wired] (Guards check event/completion shape and workload admissibility rather than authenticating externally supplied observations.); Task questions and retained traces at the automatic history-to-workload update consumer: known — learning [wired] (Learning authority here means input to a durable behavior-shaping update, not evidence of improved capacity.); Trajectories, hypotheses and rejected-round evidence at the instructed Evolver update consumer: known — learning [claimed], knowledge [claimed] (The skill requires diagnosis and reuse, but actual content-dependent interpretation is unobserved.); Checkpoint-derived reminder at the SessionStart host-context channel: known — instruction [wired] (The emitted reason/action instruction depends on accumulated checkpoint/trajectory metadata; it is not the static skill itself.); Opaque evaluation cases and file bodies behind source/artifact references: not-determinable (Gate fields and projected summaries have classified consumers; unspecified payload bodies do not establish further authority merely by being retained or pointed to.); Included InsightBench analytical-history/checkpoint payloads and fallback-associated later-consumer routes: uninspected (The accepted account explicitly retains an untraced recovery/history inventory beyond normalized trajectories. No canonical part/route declarations classify its behavioral_authority; available evidence was not independently traced. This prevents complete boundary coverage, not the positive findings on inspected routes.)	AAS-2026-10-13-evoontology-f2d0c7992c6e-01
exo	2	Retained material accumulated or revised through use: Basic artifact versions and event/recovery state; agent facts, conversation plans, skill catalog and payload; installed reusable code; mutable managed/local instructions and source changes consumed by later builds; result offloads/previews; web cache; scheduler result/pending delivery state; native replay/compaction, warm/resumed state; sandbox checkpoint registry and opaque checkpoint contents. Included but unresolved: Claude/Cursor, adapter, hosted/UI/evaluator alternatives named in the accepted reports. Excluded: static shipped identity/procedures not changed through use, ordinary turn-local RLM variables, credentials/network settings as knowledge stores, external model weights/provider internals, and standalone Harbor output trajectories without a supported later update consumer. The exporter read of retained journal events remains included.	partial coverage (Files, repository edits and a process cache have local witnesses. Native/backend state and host-specific alternatives prevent complete storage coverage; external model weights are not accumulated Exo memory.); Basic artifact metadata and bytes: facts, todos, skill index/payload, raw result offloads and snapshot registry: known — files [wired] (Version-numbered JSON metadata and bin payloads in the Basic object store establish files, independently of JSON encoding. Other storage backends are not thereby classified.); Native Basic queued TurnWork: known — files [wired] (Basic serializes the retained queue to turn_queue.json for coordinator recovery.); Installed reusable tools, dynamic local profile and guardian outcome files: known — files [wired] (Installed filesystem modules/metadata, reread local profile file and update JSON are separate retained file surfaces. Shipped static identity text is excluded.); Writable Exo source used by guardian rebuild: known — repo [wired] (Retained edits in the source repository enter a later build/module load, conditional on writable grants; no executed edit is established.); Web search result cache: known — in-memory [wired] (Process-local cache retains normalized search results for reuse until TTL/eviction.); Conversation event journal, managed configuration and scheduler retained artifacts across host backends: not-determinable (Persistence and later consumers are wired, but these records do not enumerate backend substrates for every retained part.); Codex warm/resumed/compacted native thread and backend sandbox checkpoint bytes: not-determinable (Native state persistence and backend checkpoint storage are opaque; their symbolic IDs do not establish files, service-object or another substrate.); Claude/Cursor memory alternatives; adapter lifecycle and hosted/UI/evaluator selection outside the inspected chains: uninspected (Included shipped alternatives are not exhaustively inspected. Storage substrates cannot be enumerated from Codex, scheduler and Harbor witnesses; external grading and later trajectory consumption are uninspected.)	partial coverage (Automatic model/software routes are wired. No explicit human per-content decision is recorded strongly enough to assert manual; unresolved caller/edit alternatives prevent absence or complete automatic-only coverage.); Foreground model remember/forget and todowrite admission mechanisms: known — automatic [wired] (Model proposals enter software shape checks without required per-content human approval. Generic tool authorization is not approval of each retained byte.); Foreground model skill and reusable-tool installation/replacement mechanisms: known — automatic [wired] (Software validates and publishes model-proposed contents; initiating a task does not make these admissions manual. Generic nonmodel caller control remains a separate unresolved alternative.); Result normalization/offload, event journaling, scheduler result persistence and native replay compaction: known — automatic [wired] (Software writes/derives retained state without an explicit operator content decision. Opaque native transformation details do not erase known host automation.); Web cache insertion and expiry: known — automatic [wired] (HTTP/search handling admits normalized cached results through software rules.); Guardian deployment outcome admission: known — automatic [wired] (Deferred script records pipeline status without per-outcome human approval; this does not settle admission of the source edits being built.); Nonmodel generic API/tool callers; definition changes; local profile/source editing; initial skill/fact content and checkpoint replacement control: not-determinable (Generic caller identity and physical file writes do not establish a human operator decision. Capture/rewind requests and reads do not independently prove manual checkpoint replacement. These alternative mechanisms remain unresolved.); Claude/Cursor memory alternatives; adapter lifecycle and hosted/UI/evaluator selection outside the inspected chains: uninspected (Included shipped alternatives are not exhaustively inspected. Admission control cannot be enumerated from Codex, scheduler and Harbor witnesses; external grading and later trajectory consumption are uninspected.)	partial coverage (Both requested pull and independently supplied push have local witnesses. Within a chain the requested result return is not counted as an additional push. Full-result reload and alternate hosts remain unresolved.); Instruction-time facts, open-plan list and skill catalog: known — push [wired] (Assembly automatically supplies retained entries without a foreground request for those contents. Exact-path lookup inside that assembly remains part of push, not another consumer-requested route.); Later conversation materialization, Codex replay/reuse, registry loading and dynamic guidance: known — push [wired] (Harness/runtime automatically brings selected retained history, instructions or capability state to a later invocation. No parity is inferred for every native harness.); Scheduler owner-selected wakeup: known — push [wired] (Pending result delivery automatically supplies formatted retained output to its owner conversation without that consumer requesting retrieval.); Foreground skill/body/file reads and checkpoint rewind: known — pull [wired] (Consumer chooses name/path/snapshot and requests retained material. Checkpoint read-back reaches executable filesystem consumers; introspection is a separate afforded described tool route below.); History introspection description contract: known — pull [afforded] (Registered model-facing contract affords requested retained telemetry; host-side filtering and actual returned records were not independently audited.); Harbor exporter requests recorded event page: known — pull [wired] (Later CLI export invocation requests up to 10,000 retained events by type; returning that page fulfills pull. External later use of the resulting trajectory is not established.); Search cache hit satisfying model query: known — pull [wired] (Reuse returns retained normalized results in fulfillment of an explicit query, not independent unsolicited memory supply.); General full-result payload reload and opaque backend/native alternatives: not-determinable (Artifact APIs establish storage access, not a universal model-facing full-result consumer. Native continuation is wired but additional internal requested or automatic routes are opaque.); Claude/Cursor memory alternatives; adapter lifecycle and hosted/UI/evaluator selection outside the inspected chains: uninspected (Included shipped alternatives are not exhaustively inspected. Read-back alternatives cannot be enumerated from Codex, scheduler and Harbor witnesses; external grading and later trajectory consumption are uninspected.)	partial coverage (Coarse and owner-identifier push selection are supported locally. Pull-only operations need no push signal, but their existence cannot make this partial inventory inapplicable. No inspected semantic/embedding/lexical selector is established.); Fact injection and skill catalog availability: known — coarse [wired] (Trigger: instruction assembly. Selector: agent and latest exact-path artifact, then nonempty/readable contents. Selected parts: all current facts or indexed names/descriptions. This is coarse availability, not topical identity targeting.); Plan injection: known — coarse [wired] (Trigger: assembly. Selector: current conversation/latest list plus any pending/in_progress status. Selected part: whole list including closed items when one remains open. Status filtering is coarse.); Conversation history and retained normalized previews: known — coarse [wired] (Trigger: next model request. Selector: conversation scope/cursor and code conversion/length budgets. Selected parts: materialized prior messages/results or their fixed prefixes. These are coarse scope/budget rules, not inferred relevance.); Native replay selection and interleaved compaction: known — coarse [wired] (Trigger: failure to reuse a native thread. Selector: replay-eligible history with role exclusions and 256,000-character chunk rule. Selected retained parts: call/output-preserving chunks. This classifies bridge coarse selection, not opaque internal compaction.); Scheduler delivery to owner conversation: known — identifier [wired] (Trigger: pending fire. Selector input: stored owner conversation/task and pending-delivery identity. Match selects that retained result/prompt for the actual destination. This is targeted identifier-based push, not merely an address on requested data.); Registry load and dynamic local/managed instruction supply: known — coarse [wired] (Trigger: later registry/request assembly. Selector: installed/enabled entries and configured source/path; selected parts: available definitions/modules and standing guidance. Catalog/load availability does not establish semantic targeting.); Native compacted/warm/resumed internal payload selection: not-determinable (Host selection is inspected but the native selector and selected payload are not. No inferred-judgment or embedding value follows from a model/native compaction label.); Remaining backend checkpoint and source-change internal push consumers: not-determinable (Requested rewind has no independent push-signal finding; subsequent restored-state/module consumers are insufficiently decomposed to enumerate all selectors. An upstream requested ID alone is not identifier push.); Claude/Cursor memory alternatives; adapter lifecycle and hosted/UI/evaluator selection outside the inspected chains: uninspected (Included shipped alternatives are not exhaustively inspected. Push selectors cannot be enumerated from Codex, scheduler and Harbor witnesses; external grading and later trajectory consumption are uninspected.)	partial coverage (Three continuation transformations support local yes at wired strength. Raw logs, exported trajectories and sequential trials alone are not added learning witnesses. No bounded absence record supports a local or system no; unresolved proposal/native routes remain visible.); Automatic result envelope/preview write and later model materialization: known — yes [wired] (Recorded tool output feeds a durable reduced context artifact/envelope and later model requests consume it. This qualifies even though reduction is deterministic, no new claim is generated and improvement is unobserved.); Codex automatic replay compaction and subsequent native turn: known — yes [wired] (Retained session/tool history feeds automatic native-state transformation and later turn context. The boundary establishes a durable continuation state route; opaque representation limits form/curation, not the wired existence of this trace-fed update.); Scheduler result artifact/pending fire to formatted retained wakeup and later agent turn: known — yes [wired] (Command execution feedback is automatically retained and formatted into subsequent model context. This is a trace-fed behavior-shaping continuation route, not an admitted reusable lesson or evidence of improvement.); Foreground remember, plans, skills, reusable-code/source proposals and snapshot contents: not-determinable (Software/model admission is known, but original trace inputs and a trace-to-durable-update relation are not established for every proposal. Experience may inform proposals; generic writes alone establish neither yes nor a bounded no.); Native warm/resume states apart from inspected replay compaction: not-determinable (Retention/reuse is known; additional automatic derivation and input details are opaque. The replay witness cannot classify all alternative checkpoints.); Claude/Cursor memory alternatives; adapter lifecycle and hosted/UI/evaluator selection outside the inspected chains: uninspected (Included shipped alternatives are not exhaustively inspected. Qualifying trace-fed writes cannot be enumerated from Codex, scheduler and Harbor witnesses; external grading and later trajectory consumption are uninspected.)	partial coverage (Authority is split by consumer. Instructional reliance and informational context coexist with routing, local validation/veto and trace-update learning. No complete authority union or empirical effect follows from these wired paths.); Saved facts and plans delivered as standing developer guidance: known — instruction [wired] (The later model is directed to rely on saved facts and maintain the retained plan. This is model-interpreted instruction, not truth acceptance or a hard policy gate.); Skill body/procedure at the model and catalog descriptions requesting relevant body use: known — instruction [wired] (Catalog guidance requests retrieval and the selected procedure requests compliance. Actual model compliance is unobserved.); Skill index membership at use_skill: known — enforcement [wired] (Membership permits/refuses this body-read route. Direct read_skill_file bypasses that check, so enforcement is local, not complete revocation.); Skill index names/paths at body/file lookup and artifact identities at version selection: known — routing [wired] (Retained access metadata routes fixed lookups to the requested payload/latest version; this is not evidence that catalog descriptions are automatically ranked.); Conversation history, tool-result previews, web text and scheduler outputs at later model requests: known — knowledge [wired] (Retained material supplies context or evidence to subsequent answering/action decisions; persistence confers no factual warrant.); Retained tool-result traces at normalization into durable continuation context: known — learning [wired] (The trace-fed update produces a retained behavior-shaping result envelope/preview for later model context. Learning denotes this update consumer, not demonstrated capacity improvement.); Retained scheduler execution feedback at continuation-formatting consumer: known — learning [wired] (Recorded command output/exit feedback feeds a durable formatted wakeup for the later agent turn. Learning describes the trace-fed update force, not a correctness check or demonstrated improvement.); Retained conversational replay at Codex compaction/update consumer: known — learning [wired] (Recorded messages and call/output units feed automatically updated native continuation state consumed by a subsequent turn. Native algorithm and semantic payload remain opaque.); Retained queue/recovery state, scheduler pending/owner state and native resume metadata at progression selectors: known — routing [wired] (State chooses next work, redelivery destination or recovery/replay branch; it does not establish an answer oracle.); Saved Basic progress/pending-approval state at unresolved-call resumption gate: known — validation [wired], enforcement [wired] (Runtime checks the retained call/approval correspondence before resumption and can refuse unsafe recovery. The retained match supplies a criterion and a veto within Basic only.); Installed registry identity/module references at subsequent loader: known — routing [wired] (Retained registry entries select modules/initialization and expose the corresponding capabilities. Executable effects are not themselves all classified as permission enforcement.); Managed/local instruction text at later harness/model: known — instruction [wired] (Saved definition and dynamic local profile supply standing directions; selected harness/model settings separately route execution.); Retained definition/thread settings and sandbox checkpoint registry at executor/restoration selection: known — routing [wired] (Settings select executor/resources; snapshot IDs select restored state for later commands. This does not classify opaque restored file contents.); Opaque native/snapshot contents, arbitrary tool code effects and guardian changed-source semantics: not-determinable (Executable behavior and restored bytes can have several forces, but these records do not establish every actual retained-content/consumer effect. No ranking or blanket enforcement value is inferred from executable authority.); Claude/Cursor memory alternatives; adapter lifecycle and hosted/UI/evaluator selection outside the inspected chains: uninspected (Included shipped alternatives are not exhaustively inspected. Consumer/effect inventory cannot be enumerated from Codex, scheduler and Harbor witnesses; external grading and later trajectory consumption are uninspected.)	AAS-2026-10-15-exo-f0905e2cf44b-01
Exocomp	2	Included: accumulated planner/contractor message records and returned Inquire summaries; requirement Behavior and source-derived declaration/status views; bug descriptions/status and changelog descriptions/indexing metadata; project-local role templates/grants and skill bodies/dependencies/scripts with catalog, loaded-state and separate body-consumption routes; diagnostic snapshots; page/answer content retained in messages; recorded restore and operator-display routes. Unresolved included alternatives: initial project-content provenance/control, provider-private retention, arbitrary program-created memory, cross-restart loaded-skill bookkeeping and pending-wait reload. Excluded: static shipped role instructions/grants as standalone memory, unchanged external model weights, production binaries and arbitrary product files except source supplying requirement reconstruction, ordinary live-process/HTTP state with no later retained-memory consumer, external dependency implementations and generated frontend deployment behavior.	partial coverage (The records inventory filesystem/in-process retained surfaces and their readers. Provider-private retention, arbitrary program stores and restored extension bookkeeping prevent complete substrate coverage.); Planner and collected contractor messages: known — files [wired], in-memory [wired] (Agent arrays and recovery JSON retain the recorded messages.); Inquire retained tool-result summary: known — files [afforded], in-memory [afforded] (If invoked through Session, the returned summary enters messages and can survive backup.); Requirement behavior and declaration/status view: known — files [wired], in-memory [wired] (Pending JSON and source files back the freshly merged map; implemented true status is regenerated rather than archived.); Bug description, identity/fixed status and changelog entries: known — files [wired], in-memory [wired] (Each report part is stored in project JSON and a tool map.); Project-local role templates: known — files [wired], in-memory [wired] (Project YAML is reread at construction and registered in the process map.); Skill bodies/dependencies/scripts, catalogs and loaded objects: known — files [afforded], in-memory [afforded] (Grant-dependent skill capacity has separate file/catalog/loaded/body parts; no default activation is inferred.); Diagnostic request/response snapshots: known — files [wired] (The diagnostic boundary is filesystem JSON; no automatic inference consumer is established.); Cross-restart loaded-skill bookkeeping: not-determinable (Message-body recovery does not establish persistence or reload of loaded_skills; its cross-restart substrate is not determinable.); Provider-private retention and arbitrary program-created memory: uninspected (Included opaque alternatives: no recorded inspection establishes their storage substrate. Inference and host execution are not evidence of their internal retained-memory mechanisms.)	partial coverage (Software admission is distinct from who authored content and from operator startup. No explicit per-content human decision is recorded for an additional manual finding; initial/external caller-controlled writes remain unresolved.); Turn/child-output recording, backup and recovery-content admission: known — automatic [wired] (Runtime records/imports content and parsing admits restored JSON without human per-turn approval.); Requirement source refresh, Signoff status and pending writer: known — automatic [wired] (Source-first reconstruction and filtering are software decisions, not human signoff.); Direct-method requirement definition admission: known — automatic [afforded] (Definition methods parse/admit supplied declaration/behavior without per-content human veto; generic caller identity does not prove manual content supply.); Model-proposed bug/changelog report maintenance: known — automatic [wired] (The model/tool route admits reports after software checks rather than human approval.); Explicit human supply/edit/approval on direct report API and initial project file writes: not-determinable (Generic API caller identity and arbitrary file authorship do not identify per-content human control. No manual classification follows from starting the workflow.); Role discovery registration: known — automatic [wired] (Software parse/format/role checks admit a project file into the template map; upstream authorship is separate.); Skill catalog/loaded-object and helper-body admission: known — automatic [afforded] (Catalog import, dependency gating and helper text checks admit material without human body approval; helper and qualified entry points remain separate.); Inquire derived summary admission: known — automatic [afforded] (Caller chooses a child but program/model produce and parse the summary without human content approval.); Debug snapshot replacement: known — automatic [wired] (Raw inference I/O is automatically marshaled at fixed filenames.); Provider-private retention and arbitrary program-created memory: uninspected (Included opaque alternatives: no recorded inspection establishes write/admission control. Inference and host execution are not evidence of their internal retained-memory mechanisms.)	partial coverage (Requested delivery and automatic supply are separate operations even within one chain. Known pull routes cannot resolve provider/program consumers or cross-restart extension-state uncertainty.); Restored planner checkpoint selection and subsequent full-message supply: known — push [wired] (Startup/inference automatically supplies the selected backup without a request for individual retained turns.); Await report, project List/Search and operator GET: known — pull [wired] (Consumer calls request retained material; return is fulfillment, not a second push.); Explicit Skills.List catalog request: known — pull [afforded] (An explicit request refreshes and returns the catalog; grant-dependent capacity is not default activation.); Requested Inquire summary return: known — pull [afforded] (Caller explicitly requests a summary of a selected recorded agent; its return is not unsolicited proactive retrieval.); Project role file discovery and later same-process template lookup: known — push [wired] (Construction automatically imports matching local files and later lookup uses the registered template; this does not establish fresh-child execution.); Requested skill selection and helper instruction delivery: known — pull [afforded], push [afforded] (Load/helper requests select a body/object; later inference automatically resupplies already inserted system bodies as part of accumulated messages.); Diagnostic snapshots at a later reader: not-determinable (A storage capability and hypothetical reader do not establish a read-back route.); Restored loaded-skill objects and pending human-wait state: not-determinable (Recorded message recovery does not establish these coordination/tool-state reload routes.); Provider-private retention and arbitrary program-created memory: uninspected (Included opaque alternatives: no recorded inspection establishes later read-back direction. Inference and host execution are not evidence of their internal retained-memory mechanisms.)	partial coverage (Signals classify automatic push selection only. Fixed checkpoint/wholesale message supply is coarse; role replacement uses an actual role match. Requested identifiers/searches and display filters are not inferred proactive selectors.); Startup checkpoint and later accumulated-message supply: known — coarse [wired] (Trigger: terminal/web startup and later inference. Input: backup existence/parse success and full Agent.Messages. Selected part: the project checkpoint and all accumulated planner messages; no relevance or token ranking.); Project-local role replacement: known — identifier [wired] (Trigger: Agents construction. Input: parsed valid YAML Role and matching template key. Selected part: retained project role template admitted as that role replacement.); Later wholesale supply of separately injected skill system bodies: known — coarse [afforded] (Trigger: later inference. Input: already accumulated Agent.Messages, including separately inserted body. Selected part: whole message content, not a newly identity-matched skill recommendation.); Requested Await/Inquire, project List/Search, operator GET and explicit skill catalog/Load requests: inapplicable (These local operations fulfill consumer requests. Their agent/file/symbol/name and lexical display inputs are pull selectors, so no push signal applies to these operations.); Unestablished diagnostic, restored coordination/extension and opaque-provider/program consumer routes: uninspected (Missing later-consumer/selection mechanisms prevent determining whether any additional push exists and what it selects. Known requested routes cannot make these alternatives inapplicable.)	partial coverage (The afforded Inquire chain qualifies by automatic trace-fed summary retention and later context consumption, not by measured improvement. Raw logs and source-fed refresh are separate; opaque updates and arbitrary authored stores prevent a system-wide negative or complete enumeration.); Inquire child-session summary retained through Session tool result/backup and later caller inference: known — yes [afforded] (Recorded prose feeds automatic condensation, retained tool-message output supplies later context, and raw child transcript remains separately available. No embedded grant or observed invocation is established.); Raw turn/child transcript backup and diagnostic request/response snapshot copying: inapplicable (These particular writes retain raw activity without deriving a durable behavior-shaping artifact or learned parameters. No qualifying trace-learning transformation is established by raw retention alone.); Source-derived requirement reconstruction: inapplicable (Input is project source syntax, not activity/experience traces; automatic status/declaration refresh is not a qualifying trace-fed write.); Initial or external authoring of project reports, roles and skills; arbitrary program/provider updates: uninspected (Reports and instruction files can be authored by external humans/programs/models, but the records do not trace their original inputs and transformations into a qualifying retained update. Provider-private changes and arbitrary stores remain uninspected.)	partial coverage (Authorities attach to consumed parts and effects. Known local routes do not establish compliant models, correct reports, successful persistence or forces inside opaque stores.); Recovered conversation and requested project reports/page/answer content at later model or operator: known — knowledge [wired] (Content is reference/context/advice; delivered answers do not certify truth or approval.); Surviving requirement Behavior at model/UI readers: known — instruction [wired], knowledge [wired] (Behavior describes requested policy, while declarations/status describe source structure; ordinary new Define* admission is not wired.); Recovered system-role content and local role prompt at same-process template/model consumer: known — instruction [wired] (Recovered system content retains instruction priority; local registration reaches later same-process lookup, not established fresh-child execution.); Skill body at separate helper/bare-name insertion consumer: known — instruction [afforded] (Only this alternative supplies the body at system priority; qualified Load delivers a generic notice.); Child prose at Inquire update and returned summary at caller: known — learning [afforded], knowledge [afforded] (Automatic retained summary production uses recorded experience; this authority is distinct from demonstrated improved capacity.); Bug fixed-status at List/Search filter: known — enforcement [wired] (The flag excludes descriptions from model tool lists but not whole-map operator JSON.); Requirement IsImplemented at pending writer and source-first file/symbol keys at merge: known — enforcement [wired], routing [wired] (Consumed status and keys enforce omission/precedence; they do not validate required behavior.); Changelog retained description/type/file/symbol at duplicate admission and Date at List ordering: known — validation [wired], enforcement [wired], ranking [wired] (Duplicate metadata is an admission criterion; retained dates determine relative display order.); Admitted local-role grants at toolset and HasMethod consumer: known — enforcement [wired] (Only later same-process Agent/tool construction is established for registered templates; fresh-child role execution is not inferred.); Loaded skill dependency/script state at Load and Execute consumers: known — validation [afforded], enforcement [afforded], routing [afforded] (Dependencies admit loaded-state changes; loaded-name/script/runtime checks govern execution independently of body injection.); Diagnostic snapshots at any future automatic model consumer: not-determinable (Storage is recorded but no automatic next-inference consumer is established; therefore no authority at such a consumer is determinable.); Provider-private retention and arbitrary program-created memory: uninspected (Included opaque alternatives: no recorded inspection establishes later consumers and behavioral effects. Inference and host execution are not evidence of their internal retained-memory mechanisms.)	AAS-2026-10-14-exocomp-692bb27c35f9-01
ExpeL	2	Use-derived trial histories, retry reflections and their archives, rule text and fold caches, rule counters, persisted trial/critique logs and resume state, rebuilt retrieval documents/vectors and their access metadata, and auxiliary critique dictionaries with unresolved later consumers. Includes training, extraction, retry, evaluation, checkpoint replacement and reporting routes. Excludes shipped static examples and method prompts as memory objects, ordinary unretained task frames, external model weights and provider-internal retention; includes the access structures derived from retained experience, not the embedding model itself.	known coverage (The recorded object inventory covers scoped memory storage, including conditional auxiliary objects; external weights are not accumulated memory here.); Trial histories, reflection copies, rules/caches, counters, logs and resume state: known — in-memory [wired], files [wired] (Records cover the operative retained containers and their persistence; no checkpoint instance or successful write was observed.); Rebuilt retrieval documents, mappings and FAISS access store: known — in-memory [wired], vector [wired] (Numerical encoding and vector substrate are separate classifications.); Auxiliary critique dictionaries and optional dictionary snapshots: known — in-memory [afforded], files [afforded] (Known substrate classification does not imply a functioning shipped auxiliary generation loop.)	partial coverage (Automatic admission is supported locally; replacement provenance and alternate caller control prevent a complete agency inventory.); Ordinary non-testing trial admission, reflection retention, rule updates, cache copies and snapshots: known — automatic [wired] (Computational outcome classification and syntax/counter gates govern ordinary writes.); Retrieval-access reconstruction: known — automatic [wired] (No per-document approval is recorded.); Checkpoint restoration: partial — automatic [wired] (Upstream supplying/editing/approving replacement content is unresolved: a generic caller-selected file does not establish manual admission.); Testing response supply and auxiliary caller content control: partial — automatic [afforded] (Testing input changes content supply to the caller; exact principal and explicit human decision on retained content are not established for every alternate, so manual is not asserted.)	partial coverage (Both requested loading and automatic model supply are supported; unresolved included alternatives prevent complete read-back coverage.); Requested checkpoint/log loading by drivers or reporting callers: known — pull [wired] (Explicit loading requests precede separate downstream automatic context selection.); Retry reflections, extraction evidence/existing rules, evaluation rules and demonstrations: known — push [wired] (Starting extraction/evaluation supplies workflow inputs, not a model memory query; selection and model delivery are separate from requested file returns.); Auxiliary critique and direct reflection-archive consumption alternatives: not-determinable (No shipped auxiliary decision reader or direct past_reflections reader is established; reflection_similarity lacks final dispatch, so its potential direction is not classified.)	partial coverage (Signals characterize automatic selection within push chains; requested file/log names are not counted as identifier push signals. Unresolved consumers keep this axis partial.); Retry reflection and full current-rule supply: known — coarse [wired] (Failure/retry and evaluation prompt construction are triggers; available reflection/current-rule state is the selector input.); Extraction task lookup and evaluation fold-cache rule selection: known — identifier [wired], coarse [wired] (Triggers are create_rules or fold rule load. Identifier matching is an actual retained-part lookup, not merely an identifier on a requested file; success-group task-identity holdout remains conditional.); Automatic evaluation demonstration refresh: known — inferred-embedding [wired], identifier [wired], coarse [wired] (Selected retained parts are successful-history demonstrations via document keys, mixed with shipped examples; unsupported reflection_similarity is excluded from supported selectors, not from the uncertainty inventory.); Auxiliary critique and direct reflection-archive read-back: not-determinable (An actual later push and selector are unestablished, preventing classification rather than implying pull-only inapplicability.)	partial coverage (Two positive wired routes coexist with unresolved alternatives. No bounded absence record supports a local or system-wide no; raw trajectory retrieval alone is not counted as a derived learning write.); Ordinary failed-trial reflection write followed by retry acting: known — yes [wired] (The retained list/trajectory/checkpoint chain qualifies without parameter updates or observed improvement.); Ordinary trial-conditioned rule maintenance followed by critique/evaluation: known — yes [wired] (Wired behavior-shaping artifact production is not successful capacity improvement or epistemic acceptance.); Auxiliary critiques, replacement checkpoint provenance and testing content alternatives: not-determinable (Auxiliary durable later behavior-shaping use is unestablished; external replacement generation and testing supply control are unknown. Raw retention or restoration alone does not settle a qualifying automatic trace-fed write.)	partial coverage (Inspected consumers establish multiple local forces; unresolved archive/auxiliary consumers prevent complete authority coverage. Static parser policy is excluded as a memory object.); Retained trials and current rule text consumed by the extraction critic: known — learning [wired] (Authority at the update consumer is separate from the output's acting authority and from demonstrated improved capacity.); Retained reflection and rules supplied to retry/evaluation acting: known — instruction [wired], knowledge [wired] (Instruction and advisory consumption coexist across the recorded templates; FEVER's MANDATORY text remains model-interpreted, not an action veto.); Retained demonstrations at retrieval selection and subsequent acting: known — ranking [wired], knowledge [wired], routing [wired] (Selected key and displayed trajectory can differ; ranking does not establish relevance or truth.); Rule counters at the updater: known — ranking [wired], validation [wired], enforcement [wired] (The same criterion supplies validation and an enforced membership veto; numeric counters are not delivered to the actor.); Resume state and persisted logs at restoration/extraction/reporting consumers: known — routing [wired], knowledge [wired] (A report's textual predicates do not independently authenticate benchmark success.); Auxiliary critique payloads and archived reflections without established direct consumers: not-determinable (No shipped later decision reader is established for auxiliary dictionaries or past_reflections; storage and display alone cannot assign force, and unsupported reflection retrieval is not silently counted.)	AAS-2026-10-14-expel-efd908f05069-01
Fractal	2	Fractal-retained summary, bounded detailed history including opaque RunTrace, usage parts, their checkpoint admission/recovery and later model/UI routes; changed root workspace guidance and its reload; operator-changed provider/default configuration and later resolution; accumulated release-notice cache and later fallback. Includes unresolved caller-supplied summary/history alternatives at the direct Python API as afforded consumption, not invented persistence. Excludes shipped static instructions, arbitrary application work products without an established later memory consumer, external provider weights/state, SBX contents, PredictRLM internal compaction/recursive frames and release artifacts.	partial coverage (Session parts, root guidance, configuration and version cache have inspected file retention and live host objects. The direct API permits caller-supplied context without establishing its retention substrate.); Summary, detailed history including opaque RunTrace, and usage checkpoint parts: known — files [wired], in-memory [wired] (All three parts persist in session JSON files and live in Pydantic objects. Opaque trace semantics do not make its visible checkpoint substrate unknown.); Root AGENTS.md retained guidance: known — files [wired] (The loader rereads a mutable workspace file; no repository-backed retention requirement is established.); Changed provider/default settings: known — files [wired] (TOML files retain settings for later launch resolution; live selected adapters are not a trained-memory store.); Version notice cache: known — files [wired] (A config-directory JSON file retains latest, fetched_at and install_id for later startup.); Direct API supplied summary/history retention: not-determinable (The adapter accepts supplied context but does not persist it. Caller retention and initial substrate are unspecified; this afforded alternative cannot be assigned the checkpoint substrate.)	partial coverage (Checkpoint, recovery and cache admissions are software decisions. Configuration explicitly takes operator selections. Human content control is unresolved for generic input and open workspace/direct-API alternatives.); Pending/final checkpoint admission, counts/status/usage projection: known — automatic [wired] (Software admits caller/model/event contents and finalizes matching turns without a per-content human approval; launching a task does not make these admissions manual.); Session load rejection and empty-state replacement: known — automatic [wired] (Schema/JSON/ID checks and recovery code admit or replace state automatically; requesting an ID is selection, not content approval.); Provider/default settings supplied by operator command or setup: known — manual [wired] (The operator explicitly chooses settings to save or replace. Software validity checks do not remove this per-setting operator decision.); Website assertion and cache metadata admission: known — automatic [wired] (Network/shape/fallback code writes cache contents without operator content approval.); Model editing of retained root guidance: partial — automatic [afforded] (The mutable-workspace route affords a model edit without per-edit veto, but no dedicated guidance write or instance is established, and external editing control remains unknown.); Generic caller message origination and direct-API context initialization: not-determinable (Caller identity does not prove a human per-content decision. Known automatic checkpoint admission does not resolve how caller-supplied material was initialized.)	partial coverage (Explicit session and accounting requests are pull; subsequent automatic context, guidance, configuration and cache supply are push. A separate later consumer for generic direct-API input remains unresolved.); Requested session-ID reload, session listing and usage access: known — pull [wired] (The caller requests retained state or accounting; fulfilling that request is pull, not another push because code returns it.); Subsequent turn summary/history handoff: known — push [wired] (Submit automatically supplies all retained summary turns and the recent detailed tail without a separate request for these memory parts.); Foreground exact-detail lookup within supplied history: known — pull [afforded] (The delegated model is directed to inspect retained history from Python when it wants details. That requested inspection is an afforded pull, separate from host push delivery.); Root guidance reread on agent invocation: known — push [wired] (The loader supplies retained root text on each call without a separate consumer retrieval request.); Configuration resolution and cached release fallback on later launches: known — push [wired] (Host launch/startup automatically consumes retained settings and eligible cached latest information.); Direct API supplied context across consumer invocations: not-determinable (The adapter offers caller input consumption, not established persistence or a later memory route; whether a particular later caller requests or receives retained parts is unspecified.)	partial coverage (Inspected push selectors use availability, recency, fixed root location, configuration precedence and TTL. Requested session IDs are not identifier-based push. Direct-API later selection is unresolved.); Automatic summary/history push on submit: known — coarse [wired] (Trigger: later submit. Input: selected session state. Parts: every summary turn plus the latest 20 history entries, including pending turn. Coarse session/recency selection is not semantic similarity or identity-matched subset delivery.); Root guidance push on agent call: known — coarse [wired] (Trigger: invocation. Input: presence/readability of root AGENTS.md. Part: whole text or its 20,000-character prefix. A fixed filename alone is not targeted identity-matched push.); Configuration and version-cache host push: known — coarse [wired] (Triggers: launch resolution/startup fallback. Inputs: fixed configuration layers and cache availability/age. Parts: applicable settings or cached latest identifier. Precedence and TTL are coarse eligibility rules.); Requested session reload/list/usage and foreground detail inspection: inapplicable (These are pull operations; their request IDs or model lookup choices do not classify a push signal. Other included push operations remain separately assessed.); Direct API caller selection for later supplied memory: not-determinable (No independently established later consumer/selector is recorded; cannot classify its signal or declare this alternative pull-only.)	partial coverage (Automatic derived status/accounting/context reaches a later model, supporting a narrow yes. This does not establish improved capacity. Open guidance edits and generic API alternatives cannot supply a system-wide no.); Trace-fed derived checkpoint context: known — yes [wired] (Hook observations and trace status/accounting automatically produce durable summary context for a later model invocation. This qualifies even though responses are not semantically summarized and no general lesson algorithm exists.); Raw detailed-history retention alone: inapplicable (Saving raw messages/traces without a separate behavior-shaping derivation is not itself a qualifying learning write. Its role as input/context remains distinct from the derived summary route.); Usage-only host accounting output: inapplicable (Standalone numeric usage display/aggregation is not a behavior-shaping model artifact or learned parameter update; context counts/status are assessed separately.); Operator configuration and website-version cache writes: inapplicable (These routes admit explicit selections or website assertions, not automatic trace-fed behavioral artifacts. This is route inapplicability, not a system-wide negative.); Potential trace-responsive root guidance editing: not-determinable (Model task editing can change later guidance, but no accepted record ties its proposal to qualifying trace inputs. A reread path does not establish or exclude trace-fed learning.); Direct API supplied context derivation: not-determinable (Initialization control, trace inputs, durable transformation and later caller are unspecified; the known checkpoint yes does not resolve this alternative.)	partial coverage (Authority is assigned at actual later consumers, not from container names or prompt placement. Opaque trace semantics and direct-API initialization leave possible additional forces unresolved.); Summary at later PredictRLM context consumer: known — knowledge [wired] (Prior requests/results/status/counts supply trajectory context and advice for subsequent work. Placement in an instruction docstring alone does not turn historical responses into standing directions or prove reliance.); Detailed history at afforded model Python recall consumer: known — knowledge [afforded] (The supplied tail offers exact-recall reference material. Fractal wires delivery; actual inspection is only afforded and is not observed.); Additional authority of opaque RunTrace internals: not-determinable (No accepted record identifies all interpreted trace parts or their exact dependency consumers; blanket instruction, learning or enforcement authority cannot be assigned.); Usage and retained conversation at later UI/accounting readers: known — knowledge [wired] (Retained fields provide information for conversation display and cumulative usage, not an action-approval barrier.); Retained root guidance at signature consumer: known — instruction [wired] (The model is asked to follow root guidance subject to the current request; no filesystem enforcement follows from that instruction policy.); Retained provider/default configuration at later launch resolution: known — routing [wired] (Settings select adapters and execution defaults. Routing is not parameter learning or task-answer validation.); Cached release identifier at later notice consumer: known — knowledge [wired] (The cached assertion supports an advisory upgrade notice, not automatic installation or a quality verdict.); Direct API supplied context at unspecified later consumer: not-determinable (Construction affords consumption but does not establish a separately wired later caller, content semantics or all authority paths.)	AAS-2026-10-14-fractal-ee03010af4c2-01
funes	2	Retained library originals, searchable editions, source notes, concepts, topics, registry, wiki index, saved outputs, dated health reports and changelog; their admission, editorial maintenance and later consumers. Includes setup-created local scope/deviation instructions and their future loading, but excludes unchanged shipped protocol/scaffold text as memory. Standalone helper stdout, transient extraction state, temporary images and current CI status are not retained library memory. Saved helper content is included when retained for library use. External host/provider internals, linked example libraries and unacquired user libraries are excluded, not classified as absent. Optional Git versioning is distinguished from files.	known coverage (The declared tree and its parts cover the scoped retained substrates. The external model/provider is not an opaque library storage alternative.); Library originals, editions, wiki prose, navigation, outputs, reports, changelog and setup-created local instructions on disk: known — files [claimed] (The container and separately described parts cover the included file inventory, including binary originals.); PDF and saved-caption editions produced by the executable helpers: known — files [wired] (Includes partial PDF output, alternate caption formats and caller-selected paths when admitted to the library.); Committed versions of the retained library: known — repo [afforded] (Optional versioned retention is distinct from precommit filesystem retention.)	partial coverage (Both human-controlled and computational admissions are supported locally; ambiguous offered/proposed changes retain uncertainty instead of inheriting routine edit agency.); Human supply of raw source content and explicit approval of cross-library links: known — manual [claimed] (Human approval is an instruction rule, not demonstrated interception of arbitrary writes.); Routine Librarian admission of raw persistence, transcriptions, source notes, concept revisions, topics, registry/index edits, outputs, reports and changelog: known — automatic [claimed] (Safe fixes belong here; proposed non-safe fixes and offered filing remain distinct decision branches.); PDF extraction and caption serialization/file admission: known — automatic [wired] (Unknown identity of a generic caller does not obscure the subsequent software admission mechanism.); Confirmed setup conventions and agent-authored local instructions/scaffold: known — manual [claimed], automatic [claimed] (Choice of conventions and admission of generated file contents are separate stages of this setup chain.); Durable output filing and non-safe editorial changes proposed to the human: not-determinable (The offer/proposal establishes a decision opportunity, but the records do not resolve final per-content admission across the stronger output README wording and shared protocol. No unconditional automatic installation or deployed human veto is established.)	partial coverage (Supported pull routes do not resolve the included local-guidance delivery and unspecified later report/output routes; the scoped boundary cannot be declared completely pull-only.); Question-driven wiki retrieval, compilation/source verification and requested editorial inspection: known — pull [claimed] (On-demand index/source/concept reading has no fixed numerical context budget. Standalone helper production is not later read-back.); Future loading of setup-created local guidance: uninspected (The instructions allow host loading or an explicit read request; actual host delivery and selection are uninspected. Neither unsolicited push nor exclusively requested pull is established for this included part.); Later use of old health reports/changelog and offered output filing: not-determinable (Reports/logs lack a specified recurrent trigger and outputs require an unresolved follow-up. Retention does not establish an automatic later supply chain.)	not-determinable coverage (Inspection yields no supported positive signal, while unresolved included delivery routes prevent axis-wide inapplicability. No ranking, lexical match, embedding or coarse availability selector is invented.); Question-driven wiki and requested source/editorial pull chains: inapplicable (Within these named requested-consumption operations there is no independent push selector to classify. Agent judgment and relative paths used during pull do not establish inferred-judgment or identifier push signals.); Host delivery of newly created local guidance: uninspected (No inspected trigger/selector establishes whether or how a host pushes retained local guidance; folder identity alone is not a demonstrated identity-match selector.); Unspecified later report/log consumption and output-filing follow-up: not-determinable (No complete later selection chain or automatic supply is specified, so no controlled push signal follows.)	not-determinable coverage (No qualifying positive route is established by the accepted records. This is not a system-wide no: arbitrary source origins and missing later report/log consumers prevent a complete bounded absence classification. No learning benefit or parameter change is inferred.); Source-fed compilation, subsequent editorial updates, output filing and setup-created instructions: not-determinable (Automatic source-fed writes and later knowledge consumers are documented, but generic documents/chat/pasted content do not establish original activity-trace content on a qualifying route. Neither a positive trace-fed chain nor a bounded all-source negative is recorded. Setup choices alone are not a session-log extraction chain.); Operation records in changelog/health reports and possible feedback to later work: not-determinable (Logging activity or retaining diagnoses does not close a trace-to-durable-behavior-shaping-write-to-later-consumer chain. No automatic recurrent report/log consumer is specified; the missing driver bounds implementation but does not prove every instruction-mediated update absent.)	partial coverage (Knowledge, routing and new local instruction authority are supported. Transient audit status is not a retained criterion; no enforcement, validation, ranking or learning value is inferred merely from an audit reading memory. Unspecified later consumers prevent full coverage.); Originals, readable editions and compiled source/concept content at later compilation, verification and Q&A consumers: known — knowledge [claimed] (The actual intended consumers are named; neither activation nor answer benefit is observed.); Topic maps, wiki index/connection notices and registry status at Librarian navigation and pending-source selection: known — routing [claimed] (The described selection effect supports routing at instruction strength; IDs alone do not establish ranking.); Setup-created local scope/deviation guidance at future Librarian entry: known — instruction [claimed] (Requested instruction authority is not a machine-enforced write barrier.); Saved output substance at offered filing-back and health diagnoses at editorial repair: known — knowledge [claimed] (The finding covers the documented editorial use, not every unspecified future inspection.); Changelog history and unspecified later reload of previous health reports or non-prose outputs: not-determinable (Readable retention alone does not establish a later force. No specified old-log consumer, recurrent report consumer or non-prose consumption path resolves this unit.)	AAS-2026-10-15-funes-a9e6dbb3946a-01
GBrain	2	Accumulated page/fact/fence knowledge and eligibility controls; checkpoint corpus, manifests and harvested facts; takes, dream bodies and quarantine; retrieval events/history/weights; journal, cursors, mutable shared skills, durable subagent replay/task/result state and their write, maintenance and later-consumer routes. Included unresolved alternatives: provider-internal replay state, compiled context, pinned answers, connectors/drains, other dream/optimization/evaluator/quarantine/purge paths. Excludes static shipped instructions and executable policy as memory objects, ordinary unretained current-run outputs, external host compaction and conversation stores, external inference weights not established as accumulated memory, private corpora and live services.	partial coverage (Positive stores cover the inspected object inventory; opaque provider state and other included material remain unresolved. No SQLite, repository-store or graph-store value is inferred from source code being in Git or pages having links.); Fact text, eligibility/provenance fields and vector output: known — rdbms [wired], vector [wired] (SQL and vector parts are both established; no model-weight storage is inferred.); Captured/entity page bodies and facts fences: known — rdbms [wired], files [wired] (File-backed and DB-only branches are both included; files are not universal.); Rendered checkpoint corpus and retry sidecars: known — files [wired] (Corpus is independently classified from SQL manifest state.); Checkpoint pointer manifest: known — rdbms [wired] (Readable titles do not change the SQL substrate.); Withdrawal ledger and consolidated takes: known — rdbms [wired] (These are retained negative control and derived knowledge, respectively.); Retrieval attribution, rating history and localized weights: known — rdbms [wired] (Weights outlive event/history pruning; their numerical encoding does not imply model weights.); Persistence journal and shared-skill catalog/revisions/files: known — rdbms [wired], files [wired] (Only retained publication/replay state and mutable skill content are included, not the shipped static tool policy.); Subagent replay, queued task/snapshot and retained result: known — rdbms [wired] (This covers durable GBrain replay and polling, not an external conversation runtime.); Dream page body and quarantined frontmatter: known — rdbms [wired], files [wired] (Returned unsaved answer variants are not presumed retained objects.); Opaque provider-internal replay/state alternative: uninspected (Readable SQL replay establishes no provider-native persistence substrate; external provider internal state remains uninspected.); Other included producers and consumers: patterns, concepts, atom extraction, skillopt, pinned answers, compiled context, connectors/drains, quarantine/purge and evaluator state: uninspected (Accepted records retain incomplete inspection of these alternatives. Their retained substrates are not enumerated; labels and driver calls cannot complete classification.)	partial coverage (Automatic admission and one explicitly interactive manual mechanism are established. Initial control, owner designation, host import and other producers remain separately unresolved.); Software admission of page/fact proposals, skill revisions and withdrawal: known — automatic [wired] (Authorship and human control of proposed bytes remain separate unresolved contributions.); Extraction, checkpoint corpus/manifest writes, consolidation and dream repair: known — automatic [wired] (Starting maintenance or enabling harvest is not manual admission of each result.); Feedback weight/history update, cursor advancement, replay and result settlement: known — automatic [wired] (The rating proposal may be human or computational; the bounded update admission is software.); Interactive local owner trust confirmation: known — manual [wired] (This manual witness applies only to interactive confirmation, not remote memory_confirm authority.); Remote owner confirmation: known — automatic [wired] (Software scope/state checks admit the tier update; principal identity does not demonstrate a human decision.); Caller proposal authorship, initial corpus, core owner designation and trusted reviewed skill import: not-determinable (Owner-plane labels and generic caller identity do not establish per-content human decisions; reviewed-import implementation is not traced.); Other included producers and consumers: patterns, concepts, atom extraction, skillopt, pinned answers, compiled context, connectors/drains, quarantine/purge and evaluator state: uninspected (Accepted records retain incomplete inspection of these alternatives. Their write/admission mechanisms are not enumerated; labels and driver calls cannot complete classification.)	partial coverage (Both pull and push exist. Internal selection, transport return and model-facing supply are distinguished; complete delivery coverage remains unresolved.); Requested query/image, recall, explicit pack/delta, takes and skill/result retrieval: known — pull [wired] (Code delivery fulfilling these requests is not a second push.); SessionStart pack and OpenClaw assembly selections: known — push [wired] (Upstream hook IPC is requested transport; the outer unsolicited model-facing supply is push.); Resumed subagent replay and later internal ranking/update consumers: known — push [wired] (These automatic internal selections are distinct from a final caller requesting evidence.); Compiled-file/installed-skill consumers, pinned answers and alternate provider state: uninspected (Known pull or push paths cannot resolve other included delivery mechanisms.); Other included producers and consumers: patterns, concepts, atom extraction, skillopt, pinned answers, compiled context, connectors/drains, quarantine/purge and evaluator state: uninspected (Accepted records retain incomplete inspection of these alternatives. Their later-consumer request and supply operations are not enumerated; labels and driver calls cannot complete classification.)	partial coverage (Coarse and identity-based push have local selector witnesses. Requested query ranking is not used to invent inferred push signals; reflex and alternate selectors remain unresolved.); Core designation push on assembly/pack callback: known — coarse [wired] (Designation/availability/budget selection is coarse, not query relevance or an identifier match merely because pages have slugs.); Session-scoped checkpoint push on assembly: known — identifier [wired], coarse [wired] (Actual session identity selects delivered state; page slugs inside pointers do not by themselves establish targeted full-page push.); Subagent resume push: known — identifier [wired] (Identity selects reused outcomes and recovered model context, not merely a catalog label.); Later ranking and extraction internal push selectors: known — identifier [wired], coarse [wired] (These are fixed-rule internal selectors; dense embeddings belong to other similarity consumers, not this weight lookup.); Salient entity reflex and hook entity-card targeting: not-determinable (Records establish detection/resolution and explicit entity inputs but do not establish lexical versus embedding versus model-judgment selectors for every unsolicited part.); Other compiled/skill/pinned/provider push alternatives: uninspected (No complete pull-only boundary exists; signal cannot be made inapplicable by inspected requested routes.); Other included producers and consumers: patterns, concepts, atom extraction, skillopt, pinned answers, compiled context, connectors/drains, quarantine/purge and evaluator state: uninspected (Accepted records retain incomplete inspection of these alternatives. Their push triggers, selector inputs and selected parts are not enumerated; labels and driver calls cannot complete classification.)	partial coverage (Positive wired trace-fed durable-context and ranking routes coexist with unresolved writes. No system-wide no, observed learning benefit or provider-parameter update is asserted.); Compaction corpus to harvested facts/fences and verified pointers: known — yes [wired] (Rendered trace to durable behavior-shaping context qualifies without novel knowledge or measured improvement.); Attributed retrieval event/rating to later page-weight ranking: partial — yes [wired] (Wired later consumer establishes qualification; link consumers and implicit signals remain unresolved.); Recorded transcript to dream page production/repair/retrieval: partial — yes [wired] (Producer modes/output semantics remain partly uninspected, but the trace-fed retained-context path is wired.); Mixed page/turn extraction outside checkpoint route: not-determinable (Some input may be conversational trace and some static pages; complete origins and later per-input routes are untraced. Known checkpoint witness does not classify every extraction input.); Raw subagent replay, initial authored content and opaque provider state: not-determinable (Raw replay is recovery/context retention, not independently evidenced transformation into an improved strategy or derived guidance. No bounded absence record licenses a local no for every possible producer.); Other included producers and consumers: patterns, concepts, atom extraction, skillopt, pinned answers, compiled context, connectors/drains, quarantine/purge and evaluator state: uninspected (Accepted records retain incomplete inspection of these alternatives. Their qualifying automatic trace-fed writes and later consumers are not enumerated; labels and driver calls cannot complete classification.)	partial coverage (Knowledge, instruction, learning, ranking, routing, validation and enforcement have local witnesses. Unknown asset activation, link weights and other consumers prevent a complete authority inventory.); Fact/page/take/dream content at recall, query, take and pack readers: known — knowledge [wired] (Rejected frontmatter supplies review evidence, not endorsed knowledge. Delivery establishes supply, not compliance.); Designated core prose at host assembly and pack consumers: known — knowledge [wired], instruction [wired] (Authority applies to consumed core text, not every static shipped prompt; default disabled and filtering qualify delivery.); Fact eligibility/withdrawal state at admission and delivery: known — validation [wired], enforcement [wired], routing [wired] (Negative controls have fixed-rule force; provenance strings alone do not validate truth.); Retained fact text/confidence/vector features at retrieval and consolidation selectors: partial — ranking [wired] (This is selector influence, not a learned ranking-policy claim; full question/hybrid ranking remains unresolved.); Checkpoint corpus and retrieval activity/feedback at automatic update consumers: known — learning [wired] (Learning authority at the update consumer is separate from output knowledge/ranking authority; improved capacity is unobserved.); Retrieval events at rating admission and history at retry/rate-cap checks: known — validation [wired], enforcement [wired] (These parts do not directly rank later queries.); Derived page weights at hybrid ranking: partial — ranking [wired] (Link-weight consumers remain untraced; page witness does not upgrade them.); Manifest, delivery cursors, replay, journal, skill heads and delegated job snapshot at fixed-rule consumers: partial — routing [wired], enforcement [wired], validation [wired] (Specific forces belong to the named consumers, not every field. Downstream skill-body/asset activation is uninspected.); Checkpoint title/pointers and retained replay/result text at later readers: known — knowledge [wired] (Renderer instructions are shipped static text, not independently learned standing guidance.); Published skill bodies/assets at downstream hosts and opaque provider state: uninspected (Publication/get-asset affordances do not establish installed instruction or executable enforcement consumers.); Other included producers and consumers: patterns, concepts, atom extraction, skillopt, pinned answers, compiled context, connectors/drains, quarantine/purge and evaluator state: uninspected (Accepted records retain incomplete inspection of these alternatives. Their actual consumers and behavioral force are not enumerated; labels and driver calls cannot complete classification.)	AAS-2026-10-11-gbrain-9de8e3c225d3-01
GMemory	2	Included: accumulated task payloads and interaction graphs, Chroma task access vectors, persisted task-query graph, rule text and association/survival metadata, automatic extraction/revision/merge and their later consumers, MacNet cross-task temporal histories, unresolved initialization and opaque provider retention. Excluded: comparator memories, shipped static instructions/examples as such, ordinary untransformed current-task accumulation, model weights with no evidenced use-driven update, and benchmark/provider dependency internals beyond the recorded opaque-state limits.	partial coverage (Positive units cover task documents, access graph, rules and temporal history. Hidden endpoint retention remains unresolved; dependency internals do not justify sqlite or rdbms values.); Task text, extras and interaction graphs in current objects and Chroma documents: known — in-memory [wired], vector [wired] (Text metadata and serialized interaction graphs are parts of one retained task document; JSON encoding alone does not establish a file substrate.); Task-query access graph: known — graph [wired], files [wired], in-memory [wired] (Covers query graph, not interaction-graph JSON or a hypothetical insight-edge graph.); Rule text and access/survival metadata: known — files [wired], in-memory [wired] (Current JSON and in-process list are evidenced; no version archive is inferred.); MacNet temporal input/output histories: known — in-memory [wired] (No restart checkpoint is established for these lists.); Opaque endpoint retention and any provider-side memory substrate: uninspected (Provider hidden retention and changes are uninspected. The client model-name string is not evidence of accumulated provider memory; no controlled value or absence is inferred.)	partial coverage (All inspected normal writes are automatic; initialization and external replacement control remain unresolved. No manual write witness is supplied.); Normal task capture, extraction and graph/index construction: known — automatic [wired] (Launch/configuration does not make subsequent content admission manual.); Ordinary rule operations, score updates and pruning: known — automatic [wired] (Model authorship and software admission are distinct from human approval.); Cluster merge replacement: known — automatic [wired] (The separate numbered-list gate also has no per-content human veto.); MacNet temporal appends: known — automatic [wired] (No task-success admission condition governs these appends.); Preexisting stores and external replacement admission: not-determinable (Loading existing Chroma/pickle/JSON identifies neither a human decision nor software admission of that initial content. Manual cannot be inferred.); Opaque endpoint retention and any provider-side memory admission: uninspected (Provider hidden retention and changes are uninspected. The client model-name string is not evidence of accumulated provider memory; no controlled value or absence is inferred.)	partial coverage (Pull at solving retrieval coexists with automatic maintenance and temporal delivery. Provider hidden later-consumer paths remain uninspected.); Task-start retrieval request through returned examples/rules and optional projection: known — pull [wired] (Covers all three shipped solving schedules; failed example returns are discarded.); Count-triggered maintenance selection of retained tasks and existing rules: known — push [wired] (Ordinary synthesis/comparison and cluster merge are later model consumers distinct from task-query return.); MacNet temporal predecessor delivery: known — push [wired] (Conditional on temporal edges and nonempty history; ordinary current-step accumulation is not counted as persistent task retrieval.); Opaque endpoint retention and any provider-side memory read-back: uninspected (Provider hidden retention and changes are uninspected. The client model-name string is not evidence of accumulated provider memory; no controlled value or absence is inferred.)	partial coverage (Signals classify only actual push operations. Pull relevance judgments are not counted; unresolved provider delivery prevents globally inapplicable or complete signal coverage.); Maintenance sampled-task selection: known — coarse [wired], inferred-embedding [wired] (Trigger is saved task count; inputs are sampled stored tasks/task names; selected parts are task records or rule batches. Opaque FINCH/model behavior does not become observed selection.); Maintenance related-rule selection: known — identifier [wired] (Actual task-name matches select rule text; names are not merely catalog labels.); MacNet latest temporal predecessor context: known — identifier [wired], coarse [wired] (Trigger is node execution; inputs are predecessor links/history; selected part is the latest retained output, not archived inputs.); Task-query pull chain: inapplicable (Push-signal classification is inapplicable to this fully identified request/return operation. Task embedding and model reranking here are pull selectors, not extra push signals.); Opaque endpoint retention and any provider-side memory push selection: uninspected (Provider hidden retention and changes are uninspected. The client model-name string is not evidence of accumulated provider memory; no controlled value or absence is inferred.)	partial coverage (The inspected durable task-derived and rule routes support yes. Other routes remain local unknowns; fixed client-side model wrappers do not establish a system-wide no.); Automatic task-derived extras retained for later prompts and update consumers: known — yes [wired] (Durable derived context qualifies without observed improvement; raw graph retention alone is not the warrant.); Automatic rule generation, ordinary revisions and cluster replacement: known — yes [wired] (Merge consumes trace-derived rules indirectly; this is text/data revision, not weight training.); MacNet temporal history as qualifying durable learned artifact: not-determinable (In-process append/forward is established, but no restart checkpoint or durable transformed artifact is established. This is not a system-wide no.); Externally initialized content and replacement route: uninspected (Original input, admission control and qualifying trace-to-artifact chain of existing stores are uninspected.); Opaque endpoint retention and any provider-side memory trace-fed writes: uninspected (Provider hidden retention and changes are uninspected. The client model-name string is not evidence of accumulated provider memory; no controlled value or absence is inferred.)	partial coverage (Authority is classified per actual consumer: learning, advice, selection and survival differ. Unconsumed metadata receives no force; hidden provider effects remain uninspected.); Readable successful examples and rule advice at solving agents: known — knowledge [wired] (No instruction value is inferred from the word rule; actual consumption is advisory.); Task experience and existing rules at rule-update consumers: known — learning [wired] (Update authority is separate from downstream advice and does not establish improved capacity.); Task names/vectors, query graph and positive rule associations at selectors: known — ranking [wired], routing [wired] (Survival scalar is not the rule relevance score; negative associations are not subtracted by the selector.); Retained rule scores at survival pruning: known — enforcement [wired], validation [wired] (This criterion controls survival, not semantic truth or action correctness.); Existing retained rule strings at duplicate rejection: known — validation [wired], enforcement [wired] (Merge is a different gate; AGREE targeting is not guaranteed by numeric checks.); MacNet temporal predecessor outputs at later nodes: known — knowledge [wired] (Archived inputs not selected by that getter have no evidenced downstream authority.); Opaque endpoint retention and any provider-side memory authority: uninspected (Provider hidden retention and changes are uninspected. The client model-name string is not evidence of accumulated provider memory; no controlled value or absence is inferred.)	AAS-2026-10-15-gmemory-a2192cc823f7-01
gnosis	2	Accumulated entry bodies and structural fields in .gnosis/entries.jsonl, their retained SQLite FTS projection and mtime control value, caller admission and replacement, maintenance, and later search/show/latest/topics consumers. Includes instructed external-agent consultation and repository sharing, with unresolved caller control and original provenance. Excludes static embedded doctrine as a memory object, transient topic aggregates, unused library capabilities as deployed routes, and external host windows, compaction, model weights and hidden session machinery.	known coverage (The accepted inventory covers canonical content, access projection and refresh control. No opaque provider state is included in this CLI memory boundary.); Canonical entry bodies and structural fields: known — files [wired], repo [afforded] (Records identify the canonical corpus and its instructed sharing path; unresolved authorship does not obscure its storage.); Retained lexical projection and freshness control: known — sqlite [wired] (Both operative cache parts are identified; disposability does not prevent retention between invocations.)	partial coverage (Automatic derived-state and field updates are supported; unresolved per-content caller control prevents a complete agency classification or a manual finding.); Caller decision supplying, replacing or removing retained body/metadata through CLI or direct JSONL edits/merges: not-determinable (The records leave human versus computational decision-making unresolved. An editor, author flag, Git identity or explicit command is not per-content human approval; software I/O alone does not resolve semantic admission control.); Code-generated IDs/timestamps and automatic normalization on admission or reload: known — automatic [wired] (Known only for these generated/normalized structural parts.); FTS projection and freshness timestamp admission on automatic refresh or requested reindex: known — automatic [wired] (Both refresh triggers share program-controlled derived-state admission.)	known coverage (The accepted inventory and memory annotation cover the shipped dispatch and later read routes without an unresolved alternative delivery mechanism inside this boundary. Actual invocation remains unobserved.); Later search, show, latest and topics requests, including lexical selection and formatted return: known — pull [wired] (ID/topic full reads, lexical search and recent/topic browsing cover the recorded later-consumer routes. Static help is not accumulated memory; external hooks are excluded.); Retained access state consumed during requested search: known — pull [wired] (Automatic refresh is a derived-state write inside this chain, not an independent read-back push.)	inapplicable coverage (No included unresolved push alternative remains in the recorded CLI boundary. Host-side context injection is excluded rather than declared absent.); All scoped later read-back operations: inapplicable (The completely classified direction inventory is request-driven. FTS lexical matching, ID matching and recent-entry filters are pull selectors here, not push signals.)	not-determinable coverage (No positive qualifying automatic trace-fed route is established, and unknown caller control prevents a complete negative. This is not an absence claim about external agents.); Caller-mediated session-to-entry writing and subsequent semantic replacement: not-determinable (Session-derived durable prose and a later advisory consumer are afforded, but automatic per-content admission and original trace inputs are unresolved. Thus neither yes nor a bounded no is established for this route. Missing observed improvement is not the reason for withholding yes.); Generated structural fields and JSONL-to-FTS/mtime updates: inapplicable (These automatic updates take entry values and file metadata, not an established activity-trace learning input. Indexing is not an additional trace-fed behavior-shaping write merely because upstream prose may describe experience.)	known coverage (Known covers the recorded consumers and their forces, not actual compliance. Display-only attribution/navigation does not add authority, and unknown upstream composition does not obscure these local consumers.); Retained entry prose at the instructed later coding-agent consumer: known — knowledge [afforded] (This is advisory knowledge consumption, not standing host instructions or a code-edit veto; static doctrine is excluded from accumulated memory.); Retained body/topic features and creation times at search and recent-entry ordering consumers: known — ranking [wired] (The authority is output priority, not credibility or demonstrated task influence.); Retained entry IDs/topics at lookup and membership selection; cached mtime at refresh branching: known — routing [wired] (Attribution and displayed Related fields have no additional executable supersession force; mtime routing is not semantic endorsement.); Existing retained IDs at related-reference admission and edit checks: known — validation [wired], enforcement [wired] (This force belongs to retained IDs as structural criteria, not to the prose's truth or an authenticated author. Pre-append resolution and interactive locked resolution have different timing guarantees.)	AAS-2026-10-15-gnosis-5bfb42422ce3-01
Graphiti	2	Included: accumulated episode/entity/fact content, typed and temporal/lineage access parts, persisted graph embeddings, community and saga summaries/cursors, notebook conversation checkpoint, evaluation graph/results files, and their admission, maintenance and later-consumer routes. Other shipped example/evaluator, backend and consumer alternatives remain included as unresolved coverage. Excluded: static shipped prompts/rules/schema/configuration as memory objects, transient ingestion queues and current-call objects, external provider weights/internals, proprietary Zep and unregistered historical notebook outputs. Shipped MCP rules supply evidence of an afforded consumer role, not accumulated instructions.	partial coverage (Graph, files and an in-process checkpoint are supported locally. Remaining included backend and consumer variants prevent complete inventory coverage. Provider weights, static prompts and admission queues are not accumulated knowledge stores in this scope.); Graphiti episodes, entity/fact text, attributes, identities, temporal fields, membership, vectors and community/saga state: known — graph [wired] (These operative retained parts are graph properties/relationships through driver writes; transient construction objects are not an additional retained store.); Evaluation baseline/candidate JSON exports and retained evaluation graph: known — files [wired], graph [wired] (The comparison route serializes results locally and reloads baseline files; evaluation graph state is separately retained.); Notebook MemorySaver checkpoint: partial — in-memory [wired] (Shipped code requests an in-process checkpointer; its exact internals and durability are outside the frozen source.); Remaining included backend/index serialization and other example/evaluator retained parts: uninspected (Backend roundtrips and other included consumers/evaluators were not exhaustively traced; the inspected graph/file witnesses do not close the retained-object inventory.)	partial coverage (Automatic derived admission and notebook recording are supported. Generic caller control, checkpoint admission internals and other shipped paths remain unresolved; no manual value is invented from explicit API invocation.); Model/software extraction, resolution, embeddings, summary assignment and derived maintenance: known — automatic [wired] (A caller starts these operations; subsequent per-content admission is software/model-controlled.); Notebook post-response episode admission: known — automatic [wired] (Scheduling is unconditional after each successful model response, including tool-call rounds, not just completed human turns.); Evaluation result export: known — automatic [wired] (Explicit evaluation invocation does not make software export a manual content decision.); Generic episode/direct object supply, explicit removal/clear and configuration of maintenance calls: not-determinable (Caller identity does not establish an explicit human decision supplying, approving, editing or replacing each retained item. Neither manual nor automatic content admission is established for this caller-controlled part.); Checkpoint admission mechanism inside MemorySaver: uninspected (Shipped code requests checkpointing, but the external framework admission mechanism was not inspected.); Remaining included example/evaluator admission alternatives: uninspected (No complete per-content control inventory was retained for other shipped callers.)	partial coverage (Pull applies to requested graph/file reads; push applies to automatic context supply to a later model/update/judge consumer. This does not claim an independent unsolicited service notification or successful model activation. Unresolved alternatives prevent pull-only or exhaustive-direction conclusions.); Library/REST/MCP query and direct retained-object returns: known — pull [wired] (A requesting caller selects retained material; returning that requested result is not another push.); Notebook graph search and automatic context assembly: known — pull [wired], push [wired] (Code requests graph facts, then supplies selected text to the model without a model request for those user facts; operations within one chain are distinct.); Prior episode/candidate and previous-summary supply inside ingestion and maintenance: known — push [wired] (The caller requests a write/update, not return of this context; internal selectors supply retained inputs to extraction or summarization consumers.); Evaluation baseline read and judge-context assembly: known — pull [wired], push [wired] (Reload and delivery to the judge are separate operations; the judge does not request a particular baseline object.); Checkpoint resumption and delivery of accumulated messages: partial — push [wired] (Notebook code sends accumulated messages to the model; the external framework’s reload/selection operations are not inspected.); Other included example/consumer and checkpoint read-back alternatives: uninspected (Existing request-driven graph returns do not close every later-consumer delivery path.)	partial coverage (Signals describe automatic context selection, not every requested return. Each unit identifies its trigger, selector input and delivered retained part; direction uncertainty prevents a global inapplicable judgment.); Notebook user-fact context push: known — inferred-lexical [wired], inferred-embedding [wired], identifier [wired] (Trigger is chatbot invocation; latest message text and user center node select up to five edge.fact strings delivered in a system message.); Prior episode selection for extraction: known — coarse [wired], identifier [wired] (Trigger is ingestion. Time/source/group/saga and last_n select retained prior episode text; explicit prior UUIDs can replace temporal selection.); Graph candidate selection for extraction/resolution: partial — inferred-lexical [wired], inferred-embedding [wired] (Ingestion triggers candidate searches supplying retained entity descriptions and fact text to extraction/resolution consumers. Lexical and embedding selection are supported; subsequent identity/semantic judgments concern reuse and temporal disposition, not selection for push delivery. Optional cross-encoder search is not established on this ingestion push, and alternate recipes/provider selectors remain unresolved.); Saga summary and linked episode push to summarizer: known — identifier [wired], coarse [wired] (Trigger is summarize_saga. Saga UUID/membership selects the retained summary/linked episodes; ingestion cursor, count and text/byte budgets bound the submitted window.); Entity/community retained-summary update inputs: partial — identifier [wired], coarse [wired] (Trigger is an update/rebuild. Matching node names and community memberships identify inputs; not all cluster or callback variants are inspected.); Evaluation baseline push to judge: known — coarse [wired] (Trigger is candidate evaluation; selected dataset prefix and corresponding positional zip select retained baseline results for the judge, not authenticated identity matching.); Checkpoint message context push: partial — identifier [wired] (Trigger is chatbot invocation; thread key selects conversation state at the notebook boundary; exact MemorySaver payload selection is opaque.); Remaining included selector recipes and consumer alternatives: uninspected (Other example routes, framework internals and all backend search variants remain untraced; known signals do not classify opaque alternatives.)	partial coverage (Wired yes witnesses cover conversation-derived graph memory and dataset-session graph results, not exercised or beneficial learning. Generic origins, checkpoint transformation and other shipped paths keep axis coverage partial.); Notebook automatic mixed-origin episode-to-graph-to-later-model chain: known — yes [wired] (The chain records latest human/tool/model state plus response, produces retained graph claims and later supplies selected facts to the conversational model. Scheduling/persistence success and benefit remain unobserved.); Evaluation dataset session processing into retained baseline/candidate graph: known — yes [wired] (Recorded session messages feed automatic graph construction; retained descriptions later shape judge context, not automatic installation of a production winner.); Saga/entity/community summaries from generic caller episodes and retained descriptions: partial — yes [wired] (Automatic transformations and later consumers are supported, but whether every original episode is an activity trace is unresolved.); Raw checkpoint-only retention: not-determinable (Recorded messages are retained and supplied as context, but external checkpoint internals establish no derived behavior-shaping update. Raw retention alone does not justify yes, and no bounded absence search warrants no.); Generic direct content writes and untraced shipped example/evaluator transformations: uninspected (Trace provenance and admission mechanisms are unresolved across remaining included paths; no system-wide no or complete qualifying-route enumeration is supported.)	partial coverage (Knowledge, ranking, routing, conditional enforcement, trace-update learning and afforded procedural instruction have separate consumers. No validation force is inferred from excluded static schemas; provider/host opacity is not upgraded by notebook wiring.); Episode/fact/summary text at extraction, resolution, summarization and notebook answer consumers: known — knowledge [wired] (This names actual text consumers; exposure to an unspecified API caller is not additional behavioral activation.); Retained text and vectors at hybrid search and relevance reranking: known — ranking [wired] (Ranking concerns relative retrieval priority, not truth acceptance.); Retained IDs, temporal/group selectors, memberships and saga cursor at selection/remapping consumers: known — routing [wired] (Code uses retained access state to choose scope, candidates and the next input window.); Retained episode references and temporal fields at deletion/date-filter consumers: known — enforcement [wired] (The force is conditional on invoking cleanup or a date-aware filter; no global current-truth-only enforcement is claimed.); Retained conversation episodes at automatic graph extraction and summary update: known — learning [wired] (Update-consumer force is distinct from the resulting facts’ knowledge force; no improvement or parameter training is implied.); Retained preferences/procedures at an external assistant following shipped MCP rules: known — instruction [afforded] (This is an afforded role only: host loading, tool-name mapping and compliance remain uninspected.); Retained evaluation reference results at model judge: known — knowledge [wired] (Reference content is model-produced, not independently certified truth or production admission policy.); Checkpoint message/state at resumed notebook model: partial — knowledge [wired] (Messages reach model context at the notebook boundary, while framework-internal selectors and effects remain uninspected.); Other included consumers, attribute-specific effects and checkpoint internals: uninspected (The supplied records do not enumerate every later consumer/effect; these unresolved parts prevent a complete authority inventory.)	AAS-2026-10-12-graphiti-1d535ca565a4-01
Halo	2	Included: retained execution telemetry and JSONL exports; derived trace indexes/freshness metadata; desktop user/assistant turns, result files, raw analysis events and Markdown reports; saved AppWorld evaluation artifacts; optional terminal compaction summaries and host continuation alternatives; unresolved initial caller continuations and opaque provider-side use-acquired/caching state. Excluded: ordinary disposable current-run context/tool results except checkpoint production and local backing parts of that affordance; static shipped engine/skill prompts; inference weights and installed engine code as machinery; arbitrary external harness edits/deployment and coding-tool internals.	partial coverage (The accepted memory inventory establishes local substrates; external checkpoint persistence and opaque provider state prevent a complete substrate inventory. Ordinary disposable tool results, static prompts and installed engine code are excluded.); Captured telemetry tables and exported/caller JSONL files: known — sqlite [wired], files [wired] (The inspected retained telemetry and export parts have separately identified substrates; upstream provider storage is not covered by this unit.); Trace index rows and freshness sidecars, including their loaded access structures: known — files [wired], in-memory [wired] (The record identifies both disk sidecars and loaded TraceStore memory.); Desktop conversation turns and per-turn result artifacts: known — sqlite [wired], files [wired] (SQLite text rows and result JSON are the recorded durable conversation surfaces.); Desktop analysis event payloads: known — sqlite [wired] (Live publication is not counted as another storage substrate.); Materialized report and artifact locator metadata: known — files [wired], sqlite [wired] (The file content and SQLite path/size metadata remain distinct retained parts.); Saved AppWorld evaluation artifacts: known — files [wired] (The immediate TestTracker is not classified as durable memory independently of saved reports.); Optional compaction summaries and host-persisted checkpoint continuation: partial — in-memory [wired] (Summary storage is known locally, but the checkpoint host chooses persistence; no desktop checkpoint store is wired.); Included opaque provider-side retained or caching state, if any; not the inference weights as shipped machinery: uninspected (The accepted model-component records leave provider internals uninspected. Local inference and cache-oriented configuration do not classify provider-side use-acquired memory for this axis: storage substrate.)	partial coverage (The named local admission mechanisms are classified independently of physical writes and authorship. Unknown initial input control and provider admission prevent completeness.); Explicit human supply of desktop question content: known — manual [wired] (The user supplies the retained question itself, unlike merely starting automatic diagnosis.); Telemetry normalization/upsert, export and trace-index derivation: known — automatic [wired] (Software admission is separate from unknown upstream authorship.); Desktop answer/event admission and legacy-turn reconstruction: known — automatic [wired] (Human run initiation does not approve each generated answer or event.); Report preparation and freshness replacement: known — automatic [wired] (Automatic terminal preparation and explicit UI requests use the same software content rule.); Evaluation result persistence: known — automatic [wired] (save_report is a software persistence choice, not per-result human approval.); Optional compaction summary admission: known — automatic [wired] (Successful complete units replace rendered items without semantic or human review.); Initial caller-file content and externally supplied system/continuation replacement: not-determinable (Generic caller identity and arbitrary source files do not establish a human per-content decision or software admission mechanism.); Included opaque provider-side retained or caching state, if any; not the inference weights as shipped machinery: uninspected (The accepted model-component records leave provider internals uninspected. Local inference and cache-oriented configuration do not classify provider-side use-acquired memory for this axis: write agency.)	partial coverage (Pull and push coexist within the shipped local chains. Follow-up history is push selection despite a requested new analysis; external checkpoint/provider alternatives prevent a complete direction inventory.); Requested trace evidence retrieval and desktop export: known — pull [wired] (Model/user requests select retained telemetry; code delivery fulfills those requests.); Follow-up history selection and supply to the next root invocation: known — push [wired] (The new question triggers automatic selection of prior content, rather than an explicit request for those retained turns.); Report materialization: automatic terminal route and explicit preparation route: known — push [wired], pull [wired] (The two triggers share a renderer but retain different request/selection operations.); Report reading by user-selected external coding actor: known — pull [afforded] (User-confirmed path access is afforded, not inspected receiving-tool operation.); Stored activity UI and saved evaluation inspection: known — pull [afforded] (Human consumers have retained display/report access; autonomous agent replay is not inferred.); Optional checkpoint host persistence and continuation resupply: not-determinable (A later continuation consumer is afforded, but the host request versus automatic selection/delivery mechanism is unspecified.); Included opaque provider-side retained or caching state, if any; not the inference weights as shipped machinery: uninspected (The accepted model-component records leave provider internals uninspected. Local inference and cache-oriented configuration do not classify provider-side use-acquired memory for this axis: read-back direction.)	partial coverage (Identifier matches and coarse inclusion filters are distinct supported findings on local pushes. Pull selectors do not establish push signals; host/provider uncertainty prevents system-wide inapplicability or completeness.); Automatic follow-up history supply: known — identifier [wired], coarse [wired] (Trigger: new follow-up. Selector inputs: selected run, upToTurnIndex and role/content. Selected parts: earlier user and nonempty assistant turns.); Automatic terminal report projection: known — identifier [wired], coarse [wired] (Trigger: terminal answer handling. Selector inputs: run ID, role/status/content. Selected parts: user turns and completed/incomplete nonempty assistant turns.); Requested trace/export, explicit report and human evaluation/activity reads: inapplicable (These requested operations are pull; their IDs/filters are not automatically classified as push signals.); Optional host checkpoint replay selection: not-determinable (No host selector is recorded. Item IDs printed in checkpoint messages do not establish identity-matched push selection.); Included opaque provider-side retained or caching state, if any; not the inference weights as shipped machinery: uninspected (The accepted model-component records leave provider internals uninspected. Local inference and cache-oriented configuration do not classify provider-side use-acquired memory for this axis: read-back signal.)	partial coverage (Positive local trace-fed retained-context routes coexist with unresolved alternatives. No bounded absence record supports a local or system-wide no. The axis does not assert parameter training, observed improvement, autonomy or closed self-improvement.); Trace investigation to automatically retained diagnostic answer to later desktop follow-up: known — yes [wired] (Raw trace storage alone is not the warrant: automatic model diagnosis, durable answer admission and wired later model delivery complete the scoped chain.); Conversation/tool-item compaction to host-persisted and re-supplied checkpoint: known — yes [afforded] (The write is wired, but durable host persistence and later replay are only afforded, so this chain is not upgraded to wired.); Initial externally supplied continuation and evaluation-to-durable-behavior alternative: not-determinable (Initial continuation provenance is unknown. Saved evaluation evidence and prescribed external shipping do not establish an inspected automatic durable behavior-shaping update; task-specific and receiving-tool mechanisms remain unresolved.); Included opaque provider-side retained or caching state, if any; not the inference weights as shipped machinery: uninspected (The accepted model-component records leave provider internals uninspected. Local inference and cache-oriented configuration do not classify provider-side use-acquired memory for this axis: trace learning.)	partial coverage (Knowledge, routing, validation, enforcement and learning are local consumer classifications, not system-wide compliance or benefit. Historical user requests do not establish acquired-memory instruction authority; caller-supplied retained system guidance remains unresolved. No retained ranking input with relative-priority effect is established by these accepted records.); Retained telemetry and trace views at diagnostic model consumers: known — knowledge [wired], learning [wired] (The same evidence has a diagnostic consumer and a downstream durable-answer production route; no truth acceptance follows.); Trace access rows and freshness metadata at TraceStore/startup consumers: known — routing [wired], validation [wired], enforcement [wired] (Access routing and cache admission are distinct effects; this unit does not imply model-interpreted instructions.); Retained conversation text at follow-up and report production: known — knowledge [wired] (Retained questions and diagnoses supply conversational context. Replayed historical requests do not establish standing directions, so these records support knowledge but not instruction.); Turn ordering/status and report locator metadata at selectors: known — routing [wired] (These are programmatic selection effects rather than semantic endorsement.); Retained activity payload at human display: known — knowledge [wired] (The raw event store is not wired into follow-up model replay.); Conversation report at the receiving coding actor: known — knowledge [afforded] (The report supplies findings; the static launch prompt is not classified as acquired memory.); Optional summary/tail checkpoint at a later host continuation consumer: known — knowledge [afforded] (A caller can persist and re-supply messages; ordinary desktop does not do so.); Saved evaluation outcomes at an experimenter: known — knowledge [afforded] (The inspected evaluator reports outcomes; external shipping remains separately prescribed.); Caller-supplied retained system guidance and provider-state consumers: not-determinable (A leading system message can bypass engine guidance, but acquired-content identity and authority are unspecified; opaque provider-state effects also remain uninspected.)	AAS-2026-10-14-halo-c4ed89d7f99f-01
Hermes Agent	2	Retained material accumulated or changed through use: built-in notes/profile; revised skill packages and access/lifecycle metadata; pending proposals; session traces, derived handoffs and checkpoint sidecars; optional provider state/recall; successful child-output spills; saved goal feedback/control; and named unresolved editable context-file, todo, context-engine, micro-compaction, native-thread and async-completion alternatives. Static bundled instructions/skills before use-derived change, ordinary current-run tool-call state, arbitrary application file effects, model-provider weights, plugin implementation installation, Python kernels and general command-approval policy are excluded from this compared retained-context boundary. External provider/native internals remain included as opaque operative alternatives, not attributed implementation.	partial coverage (Files, SQLite, process copies and service objects are positive witnesses, not a full inventory of the partial-loop boundary. Provider and alternate context surfaces prevent complete substrate coverage.); Built-in notes/profile, retained skill trees and pending proposals: known — files [wired], in-memory [wired] (UTF-8/JSON files retain content; live stores, prompt snapshots and index caches hold process copies. Plugin-registry skill alternatives are separately unresolved.); Session text, local summaries and Responses checkpoint sidecars: known — sqlite [wired], in-memory [wired] (Serialized session rows and working message copies retain distinct raw, derived and opaque parts; opacity does not obscure the local row substrate.); Provider-side RetainDB queue and named external memory objects: partial — sqlite [wired], service-object [wired] (The queue is local SQLite; remote messages/facts/files/profile are service-owned objects. Their underlying databases or embeddings are not established.); ByteRover persistent tree/cloud internals and optional provider implementations: not-determinable (CLI ownership and a knowledge-tree display do not identify the underlying substrate.); Successful delegated-answer spills: known — files [wired] (Full child output is written to cache files when spill succeeds, separately from its returned head/tail.); Saved goal verdict/control state: not-determinable (Manager saving is wired, but these records do not identify its durable substrate.); Plugin skill payloads, editable context files, todo rehydration, external context engines, app-server native context and durable async completions: uninspected (Named included alternatives have no complete retained-object substrate account in the accepted records. Spill reload, frontend retention and rolling-summary lifecycle are also incomplete.)	partial coverage (Both automatic and explicit per-content manual admission exist. Initial/control-unknown and remote alternatives remain separate; the tool gate is not universal and UI mutation bypasses it.); Ungated foreground memory/skill admission and review additions: known — automatic [wired] (Software/model admission can commit content without an explicit human decision about each retained item. Starting a request or choosing configuration is not manual admission.); Per-content pending approval and attended inline approval: known — manual [wired] (An explicit human-facing approve/apply decision admits a selected payload; reject leaves it unadopted. Staging alone is automatic proposal retention.); Pending proposal staging: known — automatic [wired] (Software stages the replay payload and removes it after successful apply; no human decision is required to create the proposal.); Compaction commit, deterministic aging/archive and provider queue/adapter ingress: partial — automatic [wired] (Code admits summaries, lifecycle changes and trace ingress; manual focus or enabling extraction is not approval of each result. Remote derived admission remains unresolved.); Delegated full-output spill and goal-state save: known — automatic [wired] (Host sizing rules create spill files; manager rules save verdict/status without per-content approval.); Existing files, learning-node UI mutation and hub/direct-file acquisition: not-determinable (The records identify mutation surfaces but not every per-content principal/decision; a UI label or generic caller does not prove human control.); Native checkpoint creation, remote provider derived writes, optional curator consolidation, micro/context-engine/todo and async completion admission: uninspected (Known adapter calls and local commits do not establish all alternate admission mechanisms.)	partial coverage (Both requested and unsolicited routes are supported. Known native adapter replay is distinct from opaque payload meaning; unresolved alternate chains remain included.); Whole notes/profile blocks and eligible skill index at prompt assembly: known — push [wired] (Load/rebuild automatically supplies enabled snapshots and catalog descriptions without the later model requesting each entry.); Requested skill body/support files, session search/read and explicit provider query/read: known — pull [wired] (A model/caller request is fulfilled by selection and return. Code delivery within that requested chain is not a second push.); Local handoffs, latest native checkpoint run and automatic provider prefetch: known — push [wired] (Code supplies selected accumulated context to later requests without a retained-material request by that consumer. Opaque content does not remove the known adapter direction.); Requested caller resume: known — pull [wired] (The caller requests retained active history; it does not request every archived/audit row.); Full spilled answer requested by parent: known — pull [afforded] (The file pointer affords a later read_file request, not automatic paging.); Saved goal status and generated repair prompts: known — push [wired] (Later manager boundaries consume saved verdict/status automatically; continuing code enqueues repair guidance for a later model invocation.); Pending payload display/apply command: known — pull [wired] (An explicit list/diff/apply request loads pending retained payloads; adoption and later prompt supply are separate operations.); Auto-load/slash/plugin skills, micro/context engines, editable context files/todo, native app-server state and async completion consumption: uninspected (Alternate delivery exists or is named, but its complete request/selection/delivery chain is not traced. No pull-only or read-back-absence conclusion follows.)	partial coverage (Coarse automatic selection is established locally. Query-based provider calls do not reveal their actual matching mechanism; unresolved alternatives prevent a complete signal classification.); Notes/profile whole-block push and skill-index push: known — coarse [wired] (Trigger is prompt load/rebuild; selector inputs are enable flags and catalog eligibility/platform/disabled/tool filters; selected parts are entire eligible blocks/descriptions, not query-matched entries.); Local handoff/head-tail and latest checkpoint-run push: known — coarse [wired] (Pressure/next-request triggers select by recency, grouping and budget; retained summary/tail or latest checkpoint plus plaintext budget slices are delivered. Positions and session IDs alone are not identity-targeted selection.); Provider automatic recall: partial — coarse [wired] (Nontrivial-turn/provider availability triggers consumption of caches or query responses. Coarse gates are known; RetainDB uses prior queued results and ByteRover a current-query CLI call. Exact remote matching signals are unresolved.); Remote provider context/profile selection algorithm: not-determinable (A query or prose synthesis name does not establish lexical, embedding or judgment selection of the actual retained payload. No exact targeting algorithm is recorded.); Saved goal control and repair delivery: known — coarse [wired] (Eligible goal boundaries consume status/verdict and conditionally enqueue the associated repair prompt; this is state/availability selection, not relevance ranking.); Requested skill/session/provider/pending/spill routes: inapplicable (Push-signal classification is inapplicable to these explicitly requested returns; their IDs/queries do not supply an independent push selector.); Auto-loaded/plugin skills, micro/context-engine and todo/context-file selectors, async/native-thread delivery: uninspected (Trigger/selector/selected-part triples are not fully recorded for these alternative pushes. Neither inferred-judgment nor identifier follows from model use or a catalog name.)	partial coverage (Review, local continuation handoffs and saved goal-feedback adaptation establish architectural yes findings. No system-wide no, observed learning benefit or exercised self-improvement is inferred; opaque and alternate writes remain unresolved.); Completed-turn review to admitted notes/skills and later prompt/view: known — yes [wired] (Automatic trace-fed lesson writes produce durable behavior-shaping content with later consumers; optional human-gated alternatives do not erase ungated admissions.); Local generated/deterministic continuation handoffs: known — yes [wired] (Automatic trace reduction is committed and supplied to subsequent inference. Knowledge novelty, lossless retention and measured improvement are not required.); Goal feedback to saved verdict/reason and later manager/repair input: known — yes [wired] (Automatic response/gate evaluation saves durable behavior-shaping state; later boundaries consume it and can deliver a repair prompt. This is not a claim of improved capacity.); Raw-only session archival and full-output spill: inapplicable (Raw trace retention/copy alone is not a qualifying behavior-shaping update. Compaction-derived handoffs and review writes are assessed separately.); Provider derived facts/profiles/tree and native encrypted checkpoint: not-determinable (Trace ingress and later adapters are known, but opaque transformation/admission and exact behavior-shaping output prevent a complete qualifying-write assessment.); Micro/context-engine updates, optional curator consolidation and other skill/context-file/todo/async writes: uninspected (Full raw-to-derived-to-later-consumer chains remain incompletely inspected. Foreground writes may also use selected traces but no complete trace-input inventory is recorded.)	partial coverage (Knowledge, instruction, learning, ranking, validation, enforcement and routing have local consumer warrants. Static guards and shipped instructions are not reclassified as accumulated memory; opaque and alternate consumers prevent complete coverage.); Retained note facts, recalled prose, history and readable handoffs at foreground model: known — knowledge [wired] (These parts supply reference/context at later inference. Delivery does not establish truth, compliance or benefit.); Retained preferences and skill procedures at foreground model: known — instruction [wired] (Preferences and procedures provide standing directions through prompt blocks or skill tool results, not executable enforcement.); Retained conversations at review and summary update consumers: known — learning [wired] (Trace content feeds automatic durable lessons or continuation handoffs. This authority concerns the update consumer, separate from downstream knowledge/instruction.); Retained session text and metadata at search ranking code: known — ranking [wired] (FTS text, lineage and cron demotion influence relative retrieval priority; no learned ranker is implied.); Skill lifecycle/provenance/read-mark/suppression metadata and pinned pending targets at admission/load code: known — validation [wired], enforcement [wired] (Protection/read marks and content pins constrain admissible mutations; archived/suppressed state constrains later catalog delivery. These retained controls differ from static scanner code.); Session active/compacted flags, checkpoint positions and saved goal status at runtime consumers: known — routing [wired] (Flags/positions select live projections or replay windows; saved goal state selects continuation, park or stop. Their role is operational routing, not factual endorsement.); Saved goal failure/reason text at subsequent repair model: known — instruction [wired] (Failure evidence is turned into a continuation/repair prompt delivered to a later model turn; successful repair is unobserved.); Spilled output at a later requesting parent: known — knowledge [afforded] (The footer affords paging full retained output with read_file; actual reload is not established.); Opaque checkpoints/provider internals, skill script execution, UI/context-engine/todo and async consumers: uninspected (Readable displays do not classify opaque consumers. Script support files may execute but full downstream effect is not established; alternate consumers remain incomplete.)	AAS-2026-10-14-hermes-agent-bfb4bdc54d37-01
Hindsight	2	Accumulated documents/chunks, extracted facts, observations, standing mental-model/page content, embeddings, provenance/access fields, refresh controls/grounding/history, directives, archive and bank overrides; shipped transcript admission, curation, recall/reflect, automatic maintenance, hook delivery, filesystem mirror and retained evaluation reporting. Alternative stores, attachments, other integrations/imports and embedded paths remain included but unresolved. Excludes static shipped instructions, current-request reflection working history, external host orchestration and provider-internal model weights/updates.	partial coverage (Known SQL, vector and file witnesses do not enumerate alternative stores or every shipped retained part. Provider inference weights and current-request lists are not added to the accumulated-memory inventory merely because they are used by the runtime.); Default memory-store document/chunk text, fact and observation text, access/provenance fields and source metadata: known — rdbms [wired] (The default SQL path is classified separately from the configured store alternative.); Fact retrieval vectors on the default database path: known — rdbms [wired], vector [wired] (Numerical encoding and vector access structure are distinct from the embedding model weights.); SQL standing-page prose, tree, edit baseline, refresh controls/grounding/history, directives, archive and bank overrides: known — rdbms [wired] (SQL page content survives even when an alternate store owns its search index.); Page-search vector on the SQL-backed search path: known — rdbms [wired], vector [wired] (The alternate page-index owner remains a separate unresolved unit.); CLI knowledge mirror and evaluation report files: known — files [wired] (File persistence is wired; later model reading of the mirror and full report consumers remain uninspected.); Configured memory store, configured page index and attachment payloads: uninspected (An extension may own rows or search outside PostgreSQL. Actual retained substrate and attachment storage are not established; an extension interface alone does not establish service-object or graph storage.); Additional integration-local and embedded retained state: uninspected (Additional shipped integrations, imports, attachments, embedded callers and alternate-store consumption remain incompletely inspected in the accepted records; complete system enumeration is prevented.)	partial coverage (Automatic transformation/admission is supported locally. Manual is not asserted without an explicit human content decision; unresolved initialization and API callers remain in the inventory.); Transcript-to-fact admission, observation proposals and automatic page refresh: known — automatic [wired] (Starting the host or maintenance workflow is not human approval of each admitted fact/edit.); Derived indexes, archive/restore representation rebuilding, mirror and report writes: known — automatic [wired] (The derived-output admission is separate from unresolved control of the initiating caller edit.); Generic retain ingress, caller page/directive creation, explicit curation and bank override replacement: not-determinable (The accepted records leave caller identity and per-item human control unresolved. Explicit API mutation alone establishes neither manual nor automatic admission for this initiating mechanism.); Initial content, alternate stores and additional shipped integration admissions: uninspected (Additional shipped integrations, imports, attachments, embedded callers and alternate-store consumption remain incompletely inspected in the accepted records; complete system enumeration is prevented.)	partial coverage (Both requested retrieval and independent automatic supply exist. Request, selection and delivery steps are distinguished within chains. Current-request synthesis history is not counted as a separate accumulated-memory read-back route.); Caller/model-requested recall, page search and explicit page-ID reads: known — pull [wired] (Fulfilling the requested retrieval is pull, not a second push merely because code serializes its result.); Claude Code per-prompt context injection: known — push [wired] (The host event triggers the hook; the answering consumer has not requested particular retained items.); Active directives and automatic maintenance inputs: known — push [wired] (An upstream reflect request or configured trigger does not make implicit direction/baseline loading a requested content return.); CLI server-to-disk snapshot synchronization: known — pull [wired] (This classifies snapshot acquisition/delivery, not an invented later agent file-read route.); Mirror later agent reading, report consumer details and additional shipped read-back: uninspected (Additional shipped integrations, imports, attachments, embedded callers and alternate-store consumption remain incompletely inspected in the accepted records; complete system enumeration is prevented. Mirroring capability and documented dashboards do not establish wired model consumption.)	partial coverage (Signals describe automatic supply operations only. Query/ID selectors on explicitly requested pull results are not added as push signals. Known pull paths cannot make unresolved push alternatives inapplicable.); Prompt-triggered hook selection of fact/observation context: partial — inferred-embedding [wired], inferred-judgment [wired] (Trigger is UserPromptSubmit; selector input is prompt/context-derived query plus bank/filter settings; selected parts are retained fact/observation text. Complete retrieval arms and configured ranking mode are unresolved.); Directive loading during reflect: partial — coarse [wired] (Trigger is a reflect invocation; active availability and optional all-active loading select directive prose. Exact tag-selector implementation is not recorded sufficiently to classify targeted identity matching here.); Automatic page maintenance baseline and grounding supply: known — identifier [wired] (A scheduled/explicit refresh selects its retained page baseline by model identity and uses stored source IDs to detect retracted grounding; supplied parts are baseline, query and maintenance evidence.); Other hook retrieval arms, tag targeting and additional push routes: uninspected (Complete selector implementations and alternative integrations are uninspected. Lexical selection is not inferred from an unspecified retrieval arm; identifier on a requested page alone is not a push witness.)	partial coverage (A concrete qualifying trace-fed write supports yes without proving improved capacity or parameter training. No bounded absence record warrants local or system-wide no; unresolved routes remain included.); Claude Code transcript retention into facts and optional derived observations/pages, then later host context: known — yes [wired] (The durable behavior-shaping artifact is extracted memory, not merely a raw log. Delivery is wired; actual host reliance and improvement remain unobserved.); Generic retention/refresh inputs and other integrations, imports and attachment routes: not-determinable (Generic inputs can mix static documents and activity records. Complete origin, automatic admission and later-consumer chains are not established for every included route.)	partial coverage (Authority follows the actual retained part and consumer. Evidence, standing directions, ranking inputs, update inputs and fixed-rule vetoes do not collapse into one force. Unknown downstream paths prevent a complete inventory.); Retrieved facts, observations, pages and source metadata at answer models: known — knowledge [wired] (Internal reflection and external host-context channels remain distinct consumers; delivered evidence is not guaranteed reliance.); Directive text at reflection prompt consumer: known — instruction [wired] (A prompt predicting rejection does not establish a response-level semantic veto.); Fact text, fact/page embeddings and query/candidate representations at retrieval ranking: known — ranking [wired] (Ranking modes vary; a cross-encoder is not universal to recall.); Transcript-derived facts and retained evidence/baseline at automatic memory-update consumers: known — learning [wired] (Learning authority here names input to durable artifact updates, not measured improved capacity.); Stored source IDs, grounding and recalled-target membership at update/withdrawal gates: known — validation [wired], enforcement [wired] (This authority belongs to the consumed lineage fields, not to truth of the associated prose.); Refresh query/trigger/watermark, access scopes and bank configuration at operation selectors: known — routing [wired] (Symbolic controls select operation scope/windows; query text shapes the maintenance input request.); Archive reasons, full audit history, mirror/report later consumers and remaining shipped paths: uninspected (Restoration and audit/report capability are recorded, but complete consequential downstream effects and additional integration consumers are uninspected. Unconsumed fields are not assigned authority.)	AAS-2026-10-14-hindsight-34c2ca821edf-01
HippoRAG	2	Included: imported passage content and metadata; extracted OpenIE rows; normalized entity/fact strings; passage/entity/fact vectors and access structures; source-aware graph; compatibility manifest; prepared-instance query-vector reuse; CacheOpenAI response replay; shipped StandardRAG, offline, MCP, reload and folder-maintenance routes. Alternative vector collections and uninspected provider caches/injected persistence retain their recorded limits. Excluded: static shipped prompts/skills and code as retained memory, external model weights and provider internals, general host state, ordinary current-call QuerySolution/IRCoT state and benchmark scores without a later retained consumer. Shipped fixtures/datasets do not establish trace provenance or observed operation.	partial coverage (The default object inventory is classified; alternative vector collections have recorded vector interfaces but uninspected external database semantics, and other provider caches remain unresolved. Vector-store encoding does not establish an additional database category.); Original passage and normalized entity/fact content rows in default stores: known — files [wired], in-memory [wired] (Default Parquet files reopen into process mappings; these are content parts, not model weights.); Passage, entity and fact embedding arrays and their vector access structures: known — files [wired], vector [wired], in-memory [wired] (Stored encoder outputs differ from encoder parameters; backend durability is not inferred.); OpenIE rows, chunk metadata and compatibility manifest: known — files [wired], in-memory [wired] (JSON files and loaded structures retain extraction, provenance and configuration state.); Source-aware graph: known — files [wired], graph [wired], in-memory [wired] (The pickle and operative igraph structure are both included.); Alternative Qdrant, Chroma and Milvus content/vector collections: partial — vector [wired] (Accepted inventory records vector interfaces and reload caches, but not complete external persistence semantics or all operative backend parts.); CacheOpenAI response table: known — sqlite [wired] (SQLite retains response messages and metadata independently of the index.); Prepared-instance query embedding reuse: known — in-memory [wired] (Memory annotation establishes query_to_embedding reuse and reset after mutation; no durable query-vector file is established.); Other provider response caches and arbitrary injected persistence implementations: uninspected (Other cache substrates are not recorded. Accepted records inspect CacheOpenAI, not equivalent lifecycles for every adapter or injection.)	partial coverage (Automatic derived admission coexists with claimed conditional human approval. Direct caller content admission and metadata replacement stay unresolved; physical file writes do not settle agency.); Initial caller choice/admission of passages and metadata: not-determinable (Generic caller identity does not show whether a human supplies, edits or approves each retained item.); Extraction, structural admission, normalization, encoding, graph and manifest production: known — automatic [wired] (Human workflow initiation is distinct from these software/model admission decisions.); Metadata replacement for repeated content: not-determinable (Replacement content is caller-supplied; no per-content human control fact is recorded.); CacheOpenAI admission and prepared query-vector insertion: known — automatic [wired] (Successful responses and query encodings are retained by code.); Folder workflow conditional batch/refresh approval: known — manual [claimed] (Only the prescribed explicit confirmation/refresh decision is manual; actual host execution remains unobserved.); Other provider response caches and arbitrary injected persistence implementations: uninspected (Admission control for other provider caches is not recorded. Accepted records inspect CacheOpenAI, not equivalent lifecycles for every adapter or injection.)	partial coverage (All established routes are pull. Uninspected adapter alternatives prevent a pull-only whole-boundary conclusion; no independent push is established by automatic selection within a requested return.); Requested retrieval/QA, dense baseline, IRCoT reads and MCP returns: known — pull [wired] (A question or explicit request triggers selection; copying selected context into requested QA is not another unsolicited push.); Requested initialization/indexing/maintenance reload: known — pull [wired] (Reopening an index or requesting maintenance consumes saved state as part of that explicit operation.); CacheOpenAI inference replay and prepared query-vector reuse: known — pull [wired] (A matching inference or query request triggers lookup; automatic hit delivery still fulfils a request.); Other provider response caches and arbitrary injected persistence implementations: uninspected (Read-back direction of alternative provider caches is unresolved. Accepted records inspect CacheOpenAI, not equivalent lifecycles for every adapter or injection.)	not-determinable coverage (No push signal is established. Known pull selectors do not warrant identifier, inferred-embedding or inferred-judgment here; opaque alternative directions prevent axis-level inapplicability.); Assessed requested index/QA/MCP and cache-reuse chains: inapplicable (These are pull operations, so their similarity, model judgment and exact-digest selectors do not classify a push signal.); Other provider response caches and arbitrary injected persistence implementations: uninspected (An unresolved read-back direction cannot be treated as pull-only, so push selection inputs remain unresolved. Accepted records inspect CacheOpenAI, not equivalent lifecycles for every adapter or injection.)	not-determinable coverage (No controlled trace-learning value is established by the accepted records. Unresolved input origins, the absence of canonical bounded-negative records, and alternative writes prevent complete assessment; parameter fixity and lack of measured improvement are not substitutes for trace-route assessment.); Document-derived durable index writes including arbitrary caller content: not-determinable (Automatic extraction/encoding and later consumers are wired, but accepted records do not establish whether arbitrary admitted content contains qualifying activity/experience traces. Document knowledge is not automatically trace learning.); CacheOpenAI raw-response replay: not-determinable (The recorded wrapper retains/replays raw inference output, which alone does not establish a qualifying transformation. No canonical bounded-absence record supports a controlled no finding; neither yes nor no is assigned.); IRCoT thought accumulation and benchmark score reporting: inapplicable (These parts fall outside retained-memory writes: thoughts remain ordinary current-call state and metric results have no recorded retained later-consumer route. They cannot establish a system-wide negative.); Other provider response caches and arbitrary injected persistence implementations: uninspected (Trace-fed transformations and later consumers of alternative caches/injected writes are not established. Accepted records inspect CacheOpenAI, not equivalent lifecycles for every adapter or injection.)	partial coverage (Knowledge, ranking, routing and compatibility vetoes are local consumer effects. Static shipped skills/prompts are excluded as accumulated memory. Neither retained facts nor cached output establishes instruction or learning authority by itself.); Original passage content supplied to extraction and answering consumers: known — knowledge [wired] (Passages supply source material and answer context, not an enforced instruction hierarchy.); Fact strings/vectors, passage vectors and graph at retrieval ranking: known — ranking [wired] (Fixed scoring code can give retained content ranking force without learning a ranking policy.); Entity vectors at synonym-edge construction: known — ranking [wired] (Retained similarity inputs affect which proposed connections receive threshold/top-k admission.); OpenIE rows at rebuild/reuse/deletion and missing-key selection: known — routing [wired] (Saved extraction and source mappings select work and input scope for later maintenance.); Chunk source identifiers at MCP deletion targeting: known — routing [wired] (Current source_id selects matching retained texts; metadata is not an authorization grant.); Manifest compatibility fields at initialization/reuse checks: known — validation [wired], enforcement [wired] (Equality/provenance criteria can veto incompatible state; this is not content truth validation.); CacheOpenAI raw messages replayed to extraction/filter/QA: known — knowledge [wired] (The retained output replaces fresh inference and supplies content to downstream interpretation, not standing instructions.); Response-cache request keys and prepared query-vector access state: known — routing [wired] (Exact matching chooses reuse versus generation; query vectors also influence retrieval ranking.); Passages returned for enclosing-agent answering: known — knowledge [claimed] (Wired server delivery and claimed host interpretation remain separate; host compliance is not upgraded.); Other provider response caches and arbitrary injected persistence implementations: uninspected (Other retained cache parts and their actual downstream effects are not established. Accepted records inspect CacheOpenAI, not equivalent lifecycles for every adapter or injection.)	AAS-2026-10-16-hipporag-2b37edfc7ade-01
Hyalo	2	Retained vault body/frontmatter changed or accumulated through use; derived snapshot entries/graph/BM25 and generated navigation; mutable configuration controlling later commands; optionally caller-retained Jev evidence/policy/results and prescribed trace-informed repairs. Include untraced mutation, OKF/log/changelog and maintenance alternatives as unresolved units. Exclude static shipped/installed instructions and code, current-request state unless captured for reuse, external host conversation/Claude memory persistence and provider model weights/state. External signals may feed included repairs without becoming Hyalo-owned stores.	partial coverage (Files and operative in-memory access products are positive witnesses. Optional helper artifact retention and untraced output inventory prevent complete coverage; external provider weights/state and host transcripts are excluded, not classified as Hyalo memory.); Durable vault body, frontmatter and generated navigation regions: known — files [wired] (File persistence is established for these parts; caller-owned vault is not a repository substrate merely because tidy can consult Git.); Extracted index/graph and BM25 postings, including loaded process copies: known — files [wired], in-memory [wired] (Encoding is symbolic; binary serialization does not imply a database or vector substrate.); Mutable type/schema/view configuration: known — files [wired] (Traced type/view configuration is file-backed; installed static assets are excluded.); Optionally captured Jev manifest parts and classification results: partial — files [afforded] (Optional file capture is established, but actual retention and later reuse are unknown. Request-local process objects alone are not accumulated memory.); Included OKF/log/changelog and remaining maintenance outputs: uninspected (Their Markdown surface is included, but detailed retained output inventory and specialized consumers are uninspected.)	partial coverage (Automatic derivation/admission is supported locally. Initial authorization and physical publication do not establish manual content admission. Caller-assigned and untraced mechanisms retain uncertainty.); Snapshot extraction, patch and persistent publication: known — automatic [wired] (A caller requesting a rebuild does not manually admit each product.); MADR managed-region generation and lint rule-generated metadata repair: known — automatic [wired] (Apply choice is workflow control, distinct from content admission.); Routine instructed tidy repair admission: known — automatic [claimed] (This is claimed model admission under workflow instructions, not an enforced human review gate.); Caller-assigned metadata, removal, configuration and initial body edits: not-determinable (Generic caller identity does not distinguish a human supplying/approving the content from an agent invoking the same command. No manual finding follows from CLI access alone.); Optional Jev artifact capture and initial policy admission: not-determinable (Caller capture is explicit, but caller identity and per-content approval are unresolved; provider proposals do not themselves persist corpus updates.); Remaining mutation/log/release and OKF generator admission mechanisms: uninspected (Detailed handlers are not fully inspected; known software admission elsewhere does not settle these mechanisms.)	partial coverage (Both requested retrieval and independent Pi/context/configuration supply exist. A known pull route cannot resolve optional or uninspected read-back alternatives.); Explicit CLI/Pi corpus and index query/read: known — pull [wired] (Code delivering the requested output is not a second push operation.); Pi session-start aggregate supply: known — push [wired] (The injected aggregate is fresh output from retained inputs, not an extension-owned checkpoint. Host transcript retention is excluded.); Traced generated navigation ordinary note read: known — pull [wired] (No independently scheduled navigation delivery is established.); Mutable configuration and optional saved Jev artifact consumption: partial — push [wired], pull [afforded] (Configuration load is wired; actual saved-helper reuse and all retained artifact consumers remain unresolved.); Remaining log/release, specialized generator and installed-host consumers: uninspected (Specialized later-consumer routes are uninspected. Static installed assets are excluded, but replacement/withdrawal and dynamic retained control coverage is incomplete.)	partial coverage (Coarse push is supported for session orientation and broad configuration loading. No inferred lexical signal is added for BM25 pull. Unknown alternatives prevent complete push-signal coverage or system-wide inapplicability.); Pi session-summary push selection: known — coarse [wired] (Trigger is session start; selector inputs are configured availability, vault metadata/tasks/structure and recent-file modification times; selected parts are aggregate fields and recent identities.); Automatic configuration load for a later command: known — coarse [wired] (The recorded mechanism loads configuration broadly; no recorded identity match targeting separate delivered memory parts supports identifier classification.); Explicit query/read and navigation pull operations: inapplicable (Signal classifies push selection, not requested file identifiers or BM25 query ranking.); Optional helper artifact and remaining push alternatives: not-determinable (Named saved artifacts have requested reuse, but retention and untraced automatic consumers are unresolved. A file identifier alone cannot establish targeted push.)	partial coverage (A claimed yes is local to the instructed automatic trace-informed repair chain. No local no is invented without bounded absence records. Other writes remain unresolved, so neither wired system learning nor a complete negative is supported.); Prescribed trace-informed tidy repair to later corpus consumers: partial — yes [claimed] (Starting repair is authorization, not per-content human admission. This supports the prescribed automatic trace-fed branch, not exercised learning, improved capacity or learned static skills; concrete outputs and retained rationale remain unobserved.); Corpus-derived snapshot and MADR generation: inapplicable (Named transforms derive access/navigation state from documents, not established activity traces. They do not constitute qualifying trace-fed updates in this scoped path.); Pi summary injection and Jev helper computation alone: inapplicable (Summary output is not retained by Hyalo; Jev computation does not automatically persist a durable corpus update. Neither alone is a qualifying trace-fed write.); Initial corpus, caller writes and untraced maintenance/log/release branches: not-determinable (Actual corpus traces, proposal inputs and complete admission/later-consumer chains are unresolved. The inspected positive branch does not settle other writes; no bounded system-wide no is warranted.)	partial coverage (Knowledge, routing, ranking, validation and enforcement have local code witnesses; optional saved rubrics add afforded instruction and tidy adds claimed learning authority. Static installed instructions are excluded from the accumulated-memory boundary. Missing payload/consumer coverage prevents a complete union.); Retained prose at explicit read and Jev evidence consumers: known — knowledge [wired] (Delivery establishes a context route, not reliable compliance or source truth.); Metadata and structural access fields at filters and link traversal: known — routing [wired] (This is code-level selection, not standing model instruction.); Lexical postings/text-derived retrieval features at ranked find: known — ranking [wired] (Fixed BM25 scoring still gives retained features ranking authority.); Mutable schema and view configuration: known — validation [wired], enforcement [wired], routing [wired] (Forces are consumer-local. Optional set validation is not a whole-document or all-writer guarantee.); Generated navigation at later Markdown readers: known — knowledge [wired], routing [wired] (No autonomous special TOC consumer or rationale acceptance is inferred.); Optional retained Jev controls, rubrics and decisions: partial — instruction [afforded], validation [afforded], enforcement [afforded], routing [afforded], knowledge [afforded] (Consumer mechanics are wired, but accumulation through optional capture/reuse is only afforded. Request-local guidance alone is excluded; no actual reuse is observed.); Git/task/activity evidence at instructed tidy update consumer: partial — learning [claimed] (This is learning authority at the trace-fed update consumer; output metadata has separate routing/knowledge paths. It establishes neither improved capacity nor automatic method revision.); Untraced retained payloads and specialized consumers: uninspected (Actual corpus guidance, log/release consumers and remaining rule consumers are not completely inspected.)	AAS-2026-10-14-hyalo-ffe11d6a6d56-01
Hyperagents	2	Through-use retained executable lineage and reconstructed repositories; archive/control metadata; task predictions, grading and evaluation reports; interaction/episode/test artifacts; DGM derived guidance; Genesis generated reward code, saved configurations and learned checkpoint state; their admission, maintenance and later-consumer paths. Includes unresolved bootstrap provenance, checkpoint ancillary parts, environment-context/transfer detail and possible provider retention. Excludes unchanged shipped README/tool instructions, fixed provider weights as such, fresh conversation accumulation and process-local undo/inference tensors; source availability is not evidence of through-use memory.	partial coverage (Local retained carriers are classified, including the dependency-controlled checkpoint file. Opaque provider-side retention prevents complete inventory coverage; model-weights is not inferred merely from parametric content.); Generation patches and reconstructed executable lineage: known — files [wired], repo [wired] (Both persisted patch carriers and their repository reconstruction are included; current subprocess memory is not another retained substrate.); Local metadata, prediction/report/grade files, raw interaction and episode artifacts, reward code and saved configurations: known — files [wired] (The records locate these parts in generated worktrees or output directories, including copied bootstrap files.); Genesis retained policy checkpoint carrier: known — files [wired] (The local file substrate is known independently of opaque checkpoint fields; active inference tensors are current computation, not an additional retained store.); Provider-side retained state, if any: uninspected (Provider weights and hidden retention are opaque; API inference does not determine a controlled substrate for any through-use provider memory.)	partial coverage (Automatic admissions are supported locally. Original bootstrap/checkpoint control and provider writes are unresolved, so neither a manual value nor complete automatic-only coverage is asserted.); Model code/reward/answer proposals admitted by shipped software: known — automatic [wired] (This covers the ordinary proposal/extraction/diff admissions, not unknown replacement of initial files.); Host metadata, logging, prediction checkpoints, reports, configurations and DGM statement admission: known — automatic [wired] (Physical I/O and model authorship are not substituted for admission control.); Genesis rollout training and checkpoint production through the runner: partial — automatic [wired] (The integrated training route is automatic; initialization and any separately supplied/replaced checkpoint admission are not established.); Original bootstrap-file and separately replaced checkpoint admission: not-determinable (Copying a file and selecting a checkpoint do not establish who originally approved or replaced its content; no manual finding follows from operator workflow initiation.); Provider-side retained-state writes: uninspected (Provider inference exposes neither write/admission mechanisms nor human control.)	partial coverage (Both pull and push exist at different operations, often in one chain. Unresolved environment, transfer and provider alternatives prevent complete direction coverage or a pull-only conclusion.); Foreground tool reads of surviving archive files: known — pull [afforded] (Automatic filesystem exposure and requested model-text delivery are separate operations; clipped tool output does not make the requested return push.); Scheduled archive exposure and selected executable-lineage reconstruction: known — push [wired] (The pruned filesystem view is wired; use of its text by a model remains afforded.); Scheduled resume/control/report consumption and prediction/result reuse: known — pull [wired], push [wired] (Reads fulfilling a program request are distinguished from the automatic reuse decision and downstream supply.); Optional ensemble answer requests and wrapper archive supply: known — pull [wired], push [wired] (The score choice and question-ID lookup within an answer request are not counted as extra pushes merely because code returns an answer.); DGM automatic evidence selection and statement delivery: known — pull [wired], push [wired] (Inputs are selected independently of a foreground request for particular retained passages; extraction and model compliance remain distinct.); Genesis checkpoint/configuration read-back: known — pull [wired], push [wired] (A requested checkpoint filename alone is not targeted push selection.); Remaining environment-context, transfer and provider retention routes: uninspected (Full environment-context derivation, detailed transfer consumers and provider retention were not inspected; their direction cannot be inferred from a known local route.)	partial coverage (Coarse availability/performance filtering and actual identity matches are supported locally. Requested file/checkpoint/ensemble lookups are not used as additional push witnesses. Uninspected alternatives prevent complete selector coverage.); Scheduled ordinary archive filesystem push and optional-wrapper archive supply: known — coarse [wired] (Filename pruning and mode availability select broad views, not semantically relevant passages; no archive-wide token budget is established.); Next-parent executable lineage push: known — coarse [wired], identifier [wired] (Candidate priority and the subsequent identity match are separate selectors; a path merely appearing in metadata would not by itself establish the latter.); Automatic completed-answer/result reuse: known — identifier [wired] (The retained match enters the output/skip branch automatically, not a new task-agent request.); DGM failed-case evidence push and reduced-statement delivery: known — coarse [wired], identifier [wired] (The diagnosed proposal is not an inferred-judgment retrieval selector; its new content and extraction are downstream transformations.); Genesis post-training reward-code evaluation handoff: known — coarse [wired] (The forwarded path alone is not an identifier-match signal; checkpoint/config requests are pull operations.); Unresolved environment/transfer/provider push alternatives: uninspected (Direction and selectors are unknown on these included alternatives; inspected pull routes cannot make their signal inapplicable.)	partial coverage (At least DGM and Genesis wire qualifying trace-fed behavior-shaping outputs. Ordinary evidence-shaped editing has only an afforded witness. No bounded-absence record supports a local or system-wide no; raw logs, score reuse and checkpoint selection alone are not learning writes.); DGM failure traces to logged implementation statement and later coding-agent guidance: known — yes [wired] (The qualifying write is behavior-shaping instruction production, not raw-log storage. A malformed retry limits recovery, not the initial successful branch; operation and capacity gain are unobserved.); Genesis rollout-fed retained policy parameters and later inference: partial — yes [wired] (The numerical route qualifies independently of natural-language claim acceptance. Exact optimizer/checkpoint payload and initialization are uninspected.); Ordinary history-informed durable code editing: partial — yes [afforded] (Evidence-responsive ordinary update determination is afforded, not wired by archive copying; no actual historical read or trace-dependent edit is demonstrated.); Full environment-context adaptation, detailed transfer updates and provider hidden writes: uninspected (These included alternatives are not sufficiently inspected to determine qualifying durable automatic writes or their later consumers.)	partial coverage (The inventory separates code execution, control vetoes, score ranking, answer reuse and trace-fed updates. Opaque checkpoint/provider parts and incomplete transfer/environment routes prevent a complete authority inventory.); Inherited executable lineage and reward code at Python/training consumers: partial — instruction [wired] (Instruction includes symbolic program directions here. Descendant code can change tools, admission or selectors; the records do not enumerate every possible descendant effect, so no universal enforcement or validation guarantee is assigned to code.); Metadata flags and lineage IDs at host and editable-selector control consumers: known — enforcement [wired], validation [wired], routing [wired], ranking [wired] (Criteria, vetoes, routing and ranking are separate actual consumers/effects. Archive membership is not certification of success.); Saved numerical reports at score, staged-gate and ensemble consumers: known — ranking [wired], validation [wired], enforcement [wired], routing [wired] (These effects are bounded by mode and fallback; they license neither explanation acceptance nor unseen-split generalization.); Saved answers and completed-result records at reuse, scoring and proof-grading consumers: known — knowledge [wired], routing [wired], enforcement [wired] (The stored answer is consumed directly rather than delivered as a fresh task-model conversation; report comparisons do not turn the answer itself into a validation criterion.); Historical traces/reports at DGM diagnosis and the extracted statement at the next editor: known — learning [wired], knowledge [wired], instruction [wired] (Update-consumer learning authority is separate from downstream instruction authority; neither means improved capacity.); Ordinary archive files at a tool-reading editor: known — knowledge [afforded] (No wired evidence-responsive ordinary code update is inferred from availability.); Genesis saved configuration and learned policy at evaluation: partial — routing [wired], routing [wired] (Configuration routes evaluation setup; retained policy parameters route action choice at the environment consumer. Ancillary checkpoint fields and optimizer consumers remain opaque.); Detailed transfer/analysis consumers and full environment-context derivation: uninspected (Recorded score/plot readers and environment context do not determine every retained-part effect in the uninspected transformations.); Provider hidden-retention authority: uninspected (No through-use retained part or downstream force is inspectable.)	AAS-2026-10-14-hyperagents-dd9d501b160c-01
Instinctual Memory	2	Target-retained adapted journal events; published facts, source references, controls, checkpoint, dispositions, displays/index and prior snapshots; immediate-change intents/receipts; separate tasks; generated project instruction blocks; their acquisition, update, maintenance and later extraction/host/recovery consumers. Excludes transient proposals/ranking results, static shipped skill/prompts, deployment/model-selection configuration, external inference weights/provider retention, and transient evaluator/benchmark reports. External host consumption remains included as afforded delivery, not attributed internal host implementation.	known coverage (The inventory covers the target-retained raw, derived, control, recovery, task and writeback parts. External inference parameters and provider retention are not target memory stores in this scope.); Adapted journal text, metadata and redaction markers: known — files [wired] (Day files are the operative raw input and independently reloaded retrieval store.); Published facts, evidence, controls, checkpoint, dispositions, entity bodies and index; retained prior snapshots: known — repo [wired] (All these parts are stored in published bare-Git snapshots; Markdown/JSON/YAML describe encodings, not extra substrates.); Immediate-change intents and receipts: known — files [wired], repo [wired] (Recovery/replay retain change state; the route supplies the filesystem intent and Git publisher receipt account.); Separate task titles, IDs and version: known — files [wired] (tasks.json is outside the Git memory publisher.); Generated project instruction block: known — files [wired] (The target file is separately written and can be loaded in later host sessions.)	partial coverage (Known automatic mechanisms remain positive. Human content-supplying or approval contributions at generic imports, assertions, tasks and destructive maintenance are unresolved, so a complete agency union is not justified.); Adapter parse/dedup admission, including detached sync and public append: known — automatic [wired] (Caller file selection is not per-content approval; software admits adapted events.); Rules/model consolidation and state/display updates: known — automatic [wired] (No per-fact human approval is required for candidate placement and publication.); Immediate-change structural admission and replay/recovery publication: known — automatic [wired] (The publisher is known, while actual content-supplying human control is separately unresolved.); Content-supplying decision for generic remember/correct/forget caller: not-determinable (Live-agent/caller labels do not establish an explicit human content decision; automatic publisher wiring does not resolve that contribution.); Tidy recomputed suppression admission: known — automatic [wired] (Apply recomputes the selected IDs; initiating apply and supplied keep IDs do not prove per-content approval of its new result.); Human approval of eventual tidy removals or erasure target: not-determinable (The skill requests review/confirmation, but direct executable calls do not bind an immutable approval artifact. Actual per-content human decision remains unresolved.); Erasure current-tree/control and source/history removal machinery: known — automatic [wired] (Software applies the supplied target across sequential stores; success and human target selection are separate issues.); Index/body generation and instruction-block selection/rendering: known — automatic [wired] (Invocation and target choice do not amount to individual row/statement approval.); Task append structural admission: known — automatic [wired] (Version checking is known separately from the unresolved identity of the title-supplying decision.); Per-content human decision supplying task title or initial imported material: not-determinable (Generic callers and source role labels do not establish explicit human approval/authorship; import initiation alone is insufficient.)	known coverage (Requested returns and independent automatic context supply coexist. Wiring establishes hooks/extractor delivery; the instruction-file host load and external inspection remain afforded rather than observed.); Requested search, ID/entity/history reads and CLI source enrichment: known — pull [wired] (These fulfill an explicit consumer query or identifier lookup; delivery is not a second unsolicited supply.); Automatic SessionStart preference delivery: known — push [wired] (Host callback triggers an independent context selector rather than a host request for specified memory.); Automatic UserPromptSubmit fact delivery: known — push [wired] (Prompt text is selection input; the agent did not explicitly request these facts.); Later consolidation reload of journal, facts, controls and checkpoint: known — pull [wired], push [wired] (The updater explicitly loads its retained inputs; it then automatically chooses an extraction context.); Writeback request for retained statements: known — pull [wired] (The command requests a generated block; this requested return is distinct from eventual unsolicited host input supply.); Generated block at later host instruction-file load: known — push [afforded] (Actual host loading is external and unobserved; the retained block affords later standing context.); Task reads and immediate-change recovery/replay state reload: known — pull [wired] (Later operations request the retained task/protocol state they need.); External snapshot inspection of evidence, dispositions and entity/index displays: known — pull [afforded] (Repository/file access affords later requested inspection, not automatic host deployment.)	partial coverage (Known hook and extractor selectors coexist with an external host-load selector whose actual matching/budget behavior is not recorded. A requested target-file identifier does not establish identifier-based host push.); SessionStart preferences push: known — identifier [wired], coarse [wired] (Trigger: SessionStart; selector: active-count availability then pref_user entity identity and a 20-statement cap; retained part: preference statements.); UserPromptSubmit context push: known — inferred-lexical [wired], coarse [wired] (Trigger: eligible prompt callback; selector input: up to 600 prompt characters; retained part: up to six lexical/meaningful-word-overlap fact statements.); Automatic event and known-entity selection for extraction: known — coarse [wired], identifier [wired] (Trigger: batch; inputs: checkpoint sequence, scope/role eligibility and current entities; selected parts: bounded event text and active statement summaries.); Generated block at later host instruction-file supply: not-determinable (Trigger: subsequent instruction-file loading; supplied block is available but host selection internals are not frozen.)	partial coverage (Qualifying automatic session and event-derived fact routes establish yes. Unresolved generic origins remain local uncertainties; no absence record supports a system-wide or local no.); Rules/model user-session extraction into facts and next-update/host context: known — yes [wired] (Automatic trace-fed writes persist standing statements consumed in subsequent extraction, search or hooks; faithful semantics and improvement are not required for wired classification.); Calendar/voice Note extraction into facts: known — yes [wired] (Note framing is not source origin: timestamped experience can feed the same automatic durable fact route.); Opaque prior origin of generic Note/export and direct caller/task content: not-determinable (Generic imported/caller content may derive from activity, but records do not trace every origin and later behavior-shaping automatic write. A system-wide no or complete route inventory cannot be inferred.); Derived instruction-block feedback into later Markdown extraction: not-determinable (The feedback path is wired, but file origin does not distinguish trace-derived statements from static imported rules. The qualifying trace-fed subset cannot be completely identified from Role::Note alone.)	partial coverage (Consumer/effect units retain knowledge, instruction, learning, ranking, routing, validation and enforcement independently. Unconsumed reason/evidence alternatives cannot be upgraded from their readable encoding.); Raw journal text at later answering host: known — knowledge [wired] (Requested search/read delivers raw context, not enforced instructions.); Eligible retained experience at extraction updater: known — learning [wired] (Selected session/calendar/voice text feeds durable standing statements subsequently consumed by extraction or hosts.); Journal text/metadata and fact text at lexical or reranking selectors: known — ranking [wired] (Relative priority depends on retained content under fixed lexical rules or optional model judgments.); Fact statements at host search/read and prompt-context consumer: known — knowledge [wired] (Delivery establishes context capacity, not host compliance or benefit.); Standing preference statements at SessionStart and generated instruction-file consumer: known — instruction [wired] (Preference guidance is pushed through additionalContext; instruction-block authority is conditional on later host loading.); Existing fact statements at next extraction and duplicate check: known — learning [wired], validation [wired], enforcement [wired] (Known-entity summaries guide new proposals; existing text can veto near-duplicate placement.); Fact status/date/visibility and suppression/deletion controls at read/admission gates: known — validation [wired], enforcement [wired] (Eligibility differs by route; writeback does not load Controls and entity reads do not uniformly apply temporal bounds.); Source IDs at changed-file retirement and erasure: known — routing [wired] (References select old-source facts and journal redaction targets, not proof of entailment.); Checkpoint sequence and derived lag at consolidation and hook sync: known — routing [wired] (Checkpoint is progress metadata, not a semantic continuation summary.); Intent/receipt state at recovery and replay: known — routing [wired], validation [wired], enforcement [wired] (Recovery consumes retained changes and base/request identity rather than fact truth.); Task title at later host and task version at append: known — instruction [wired], validation [wired], enforcement [wired] (Title interpretation does not enforce task execution; the expected-version check does enforce append concurrency.); Entity/index display at external inspection: known — knowledge [afforded] (Human or external snapshot inspection is supported as an affordance, not demonstrated deployment.); Source evidence text, suppression reasons and disposition explanations as later automatic guidance: not-determinable (Evidence/reasons are retained, but ordinary delivery omits evidence and no later criterion-revision consumer of reasons/dispositions is established. Their existence cannot establish learning or instruction authority.)	AAS-2026-10-12-instinctual-memory-af48437aa60c-01
jaz-evals	2	ACE accumulated advice, symbolic bullet metadata, update traces and snapshots; JAZ trace-responsive guidance/helpers and explicit history/progress handoffs; smolagents wrapped history and continuation summaries; Letta core blocks as possible editable state, retained conversations, compaction checkpoints and opaque archival/search state; retained outcome/trace diagnostic consumers. Excludes shipped static instructions, unmodified model/encoder weights, benchmark world state, ordinary transient turns without later retained consumption, and fresh-frame/CodeAct eviction alone. Initial provenance and all named opaque alternatives remain included uncertainties.	partial coverage (Positive local classifications coexist with the unresolved included alternatives named below; no complete system inventory or observed operation is asserted.); ACE live advice and symbolic bullet metadata: known — in-memory [wired], files [wired] (Live strings carry updates; text snapshots retain advice and counts.); ACE diagnostic artifacts and retained task verdicts: known — files [wired], in-memory [wired] (Attempt-local files coexist with in-process outcome dictionaries.); JAZ and smolagents explicitly forwarded history, summaries and revised guidance: known — in-memory [afforded] (Requested handoffs retain executor/invocation objects across successor calls; no cross-run reload is established.); Letta host conversation exports: known — files [wired] (Host JSONL exports are diagnostic retention, not restorable server state.); Letta operative core blocks, server conversation, summary checkpoint and archive: not-determinable (External server state and a transient summary return are exposed, but operative retained substrates and replacement persistence are not established. A server DB label does not identify a controlled substrate.); Additional method-specific viewers and publication consumers: uninspected (Their implementations were not audited; further retained parts cannot be enumerated.)	partial coverage (Positive local classifications coexist with the unresolved included alternatives named below; no complete system inventory or observed operation is asserted.); ACE trace-to-ADD admission and optional replacement: known — automatic [wired] (Models propose and software admits without per-content human approval.); ACE seed admission decision: not-determinable (Configured file supply identifies neither human approval nor the controlling principal; the read is not evidence of manual writing.); ACE diagnostics and retained verdict recording: known — automatic [wired] (Code records activity and outcomes without per-record operator approval.); JAZ/smolagents trace collection, progress summaries and guidance revision: known — automatic [afforded] (Model-mediated subsequent content decisions are computational, not human per-change approval.); Letta task-message delivery and host exports: known — automatic [wired] (Software selects new task text or substitutes nudges and exports diagnostics.); Letta initial block approval and subsequent memory/archive/checkpoint admission: not-determinable (Configuration does not reveal per-content approval; server edit and installation controls remain opaque.)	partial coverage (Positive local classifications coexist with the unresolved included alternatives named below; no complete system inventory or observed operation is asserted.); ACE live advice to later solver/reflector/curator prompts: known — push [wired] (Driver automatically supplies all current count-stripped advice on later rounds.); ACE optional snapshot/seed request: known — pull [wired] (Explicit initial_playbook configuration requests file content; loading is not a second automatic targeted push.); JAZ/smolagents successor handoff and requested hit windows: known — pull [afforded] (Foreground model explicitly forwards and searches retained history; returned windows fulfill those requests.); JAZ revised prompt/helpers supplied to later solver calls: known — push [afforded] (Root selects revised guidance for later delegates without those delegates requesting retained material.); Operator requested retained-artifact analysis: known — pull [wired] (Explicit file/attempt/run path requests diagnostic consumption.); Letta conversational recall request: known — pull [afforded] (Prompt affords targeted conversation_search on retained messages.); Letta core-block context supply, summary insertion and archive retrieval alternatives: not-determinable (Exact automatic delivery and request boundaries remain external; a known conversation-search request does not resolve these alternatives.)	partial coverage (Positive local classifications coexist with the unresolved included alternatives named below; no complete system inventory or observed operation is asserted.); ACE automatic advice supply: known — coarse [wired] (Each later prompt triggers inclusion; selector input is current playbook availability, and selected retained part is all advice after count removal, not task-keyed retrieval.); JAZ root selection of revised guidance/helpers for later delegates: known — inferred-judgment [afforded] (Later batch delegation triggers root choice of current prompt/helper material from history-based diagnosis; model judgment selects delivered revisions.); Explicit seed, history search, operator analysis and Letta conversation-search request returns: inapplicable (These operations fulfill explicit requests, not unsolicited push. Substring matching in requested recall is not a push signal.); Letta potential core/summary/archive push selectors: not-determinable (Trigger, selection inputs and selected retained payload are not established; no identifier signal follows from labels alone.)	partial coverage (Positive local classifications coexist with the unresolved included alternatives named below; no complete system inventory or observed operation is asserted.); ACE automatic trajectory/feedback to retained advice to later solver: known — yes [wired] (Code admits trace-conditioned advice, carries live state and snapshots, and supplies later prompts. No observed benefit is required for this wired route.); JAZ trace-responsive guidance/helper update and continuation summaries: known — yes [afforded] (Model-directed writes retain trace-derived guidance or progress for later invocations; actual artifacts and compliance are unavailable.); smolagents trace-derived retained continuation summary: known — yes [afforded] (Requested progress summary from recorded activity enters child context; warning alone does not establish execution.); Letta summary and archive/core update alternatives: not-determinable (Summary production seam lacks established source selection, retained installation and later consumer; archive writes remain opaque. No local negative is inferred.); Diagnostic/viewer paths beyond inspected analysis dispatch: uninspected (Recording raw logs alone does not qualify; unaudited consumers prevent a complete negative about additional trace-fed adaptation.)	partial coverage (Positive local classifications coexist with the unresolved included alternatives named below; no complete system inventory or observed operation is asserted.); ACE advice at subsequent solver and adaptation models: known — knowledge [wired] (Explicitly advisory guidance supplies context, not standing mandatory directions.); ACE retained trajectories and feedback at reflector/curator update consumers: known — learning [wired] (Recorded experience feeds durable advice updates for later solvers; improved capacity is not established.); ACE retained advice bodies at similarity grouping: known — ranking [wired] (Advice text influences similarity-based group selection for refinement.); ACE retained advice at merge model: known — knowledge [wired] (Selected retained bullets supply content for replacement wording.); JAZ histories at root revision consumer and revised guidance/helpers at later solvers: known — learning [afforded], instruction [afforded] (Returned experience supports requested prompt/helper adaptation, and revised prompts direct later solvers.); JAZ/smolagents histories and progress summaries at successors: known — knowledge [afforded] (Forwarded experience supplies context and advice for continuation.); Retained traces and results at post-run analysis/operator: known — knowledge [wired] (Diagnostics expose evidence and reference information; no automatic solver update follows.); ACE stored helpful/harmful counters: inapplicable (Counts are not measured reliability and are removed from ordinary solver/adaptation input; no retained-count authority is established on these assessed paths.); Letta core blocks, searchable conversation, installed summaries and archive at external consumers: not-determinable (Initialization and requested search do not resolve actual context force, archive ranking, editing vetoes or summary injection; no inferred authority from memory labels.)	AAS-2026-10-15-jaz-evals-17ab95f93794-01
JEPA-Anything	2	Includes use-produced design/configuration/assumption and generated-package parts, retained diagnostics, Welford history, mutable projection/EMA module state and maintained checkpoint provenance. Includes their documented alternative writes and later consumers, retaining unresolved initial numerical provenance and future checkpoint-consumer limits. Excludes shipped static Skill/reference methodology, enclosing host/provider internals and session management, ordinary current-call rollout states, and unpublished checkpoint payloads that do not exist in the current entry. Optional report files remain in inventory without an invented later consumer.	partial coverage (The recorded products and numerical state have classified storage. The package and diagnostic containers lack separately declared material parts; no uniform payload or consumer account is inferred from those containers.); Task design files, normalized configuration, generated programs, documentation and copied diagnostics/file list: partial — files [wired] (Filesystem persistence is wired. Missing separate diagnostic, executable-contract and file-list object declarations prevent complete part-level coverage.); OnlineVarianceTracker count, mean and M2: known — in-memory [wired] (Registered buffers retain history across monitor calls; a state_dict example does not establish disk save/reload.); Projection basis and EMA target-encoder parameters: known — model-weights [wired], in-memory [wired] (Local tensor/module storage is specified even though actual values and original training provenance are not.); Checkpoint publication/provenance metadata: known — repo [wired], files [wired] (The current JSON entry is repository metadata, not binary checkpoint storage.); Optional recipe and verifier diagnostic reports: known — files [wired] (Optional writes establish file persistence, not an automatic later reader.)	partial coverage (Known software admissions and explicit editing affordances coexist with unknown initialization control. Neither generic caller identity nor starting the workflow establishes manual admission.); Requested host proposal and diagnostic-driven design repair: known — automatic [claimed] (The host chooses fields/revisions without mandatory per-content human approval.); Explicit human configuration or generated-file editing: known — manual [afforded] (This is a separate editing affordance, not inferred human approval of automatic rendering.); CLI/direct-render package admission and optional diagnostic/report writes: partial — automatic [wired] (Software gates and writes are known; separately declared report/program/file-list parts are missing. Direct rendering does not inherit the CLI validity gate.); Welford update and reset: known — automatic [wired] (Caller flags select invocation; code decides numerical admissibility and replaces buffers.); Subsequent EMA and explicit projection retraction: known — automatic [wired] (The numerical mutation mechanism is known independently of initial weights and external optimizer control.); Initial basis/encoder content and external optimization inputs: not-determinable (Caller-supplied state does not establish whether a human or software admitted the original content.); Checkpoint publication metadata editing and initial admission: partial — manual [afforded] (Explicit maintainer editing is possible; the initial review/approval mechanism is not implemented.)	partial coverage (Concrete requested routes are classified without converting internal return delivery into push. Unestablished semantic reload/report consumers and future checkpoint loading prevent complete coverage.); Explicit config-path load, render and generated-module/test invocation: known — pull [wired] (The caller requests the design/package consumer; no independent relevance selector supplies extra memory.); Caller-selected tracker, projection module and target encoder at subsequent numerical calls: known — pull [wired] (All retained numerical state of the selected module is used; callback return horizons are not accumulated memory.); Checkpoint manifest and local bytes at verifier invocation: known — pull [wired] (This is manifest/file verification, not deserialization or automatic installation.); Future checkpoint consumer described by verify-before-load guidance: known — pull [afforded] (This is an afforded consumer role, not a wired weight loader; the current entry has no payload.); Copied assumptions, retained diagnostic reports and file-list at later host/semantic consumers: not-determinable (No cross-session reload selector or automatic later report consumer is established. Missing part declarations prevent resolving the mixed containers into complete later-consumer units.)	not-determinable coverage (Inspected requested routes do not need a push selector, but unresolved later-consumer alternatives prevent a whole-boundary pull-only or absent-read-back conclusion.); Inspected config/package, numerical-module and manifest verification requests: inapplicable (These are requested pull operations. A config path or module identity on a request is not identifier-based push.); Unestablished host semantic reload and retained-report/file-list consumer alternatives: not-determinable (No trigger, selector input and selected retained part are recorded for a push here. This uncertainty does not establish absent push or any controlled signal.)	partial coverage (The requested diagnostic-repair chain supports a claimed positive, not observed learning or improved capacity. Numerical input provenance and checkpoint persistence are unresolved, so there is no system-wide negative.); Host consumes validation execution feedback, repairs design and emits later-consumed generated contracts: known — yes [claimed] (This is the declared host workflow, not implemented host orchestration. Trace learning here means feedback-fed durable artifact revision, independently of beneficial learning.); Welford historical-statistics route: not-determinable (The fixed update and later loss consumer are wired, but original batch activity provenance and a durable save/reload route are not established. No qualifying trace-fed durable update or bounded negative can be concluded.); Projection/EMA state, initial encoders and external optimization: not-determinable (Original experience/optimizer inputs and durable checkpoint write routes are uninspected. Parameter mutability alone establishes neither qualifying trace learning nor its absence.); Checkpoint metadata and retained audit/package auxiliary outputs: not-determinable (These records do not establish an automatic raw-trace-to-durable-behavior-shaping write and later consumer for every included part. The missing part declarations and future payload limit prevent a blanket negative.)	partial coverage (Generated contracts have concrete fixed-rule consumers. Inspection-only prose/reports, unresolved initial model content and undeclared package/report parts prevent a uniform complete authority account.); Generated usage-mode constants at callback dispatch: known — routing [wired] (The dispatcher consumes persisted constants, not the copied diagnostic report.); Generated contract criteria at invoked tests and audit helpers: known — validation [wired], enforcement [wired] (Force is local to invoked checks; callback execution can bypass them.); Diagnostic feedback at host design repair: partial — knowledge [claimed], learning [claimed] (A validity predicate gates CLI rendering, but no separate retained diagnostic-report object is declared. The gate uses a fresh report, not an old report file.); Welford historical statistics at activity-penalty computation: known — knowledge [wired], learning [afforded] (The next monitor call consumes historical numerical evidence. Actual external optimization remains unobserved.); Checkpoint metadata at integrity/publication consistency verifier: known — validation [wired] (A failed check yields diagnostic failure/status, not revoked loading authority.); Copied assumptions, package file-list, archived reports and generated prose at later semantic consumers: not-determinable (Inspectability and optional persistence do not identify a later semantic consumer. Undeclared diagnostic/file-list parts cannot be assigned the executable constants authority.); Projection and target-encoder retained state at numerical forward calls: not-determinable (Changed tensors influence later calculations, but records do not establish a controlled authority category for each opaque parameter/buffer part at an actual learning or decision consumer.)	AAS-2026-10-15-jepa-anything-434b01ab7774-01
KBLaM	2	Retained Entity/DataPoint content and structure; paired base key/value arrays; trained encoder/model/checkpoint and KB configuration state; exported answers, attention snapshots, reduced metrics and the separate reference/open-ended judgment parts; local authentication access metadata and unresolved SDK/provider retained-state alternatives. Includes their writes, maintenance and later-consumer chains, plus unresolved initial corpus/checkpoint/query-head histories. Excludes shipped static prompts/instructions as accumulated memory, ordinary current-generation KV state, ephemeral training prefix truncation, unshipped hosts and later revisions.	partial coverage (Files, parameter state and operative judgment lists are supported locally. Provider retention and SDK cache internals prevent complete substrate coverage; dense .npy encoding alone is not vector storage.); Retained Entity and DataPoint text/field records: known — files [wired] (Classifies retained disk objects, not their temporary Python dictionaries.); Retained base key and value arrays: known — files [wired] (Keys and values remain distinct parts even though both have the same file substrate.); Trained encoder and Llama/Phi parameter checkpoints: known — files [wired], model-weights [wired] (Known local save representation; externally supplied checkpoint contents and history remain unavailable.); Persisted KB configuration: known — files [wired] (Evaluation constructs its own configuration; exact reuse is not implied.); Attention snapshots, reduced metrics and exported answer traces: known — files [wired] (Current-call text KV caches are excluded; sweep snapshots have a bounded deletion horizon.); Reference-answer judgment part: known — in-memory [wired], files [wired] (The list is an operative result at a later reporting function, not a claim of a subsequent saved-file reload.); Open-ended judgment part: known — in-memory [wired], files [wired] (Normal completion returns None, so completed mean reporting is not established.); Local authentication record: known — files [wired] (Local access metadata is included separately from semantic KB content; cache internals are not established.); SDK credential-cache internal storage: uninspected (The client requests persistence but the SDK implementation and physical cache representation were not inspected.); Included provider-side retained state behind chat/embedding endpoints: uninspected (Provider retention and internals were not inspected; no storage substrate classification of that alternative follows from local client behavior.)	partial coverage (Automatic local admission is supported. No manual finding is inferred merely from operator launch, path configuration, replaceability or checkpoint selection; unresolved initial/external writes remain included.); Entity parsing/append admission: known — automatic [wired] (Launching the script is not manual content approval.); Template, perturbation and extension field admission: known — automatic [wired] (The ordinary generator CLI defect limits reachability, not the implemented admission mechanism.); Embedding and paired-split admission: known — automatic [wired] (No multi-file atomicity or human curation follows from configured paths.); Optimizer updates, periodic checkpoint and config saves: known — automatic [wired] (Explicit checkpoint loading is not a separate evidenced manual write.); Answer/attention/metric and judgment persistence: known — automatic [wired] (Reference post-batch saving and open-ended per-record saving retain their distinct failure limits.); Authentication serialization and sweep cleanup: known — automatic [wired] (User authentication is not approval of generated KB content.); Initial corpus/checkpoint and alternative query-head admission or replacement: not-determinable (Generic caller path selection and possible file editing do not identify an explicit human per-content decision. Manual editing is a capacity, not demonstrated admission control.); Manually runnable notebook scratch deletion: not-determinable (The records identify an operator-run shell cleanup cell but do not establish a per-content retention decision; launching code alone cannot classify manual admission.); Included provider-side retained state behind chat/embedding endpoints: uninspected (Provider retention and internals were not inspected; no write agency classification of that alternative follows from local client behavior.)	partial coverage (Request-driven loading coexists with automatic row delivery and optional internal pruning. Current-generation cache continuation is excluded. Unknown provider routes prevent a complete direction inventory.); Explicit file/checkpoint loads, generator reload and scorer/converter requests: known — pull [wired] (Conditional implemented branches only: unchanged evaluator entry, resume and scorer defects prevent claiming successful execution.); Driver-selected KB and supervision supplied to model/optimizer: known — push [wired] (Upstream operator configuration supplies budgets and paths; requested disk load and subsequent unsolicited model-context supply are different operations.); Llama optional query-conditioned top-k delivery within attention: known — push [wired] (Phi has no corresponding inspected pruning branch; Llama score/value alignment is qualified.); Authentication record reuse by initialized client: known — pull [wired] (Provider retention is separate from local requested reuse.); Included provider-side retained state behind chat/embedding endpoints: uninspected (Provider retention and internals were not inspected; no read-back direction classification of that alternative follows from local client behavior.)	partial coverage (Both coarse driver selection and inferred-embedding Llama selection are supported. Pull-only operations need no signal, but they cannot make the opaque provider alternative inapplicable.); Automatic driver context/supervision selection: known — coarse [wired] (Array indexing implements a coarse/random selection decision, not a semantic identity match; requested filenames alone add no identifier push.); Llama optional sparsification: known — inferred-embedding [wired] (Top-k is inference-time selection, not durable deletion; its score-return slicing limits correctness, not the existence of numerical selection.); Explicit reload/scorer/converter/authentication request operations: inapplicable (These operations fulfill requests rather than supply unsolicited selected context, so a push-selection signal is inapplicable locally.); Included provider-side retained state behind chat/embedding endpoints: uninspected (Provider retention and internals were not inspected; no push-selection signal classification of that alternative follows from local client behavior.)	uninspected coverage (No controlled trace-learning value is established. Local recorded routes do not establish qualifying trace-fed behavior-shaping writes, but the set contains no bounded-absence declarations licensing no. External/provider routes remain uninspected; ordinary supervised adaptation is not denied.); Inspected synthetic/imported dataset-to-embedding/parameter routes: not-determinable (The described local update uses supplied/generated Entity/QA content and expected tokens, not an established trace source. Computing token errors alone does not establish recorded activity input. The accepted set declares no bounded-absence record to license a controlled no, and external input histories remain unresolved.); Inspected answer/attention-to-metric/judgment reporting routes: not-determinable (Activity records produce retained analysis, but no recorded downstream path makes scores/reasons durable answering behavior or admits successor checkpoints. This does not establish yes; without a canonical bounded-absence record it also does not license a controlled no.); Initial imported corpora, external weights and alternative query-head histories: not-determinable (The original inputs to externally supplied artifacts are not authenticated. A known local non-trace route cannot exclude trace-fed upstream production.); Included provider-side retained state behind chat/embedding endpoints: uninspected (Provider retention and internals were not inspected; no trace learning classification of that alternative follows from local client behavior.)	partial coverage (Knowledge, ranking, learning, validation, routing and access enforcement apply to the named consumers only. Parameter machinery and reporting are not assigned blanket authority; no judgment-to-successor selection is established.); Retained Entity properties at QA construction: known — knowledge [wired] (Input reliance is not a factual endorsement.); DataPoint knowledge at ICL and numerical inference: known — knowledge [wired] (No standing-instruction priority or truth guarantee is established.); Retained key text/base key vectors at matching consumers: known — ranking [wired] (Fixed attention code still uses retained features as ranking inputs; pruning alignment is qualified.); Retained base value vectors at aggregation: known — knowledge [wired] (Numerical influence is wired, specific answer dependence is unobserved.); Retained QA targets at supervised optimizer: known — learning [wired] (Learning authority does not itself establish a trace-origin learning route or realized improvement.); References and row labels at answer/attention evaluators: known — validation [wired] (Bounded assessment criteria, not an answer-release veto or independent truth oracle.); Trained encoder and learned query-head parameters at projection/matching: known — ranking [wired] (This describes the scoring consumer, not arbitrary external weights or successful checkpoint reload.); Other retained model/value-projection parameter effects at token computation: not-determinable (The records establish behavior-determining machinery, but do not individuate which controlled knowledge/instruction/other force covers all externally loaded numerical content.); Retained KB configuration at attention/resume: known — routing [wired] (Only actual config consumers are classified; evaluation may construct a different config.); Exported answers and reference/open-ended evidence at scorer: known — knowledge [wired] (Scorer inputs are not automatically accepted propositions.); Judgment score/reason and reduced metric outputs at reporting: not-determinable (Reporting/mean/conversion force is established, but no controlled authority value follows solely from being a report. Saved reasons have no inspected later inference consumer.); Local retained authentication data at service access: partial — enforcement [wired] (Wiring of access is established; actual grants, expiry and external enforcement contracts are not inspected.); Included provider-side retained state behind chat/embedding endpoints: uninspected (Provider retention and internals were not inspected; no behavioral authority classification of that alternative follows from local client behavior.)	AAS-2026-10-15-kblam-5a0c2fe96cc1-01
kgai	2	Included: accumulated decision titles/rationales and symbolic events, Kuzu and reader projections, configured conventions, local approval and autosync control/diagnostic state, their admission/maintenance/later-consumer routes, opaque host-managed task/session context consumed by extraction, and unresolved external host/provider numerical-state alternatives identified by the accepted inventory. External parameter storage, provenance and update routes are uninspected rather than credited to kgai. Excluded: shipped static skill/capture prompts, transient turn markers and pending proposals, UI-compaction bookkeeping, installed executable replacement, and hypothetical export re-import. External host compaction is not an established kgai memory route.	partial coverage (The records cover decision shards, both projections, conventions, approval and sync-control files. The provider-managed numerical state has no inspected physical substrate; this prevents a complete inventory of substrates.); Retained decision text and symbolic event shards: known — files [wired] (NDJSON is an encoding of persisted files, not a separate substrate.); Kuzu projection, including decision text fields: known — graph [wired] (The persisted graph is rebuildable from the authoritative log.); Reader replay projection: known — in-memory [wired] (The shipped human-facing reader independently materializes accumulated events.); Configured conventions and local approval registry: known — files [wired] (The convention text and fingerprint registry are distinct content and control files.); Autosync outcome and cooldown state: known — files [wired] (Local diagnostics are not engineering reasons but affect later status and control.); External host/provider numerical state used by capture: uninspected (The component identifies external provider/host-managed parameters, but not their storage implementation. No model-weights substrate is inferred merely from numerical form.); Host-managed retained task/session context copied into the capture fork: not-determinable (The external host owns copied input and its retained representation/provenance or maintenance. The capture prompt and completed-turn display do not classify the opaque payload.)	partial coverage (Automatic and human-approved decision routes coexist. Generic CLI caller identity, convention admission and external parameter control remain unresolved; initiating a coding task is not per-content approval.); Routine foreground skill/review decision admission: known — automatic [wired] (The engine has no per-decision human-consent check on routine capture.); Explicit decision-command confirmation and human supply/edit: known — manual [afforded] (The explicit command offers a separate per-content operator decision; source does not demonstrate its use.); Optional fork auto mode and unsupported-surface confirm fallback: known — automatic [wired] (Headless confirm follows automatic admission rather than implying a human gate.); Optional fork confirm mode with terminal/desktop UI: known — manual [wired] (Record admits the particular proposal; Skip does not create memory.); Generic direct API/shell decision submissions: not-determinable (An arbitrary caller may be human or software; physical append and author labels do not resolve admission control.); Sync import, envelope construction, projections and autosync state: known — automatic [wired] (Transport approval is not per-reason approval; derived state is constructed by software.); Convention text and local approval admission: not-determinable (The skill prescribes consent but kg trust authenticates no human; authoring a config file and approving its use are separate unresolved controls.); External numerical state initialization or updates: uninspected (No parameter-admission mechanism or provider control is inspected.); Host-managed retained task/session context copied into the capture fork: not-determinable (The external host owns copied input and its retained representation/provenance or maintenance. The capture prompt and completed-turn display do not classify the opaque payload.)	partial coverage (Demand decision/structure reads and automatic convention injection are supported. Requested selection and returned JSON remain pull even when code ranks the answer. Provider-managed memory delivery alternatives remain uninspected.); CLI context/search/history/query/as-of and editor requests: known — pull [wired] (These shipped callers consume accumulated state; sync only changes its availability.); SessionStart convention injection: known — push [wired] (The hook supplies retained convention text without a consumer topic request.); Explicit raw convention reload and trust/status inspection: known — pull [wired] (These returns fulfill an explicit retained-state request; they are not additional pushes.); Local control metadata consumed by automatic session/lifecycle operations: known — pull [wired] (Code loads approval and cooldown state as inputs to its own requested resolution/control operations.); External host/provider retained-context or numerical-state delivery: uninspected (The records do not expose an independently classifiable accumulated-memory request/delivery chain inside the host/provider.)	partial coverage (The assessed unsolicited convention supply is coarse, not topic-sensitive recall. Pull-only units are locally inapplicable; they do not make unresolved provider alternatives inapplicable.); SessionStart effective convention push: known — coarse [wired], identifier [wired] (Trigger is SessionStart; selector inputs are layer precedence, allowed settings and approval; selected retained part is the effective prompt text, capped by Bash string length.); Inspected demand decision/structure/control reads: inapplicable (These operations fulfill requests or code-owned reads rather than independently push selected memory to a consumer; lexical scoring on a pull does not establish a push signal.); External host/provider unsolicited memory selection: uninspected (No trigger, selector input or delivered retained part is exposed for this alternative.)	partial coverage (Automatic extraction can create durable decision context for later consumers without demonstrated improvement. Human-approved and arbitrary-caller alternatives, local control bookkeeping and external parameters do not become system-wide negatives.); Optional auto/headless fork extraction to admitted decisions and later recall: known — yes [wired] (Eligible task experience feeds automatic durable extraction; later CLI/editor consumption is wired. Exact fork context remains opaque but does not erase this write/consumer chain.); Routine foreground capture from coding experience to later reasons: partial — yes [afforded] (The automatic route is provided by model-facing capture policy; actual input/output trace lineage is not supplied for each submitted decision.); Per-content human supply/confirmation routes: inapplicable (These locally scoped writes require an explicit human decision and therefore are not qualifying automatic trace-fed writes. This does not establish an absence on other modes.); Generic submissions, convention initialization and local sync/approval control metadata: not-determinable (Generic source/control is unresolved. Stored diagnostics and cooldown updates are documented, but no recorded extraction/update-to-durable behavior-shaping artifact beyond operational bookkeeping warrants a trace-learning value for them.); External provider numerical-state learning: uninspected (Parameter mutability and learning routes remain uninspected; fixed-weight or no-learning conclusions cannot follow from the absence of exposed update code.)	partial coverage (Authority is classified at each actual consumer. Static capture instructions are excluded; their instruction to respect recalled choices does not convert every reason into an enforcement mechanism. External numerical authority remains unclassified within the memory boundary.); Decision text at later CLI model and editor human consumers: known — knowledge [wired] (Informational delivery is wired; activated coding dependence and benefit remain unobserved.); Retained lexical fields and authority/recency metadata at retrieval selection: known — ranking [wired], routing [wired] (Retained historical text can affect rank even when it is not delivered as a head reason.); Event mutations and projected graph at replay and structural-read consumers: known — knowledge [wired], routing [wired] (Graph authority governs encoded state and recall eligibility, not permission to change source code.); Event identities and parent-chain state at covered integrity admission gates: known — validation [wired], enforcement [wired] (The gate applies to covered engine/object-sync paths, not uniformly to the editor.); Configured conventions at foreground model injection: known — instruction [wired] (Standing directions are supplied, but reliable compliance is unobserved.); Approval registry at project settings loaders: known — validation [wired], enforcement [wired], routing [wired] (Approval state is not authenticated human consent and does not certify convention quality.); Autosync persisted timestamp and outcome at later hook/status consumers: known — routing [wired], knowledge [wired] (Control and diagnostics have distinct effects.); External numerical state at model inference or updates: uninspected (Inference is used by capture, but accepted records do not isolate memory authority or a parameter-update consumer for this opaque component.); Retained task/session experience at the automatic fork extraction consumer: partial — learning [wired] (The update consumer receives opaque host-forked experience and can produce durable reasons; exact included parts and additional forces are unresolved.)	AAS-2026-10-14-kgai-54bcae52e039-01
Kompl	2	Included: converted raw sources; retained extractions, aliases and mentions; draft prose and cross-invocation summaries; current/outbox/previous wiki prose and access metadata; SQLite FTS; Chroma embeddings, document prefixes and metadata; generated or restored schema; chat turns and FAQ candidates; activity/lint/failure records and saved-links projections; accumulated plan/progress/settings only at recorded later consumers; archive and direct replacement alternatives; unresolved provider-hidden continuation/checkpoint state on included calls. Excluded: static shipped code/prompts/instructions, fixed installed/provider model weights, transient current-call values with no later consumer, external MCP agents' own storage/context decisions, and unrelated product-update machinery. Provider-hidden memory is an unresolved alternative, not an asserted retained object.	partial coverage (The recorded local memory inventory has file, SQLite and vector substrates. Hidden provider continuation state is unresolved; static model weights and shipped code are excluded, not classified as accumulated memory.); Converted raw source payloads: known — files [wired] (Gzip encoding does not change the filesystem substrate.); Current and previous wiki prose and page outbox: known — files [wired], sqlite [wired] (The file archive and SQLite pending-content are separate retained copies, both operative.); Extractions, aliases, mentions, FTS, plans/drafts, chat and maintenance/control metadata: known — sqlite [wired] (These are app-owned SQLite objects. Progress and settings are included only where accumulated values drive later retry or overdue checks.); Generated or restored maintenance schema: known — files [wired] (All recorded installation alternatives target a text file.); Dense vectors and their IDs, metadata and document prefixes: known — vector [wired] (Chroma retention is recorded; its internal persistence implementation is not reclassified as an additional database substrate.); Provider-hidden continuation/context state on the included model-call paths, if retained across invocations: not-determinable (The reports do not inspect provider-hidden checkpoints or context. They establish neither this alternative's existence nor its storage substrate; it remains included as an unresolved alternative, not inferred provider memory.)	partial coverage (Automatic admission is wired on several independent mechanisms; explicit human admission is afforded, not observed or universally attributable at the API. Caller-driven restoration/replacement and hidden state remain unresolved.); Conversion/extraction/index/alias/mention writes and automatic ordinary compile admission: known — automatic [wired] (Starting compile or selecting sources is not approval of each resulting retained item.); Schema bootstrap, vector upsert/backfill and saved-links regeneration: known — automatic [wired] (These mechanisms automatically accept nonempty/structurally valid output or execute fixed projection; no per-item semantic approval is recorded.); Chat turn storage and automatic FAQ proposal creation: known — automatic [wired] (Proposal insertion is automatic; later wiki installation is separately constrained.); Deletion-triggered surviving-page redraft and software cascade: known — automatic [wired] (Caller initiation does not make each replacement-prose or cascade admission a human decision.); Shipped pending-draft approve/reject and Save to wiki human-facing decision paths: known — manual [afforded] (The UI affords explicit human content selection or approval, not proof that any API invocation came from a reviewing human.); Actual approval/save caller, archive-supplied contents and direct file replacement: not-determinable (The records identify caller control and fixed guards, not whether a human made the explicit content decision. Actual manual versus software-controlled admission cannot be determined from generic caller identity.); Provider-hidden continuation/context state on the included model-call paths, if retained across invocations: not-determinable (The reports do not inspect provider-hidden checkpoints or context. They establish neither this alternative's existence nor its admission control; it remains included as an unresolved alternative, not inferred provider memory.)	partial coverage (Requested wiki/query returns are pull; automatic schema, historical-turn and prior-draft context augmentation are push operations at the model-consumption boundary. Unknown provider continuation prevents a complete direction inventory; no external unsolicited push is asserted.); Requested wiki/MCP/previous-version/history/diagnostic reads: known — pull [wired] (Exact ID or query lookup fulfills a consumer request; delivery of the response is not a second push.); Question-driven page retrieval and compile/source-ID data loading: known — pull [wired] (The automatic selection inside a requested question response remains pull, as recorded in the memory account.); Automatic standing-schema, dossier and prior-layer/old-page context supply at drafting invocation: known — push [wired] (Separate upstream code reads from downstream model delivery: the drafting model does not request these remembered inputs; application context construction supplies them automatically.); Automatic stored-history supply to answer-model invocation: known — push [wired] (The user question requests an answer, not the selected old turns; code augments the answering invocation with stored role/content history. This is distinct from requested wiki retrieval.); Independent external injection and provider-hidden read-back: not-determinable (No independent unsolicited external-agent delivery was supplied. Hidden provider continuation is uninspected; absence of a deployed push trace is not a bounded absence warrant.)	partial coverage (Signals describe automatic context augmentation, not IDs on requested MCP pages. Query retrieval has recorded model, lexical and embedding ranking but remains a requested pull operation; it is not relabeled as push merely to add signal values. Provider alternatives remain unresolved.); Standing schema push into each later page-draft call: known — coarse [wired] (Trigger: page drafting. Selector input: schema file availability/existence. Selected part: the installed whole convention document; budget clipping does not create targeted relevance selection.); Existing-page context push into update drafting: known — identifier [wired] (Trigger: update plan. Selector input: target page ID. Selected part: that page's retained prose; identity matching is operative, not just an ID in a catalog.); Related earlier-layer draft summaries pushed into later drafting: known — identifier [wired] (Trigger: later-layer task. Selector input: related plan relationships/IDs. Selected part: persisted earlier draft summaries; layer order and availability separately bound supply.); Dossier augmentation at drafting: known — inferred-lexical [wired], coarse [wired] (Trigger: page drafting. Selector input: page relevance against source extraction dossiers. Selected part: capped dossier material, distinct from the separately supplied raw-source list.); Chat-history context push: known — identifier [wired], coarse [wired] (Trigger: next question. Selector input: session equality, ascending timestamp, LIMIT 11 and removal of final returned row. Selected part: role/content of up to ten earliest rows, not citation metadata or a summary.); Signals of provider-hidden continuation/read-back: not-determinable (Provider checkpoint forms, triggers and selectors were not inspected. Known local push selectors cannot establish these alternative signals or make the axis pull-only inapplicable.)	partial coverage (Two local positive routes survive, but ordinary FAQ wiki adoption, alternative caller content, open source origin and hidden provider updates remain unresolved. No system-wide or local no is asserted without bounded absence records; learning here does not assert improved capacity.); Saved-links automatic failure-trace projection for later wiki/query consumers: known — yes [wired] (Automatic trace-fed reduction creates a durable context artifact with later readable/retrievable consumers. No novel knowledge, parameter update or observed benefit is required.); Automatic chat-answer FAQ candidate and exceptional wiki-adoption paths: partial — yes [wired] (Automatic durable proposal construction and later review are wired; ordinary disjoint page-ID provenance blocks wiki adoption. Only exceptional coincident IDs permit that approval path, not a universal ordinary compounding cycle.); Raw chat/lint/activity retention and their possible further updates: not-determinable (Raw logs and ordinary history replay alone do not establish the qualifying durable derived update. The reports establish diagnostics/replay but no automatic lint-to-new-guidance route; this is not a bounded negative for every possible trace-fed update.); Caller-selected answer saves and direct/archive replacement: not-determinable (Supplied content need not be a stored answer; human versus automatic caller admission and original input provenance are unresolved. Manual save cannot be universally counted as automatic trace learning.); Ordinary source-derived pages/schema and possible trace-valued source inputs: not-determinable (Normal sources are recorded as imported knowledge, not agent-action traces. The actual open-ended source content and schema metadata origin do not establish a qualifying automatic trace chain for every included input.); Provider-hidden continuation/context state on the included model-call paths, if retained across invocations: not-determinable (The reports do not inspect provider-hidden checkpoints or context. They establish neither this alternative's existence nor its trace-fed update and later consumer; it remains included as an unresolved alternative, not inferred provider memory.)	partial coverage (Authority is assigned at named consumers: context, conventions, ranking, routing, validation, enforcement and trace-fed updates coexist. No finding claims observed compliance, epistemic acceptance or improved capacity; imported/hidden alternative consumers remain unresolved.); Raw sources, extraction prose, existing pages, earlier draft summaries and chat history at later drafting/answering consumers: known — knowledge [wired] (This records supplied context/reference, not truth acceptance or model compliance.); Wiki and prior-version prose at explicit human/MCP readers: known — knowledge [wired] (External agents receive requested tool results; their subsequent decisions remain outside the boundary.); Standing schema at later drafting model: known — instruction [wired] (The instruction request is wired; it does not enforce deterministic planning and actual compliance is unobserved.); FTS text, Chroma dense features and page title/summary/source-count/recency at query candidate selection: known — ranking [wired] (Fixed scoring gives retained inputs ranking force even though no retained learned ranking policy is established.); Aliases, mention counts, plan/session IDs, settings and progress at subsequent resolution/planning/retry/startup consumers: known — routing [wired] (Only state with a recorded later operational consumer is included, not all database metadata.); Pending-plan status and source/provenance dependencies at approval and cleanup: known — validation [wired], enforcement [wired] (Fixed pending-state and dependency predicates have both criterion and veto consequences. These are operational, not factual acceptance criteria.); Archived-source status at default page-index selection: known — enforcement [wired] (Default suppression is enforced; direct-ID prohibition and semantic removal are not established.); Saved-links failure records at projection and derived page at later wiki/query consumers: known — learning [wired], knowledge [wired] (Update-consumer force and downstream reference force are distinct.); Recorded answer at automatic FAQ proposal update and pending review consumer: partial — learning [wired], knowledge [wired] (The durable proposal route is wired. Ordinary page-based adoption fails; exceptional empty/coincident-ID admission must not upgrade it to a complete ordinary wiki cycle.); Triage/lint findings and reasons at operator/sidebar display: known — knowledge [wired] (Advisory diagnostic force is distinct from the planner consuming the immediate triage result.); Imported vector documents and unconsumed metadata beyond recorded readers: not-determinable (The records establish access/display/backup roles but not every actual consumer or behavioral force of restored payloads; no authority is inferred for unconsumed metadata.); Provider-hidden continuation/context state on the included model-call paths, if retained across invocations: not-determinable (The reports do not inspect provider-hidden checkpoints or context. They establish neither this alternative's existence nor its consumer and effect; it remains included as an unresolved alternative, not inferred provider memory.)	AAS-2026-10-14-kompl-9cffb1e5eb86-01
LACP	2	Accumulated or revised memory: native keyed JSON and saved conversations; Stop-derived handoffs, SMS episodes/self-model/narrative/epochs; research staging/inbox/registry/graph; daily captures and adopted durable text; hybrid chunk/sparse/dense index; captured learning events and retained promotion controls; workflow transition state; improvement source/history; their access, admission, maintenance and later-consumer routes. Includes unresolved session/article/repository/code synchronization, research consolidation, gap/link/review generation, MCP-memory rendering, quarantine/threshold maintenance, ingestion delegates, revised installed capabilities and external-host checkpoint alternatives. Excludes unchanged shipped instructions/scaffolds, ordinary transient current-turn state, unrelated runtime storage, and provider inference weights not established as accumulated or changed memory. External memory internals remain unresolved rather than assigned the local adapter properties.	partial coverage (Known file-backed, graph, vector-access, in-memory and repository parts coexist with uninspected offline/external alternatives; this is not a complete substrate union.); Native keyed entry files: known — files [wired] (Per-key JSON files are the operative store; a key namespace does not establish a separate KV service.); Native conversation across save and resume: known — in-memory [wired], files [wired] (Messages reside in RAM during operation and in saved JSONL for later resume.); Handoff JSON: known — files [wired] (Latest per-cwd handoff and hook-contract JSON persist on disk.); SMS episodes: known — files [wired] (Append-only episode JSONL is retained under SMS_ROOT.); SMS self-model: known — files [wired] (The operative pattern/preference object is self-model.json.); SMS narrative: known — files [wired] (Narrative has a known JSON-file substrate despite unresolved admission.); SMS epochs: known — files [wired] (Epochs are appended to a distinct JSONL file, not the narrative file.); Research staging, inbox, registry and graph notes: known — files [wired], graph [wired] (Staging JSONL, inbox Markdown and registry JSON are files; materialized linked graph Markdown supplies a graph access structure, not a graph database claim.); Daily captures, suggestion reports and adopted durable bullets: known — files [wired] (Daily and durable Markdown and suggestion reports are retained files.); Hybrid retained chunk index: known — files [wired], vector [wired] (JSON retains chunks, sparse weights and optional dense vectors; vector denotes the vector retrieval structure, not a separately established vector database.); Cross-CLI captured events and promotion state: known — files [wired] (Canonical event JSONL, promotion-state JSON and receipts persist on the filesystem.); Workflow progression state: known — files [wired] (Persisted state.json and transition history govern later advance/status invocations.); Improvement source, logs and result history: known — files [wired], repo [wired] (Changed source and Git commits are retained in the repository; logs and TSV results persist as files. Static shipped program content alone is not accumulated memory.); Included alternative offline and external memory paths: uninspected (Session-history mirroring, article/repository/code synchronization, research consolidation, gap/link and review-queue generation, generated MCP memory, quarantine hygiene, threshold recalibration, ingestion delegates and deployed external MCP/QMD memory remain incompletely inspected. Their retained outputs, transformations, admission and later consumers cannot be completely classified from the accepted records. These are included alternatives, not excluded stores or evidenced absences.); External-host retained conversation or compaction state: not-determinable (Host persistence and checkpoint internals are outside the inspected adapter; returned text saved in native messages does not classify the host store.)	partial coverage (Automatic admission is supported on specific routes. No manual value is inferred from apply flags, caller identity, role names or manual-gate labels; unresolved initial/replacement and caller-controlled mechanisms remain included.); Foreground model keyed-memory admission: known — automatic [wired] (Model-selected content is written after nonempty checks without per-content human approval.); Native conversation save: known — automatic [wired] (Software saves the accumulated message list on its save paths; writing user turns is not a separate explicit human approval of the saved whole.); Stop handoff, episode and self-model admission: known — automatic [wired] (Stop capture/heuristics select retained content before later quality judgment, without a per-content human gate.); Narrative replacement admission: not-determinable (Library physical replacement does not identify the decision maker; no inspected constructing/admitting caller establishes manual or automatic admission.); Initial self-model fields and direct replacement: not-determinable (Heuristic subsequent updates are automatic, but initial preference/bias content and separately supplied replacements have unresolved control.); Epoch derivation and append: known — automatic [wired] (Caller starts synthesis; software selects episodes and copied decisions, not a human deciding each retained epoch content.); Trace staging and research synchronization: known — automatic [wired] (Heuristics, graph admission and apply-enabled publication select content; starting apply is not per-candidate human approval.); Daily capture and optional durable-bullet apply: known — automatic [wired] (Software chooses and appends eligible bullets when enabled; pipeline suggestions without apply are a separate non-adopting invocation.); Hybrid index admission: known — automatic [wired] (Builder admits current eligible filesystem corpus and derived features without per-document operator approval.); Rename, prune, cap and archive selection: known — automatic [wired] (Software selects changes/discards after enablement; driver/parser mismatch limits actual archive-driver activation, not helper selection agency.); Direct resolution status/reason admission: not-determinable (Caller supplies the actual status/reason; the records require no authenticated human and do not establish whether a human or agent decides the supplied content.); Cross-CLI caller capture and supplied promotion evidence: not-determinable (Generic caller provides task/outcome/evidence/names. Code checks form/gates, but manual-labelled inputs are not proof of a human decision.); Workflow supplied decisions and note content: not-determinable (Actor strings and approve/reject values do not identify human versus computational admission of the recorded decision.); Idle delegated source edits and numeric keep/discard: known — automatic [wired] (User enables experiments but model/host proposes edits and code decides score disposition without per-edit human approval.); Installed revised skill/config admission: not-determinable (Configuration and skill files can be reloaded, but their revision writer and factory admission are unresolved.); Included alternative offline and external memory paths: uninspected (Session-history mirroring, article/repository/code synchronization, research consolidation, gap/link and review-queue generation, generated MCP memory, quarantine hygiene, threshold recalibration, ingestion delegates and deployed external MCP/QMD memory remain incompletely inspected. Their retained outputs, transformations, admission and later consumers cannot be completely classified from the accepted records. These are included alternatives, not excluded stores or evidenced absences.)	partial coverage (Both requested reloads and unsolicited host recovery delivery exist. Query return is not push merely because a selector ranks results; unknown external/offline delivery alternatives prevent complete direction coverage.); Requested native key/read/search/list: known — pull [wired] (Foreground consumer explicitly asks for retained material; returning the result is pull, not an additional push.); Explicit native conversation resume: known — pull [wired] (Startup resume or /resume reloads the selected saved conversation at the consumer request. No automatic unrelated memory injection is inferred.); Host-event recovery and SMS supply: known — push [wired] (Startup/clear/compact automatically selects eligible retained parts and supplies systemMessage without a content request from the receiving host.); Explicit SMS/handoff/epoch/vault display: known — pull [wired] (CLI caller requests retained records or search results; display is pull even when software selects matching results.); Requested hybrid query and evaluator loads: known — pull [wired] (CLI/probe/benchmark requests load retained payload and return selected chunks. Incoming read-only consumption counts despite no persisted new query output.); Research sync, daily suggestions and maintenance input reload: known — pull [wired] (Invoked processing consumers explicitly load their retained registry/daily/probe inputs; later foreground delivery is not automatically wired by this fact.); Cross-CLI query and evaluation: known — pull [wired] (Requested advisory retrieval and evaluator reload consume retained events. Neither establishes unsolicited foreground hint injection.); Workflow and promotion reload: known — pull [wired] (Advance/status/re-promotion/rollback requests consult persisted state and controls.); Improvement history and source reload: known — pull [wired] (Experiment setup reads saved history; subsequent process import loads retained source. Static program loading alone is outside accumulated memory.); Revised installed skill/config and external-host checkpoint delivery: not-determinable (On-demand skill loading is known but whether a revised through-use artifact reaches a later invocation, and how opaque host checkpoints are supplied, remain unresolved.); Included alternative offline and external memory paths: uninspected (Session-history mirroring, article/repository/code synchronization, research consolidation, gap/link and review-queue generation, generated MCP memory, quarantine hygiene, threshold recalibration, ingestion delegates and deployed external MCP/QMD memory remain incompletely inspected. Their retained outputs, transformations, admission and later consumers cannot be completely classified from the accepted records. These are included alternatives, not excluded stores or evidenced absences.)	partial coverage (Known automatic host supply uses identity plus coarse gates. Hybrid lexical/embedding query ranking is retained under authority, not misclassified as an unsolicited push signal; other push alternatives remain unresolved.); Host handoff push: cwd identity match: known — identifier [wired] (Trigger is host startup/clear/compact. Selector hashes current cwd and reads its matching latest handoff; the matched retained handoff supplies recovery text. This actual match, not merely a printed ID, establishes identifier selection.); Host handoff push: age and priority budget: known — coarse [wired] (Same event uses under-24h availability and priority/budget inclusion for the selected handoff. Coarse filtering remains distinct from its cwd identity match.); SMS arc push: known — coarse [wired] (Host event selects nonempty current_arc and truncates to 200 characters; selection is availability/length, not inferred goal relevance.); SMS self-model push: known — coarse [wired] (Host event uses confidence >0.4 and fixed field/list slices to supply preferences, biases and failures; no semantic or identifier-based project targeting is established.); SMS episode push: known — coarse [wired] (Host event supplies last three episodes from seven-day, decayed-significance>=0.6 selection, then priority budget. Agent/project arguments are not applied as identity filters.); Inspected explicit pull-only query/display/control routes: inapplicable (These particular consumers request retained data; push-signal classification is inapplicable locally. Lexical/dense ranking on requested hybrid queries is not a push signal.); Unresolved external-host checkpoint and revised-extension push: not-determinable (Known pull routes do not establish the signal of opaque or unresolved alternatives.); Included alternative offline and external memory paths: uninspected (Session-history mirroring, article/repository/code synchronization, research consolidation, gap/link and review-queue generation, generated MCP memory, quarantine hygiene, threshold recalibration, ingestion delegates and deployed external MCP/QMD memory remain incompletely inspected. Their retained outputs, transformations, admission and later consumers cannot be completely classified from the accepted records. These are included alternatives, not excluded stores or evidenced absences.)	partial coverage (Positive wired trace-fed routes coexist with unresolved admissions, raw-log-only routes and opaque alternatives. No global no, observed improvement or parameter training is inferred; no local no is invented without bounded absence records.); Stop-derived recovery handoff to subsequent host: known — yes [wired] (Automatic capture compresses session activity into durable recovery text subsequently supplied as host context. It qualifies without new knowledge or observed improvement.); SMS episode capture to later host context: known — yes [wired] (Automatic trace-fed summary capture persists episodes and later host-event selection delivers them as context.); SMS self-model update to selected later guidance: known — yes [wired] (Recorded summary/outcome signals automatically revise durable preference/failure fields consumed in later host guidance. Count-based confidence is not a truth warrant.); Assistant/history-to-research-to-index-to-caller: known — yes [wired] (Implemented extraction/admission yields durable graph/reference material with later query consumers. Hook registration/activation remains unresolved, while the implemented qualifying chain remains wired.); Daily trace extraction and optional durable promotion: known — yes [wired] (Automatic extraction and enabled software append produce durable context/reference artifacts reachable by later suggestion/sync/index consumers; the default pipeline does not enable append.); Epoch grouping to CLI display: not-determinable (Automatic trace-derived grouping and persistent display are established; behavior-shaping guidance beyond display is not established. Do not transfer the episode host consumer to epochs.); Narrative initial/replacement route: not-determinable (Startup arc consumption is known, but a trace-fed automatic admission path is not; the epoch writer does not establish one.); Keyed entries and initial self-model content: not-determinable (Model/caller admission can be automatic, but supplied arbitrary text/initial fields lack a demonstrated original trace input and qualifying derivation chain.); Raw native conversation retention: not-determinable (Raw save/resume is not by itself a demonstrated trace-derived behavior-shaping update; these records do not establish an exhaustive bounded absence of all qualifying native updates.); Cross-CLI events and promotion: not-determinable (Captured summaries/outcomes and later hints/evaluation are established, but generic caller identity leaves automatic trace-fed content admission unresolved. Mode promotion is not by itself a content-learning route.); Idle source changes and count history: not-determinable (Score-selected source revision and recurrence are wired, but retained logs/scores are not systematically fed into guidance content; original qualifying trace input to edits is not established.); External host checkpoint generation: not-determinable (Host compaction/summary implementation is opaque; local adapter returns do not establish trace-fed durable checkpoint generation.); Included alternative offline and external memory paths: uninspected (Session-history mirroring, article/repository/code synchronization, research consolidation, gap/link and review-queue generation, generated MCP memory, quarantine hygiene, threshold recalibration, ingestion delegates and deployed external MCP/QMD memory remain incompletely inspected. Their retained outputs, transformations, admission and later consumers cannot be completely classified from the accepted records. These are included alternatives, not excluded stores or evidenced absences.)	partial coverage (Authority belongs to the named consumed part and effect. Retained text can rank before supplying knowledge; trace inputs can supply learning at an update consumer without proving improvement or downstream enforcement.); Keyed content returned to foreground model: known — knowledge [wired] (Actual tool-result delivery supplies retained reference/context; action changes remain unobserved.); Saved messages restored as later provider conversation: known — knowledge [wired] (Resumed content provides conversational knowledge/context, not an independent acceptance criterion.); Handoff and SMS episode/arc text at later host: known — knowledge [wired] (Selected recovery summaries, episodes and current arc supply interpreted context through systemMessage.); Selected SMS preferences and caution at host: known — instruction [wired] (Preferred approaches, biases and failures supply standing behavioral directions; retained success labels are not all delivered.); Episode and session signals at self-model/recovery update: known — learning [wired] (Retained experience feeds durable pattern/preference and recovery-artifact updates; learning authority at update is distinct from instruction/context at host.); Retained self-model confidence and episode significance at startup selector: known — ranking [wired] (Confidence gates field delivery and significance/recency prioritize episode eligibility; these retained fields affect selection priority rather than prove truth.); Retained handoff cwd/time and SMS confidence at delivery routing: known — routing [wired] (Cwd identity and age select the recovery artifact; confidence selects whether a self-model context part is eligible.); Epochs at CLI display: known — knowledge [wired] (Later CLI exposes copied episode decisions as reference for caller interpretation; no startup narrative consumer is established.); Research text, counts and access features at admission/publication: known — ranking [wired] (Counts, confidence and retained text affect selection/priority of registry candidates and graph publication.); Research blocked IDs at ingestion: known — validation [wired], enforcement [wired] (A retained blocked-ID criterion rejects matching candidate admission: it both validates admissibility and enforces a veto, not factual truth.); History/daily text at durable extraction and promotion: known — learning [wired] (Retained historical inputs feed durable guidance/reference updates even without observed improvement.); Hybrid retained text and sparse features at ranking: known — ranking [wired] (Chunk text and access features affect relative retrieval score under fixed code; no retained learned ranking policy is required.); Hybrid retained dense vectors at numerical ranking: known — ranking [wired] (The separate dense part influences similarity priority when enabled; fallback does not erase its supported optional path.); Selected hybrid chunks at CLI and evaluator: known — knowledge [wired] (Returned retained text supplies reference/context; benchmark/probe interpretation is wired, foreground task-model integration only afforded.); Retained probe results and registry features at forgetting: known — ranking [wired], validation [wired] (Probe gap/importance and registry strength/edges protect or prioritize archival candidates. They constrain maintenance rather than establish source truth.); Cross-CLI retained event fields at advisory ranking: known — ranking [wired] (Summary tokens, trust, status/score and recency affect relative priority and eligibility of hints.); Selected event summaries at advisory caller: known — knowledge [wired] (Printed retained summaries offer advice/reference; query is explicitly non-routing and does not read enforce-state as activation.); Retained event labels at metric gate: known — validation [wired] (Reloaded event labels yield metric predicates used by promotion admission. This is validation for mode transition, not acceptance of each assertion.); Promotion state at re-promotion and rollback: known — routing [wired], enforcement [wired] (Saved mode/active-promotion state selects later status, re-promotion refusal or rollback progression; downstream host enforcement remains unresolved.); Workflow retained stage/token at next advance: known — routing [wired], validation [wired], enforcement [wired] (Stored stage/owner/token controls which transition is admissible and can refuse advancement; history notes do not verify completed work.); Improvement TSV count at experiment scheduler: known — routing [wired] (Retained history count selects whether further experiments are eligible, not content-directed criticism.); Changed source at later imports: not-determinable (Executable authority is established, but the exact changed functions/effects are not individuated; execution is not itself one controlled authority value. Static program instructions are excluded unless revised through use.); All retained SMS fields and initial narrative at other consumers: not-determinable (Full CLI displays and startup arc/preference subsets differ. Unconsumed success/effective-context metadata does not acquire startup authority by association.); Included alternative offline and external memory paths: uninspected (Session-history mirroring, article/repository/code synchronization, research consolidation, gap/link and review-queue generation, generated MCP memory, quarantine hygiene, threshold recalibration, ingestion delegates and deployed external MCP/QMD memory remain incompletely inspected. Their retained outputs, transformations, admission and later consumers cannot be completely classified from the accepted records. These are included alternatives, not excluded stores or evidenced absences.)	AAS-2026-10-14-lacp-91efd16f7a35-01
Letta Code	2	Retained core/deferred MemFS bodies and access metadata; agent and attached skill content changed or accumulated through use; raw local session history, derived compaction summaries and their checkpoint/access structures; reflection payload/progress, arena disposition records and mutable memory policy; their admission, maintenance and later consumers; opaque hosted agent/conversation and attached non-skill memory alternatives. Excludes shipped static prompts/bundled skills, arbitrary uncaptured extension code, package replacement and the separate executable mod-learning builder. Optional mirrors and feedback uploads have no established later memory consumer; their storage alone is not read-back. Provider inference weights are not an established accumulated-memory surface; external memory update internals remain unresolved.	partial coverage (Local retained files, Git access structures and resident transcript projections are classified; external retained state has no established substrate. Optional mirrors have no recorded later consumer and are not additional operative memory units.); Local core/deferred MemFS and agent or attached skill files: known — files [wired], repo [wired] (Bodies and their access paths are retained in files and Git checkouts; project/global acquisition is uncaptured but does not obscure the recorded file substrate.); Raw local transcript, retained summary and active cached projection: known — files [wired], in-memory [wired] (JSONL persistence and resident message/index caches are both recorded; this does not classify the uninspected replay algorithm.); Reflection payload, manifest and successful-progress checkpoints: known — files [wired] (Local slice files and locked progress updates are the inspected retained parts.); Mutable MemFS policy and arena disposition records: known — files [wired], repo [wired] (Tracked JSON policy/Git metadata and local run JSON/choice JSONL are retained; no feedback-training substrate is inferred.); Hosted agent/conversation and non-skill attached-memory payloads: not-determinable (The client contract and readable mounts do not reveal external persistence; no service-object, database, vector or weight substrate can be established.)	partial coverage (Automatic trace-fed and hook admission paths are established. Human-facing arena selection is an affordance, not evidence authenticating generic callers; initialization, restore and external controls remain unresolved.); Compaction summary retention after model return: known — automatic [wired] (The backend trims/clips and persists returned text without a per-content human decision.); Default reflection proposal integration and agent-mediated explicit integration: known — automatic [wired] (Default clean proposals merge mechanically; explicit integration asks another agent not to wait for the user.); Hook-covered model edit admission and successful-progress update: known — automatic [wired] (This unit names inspected software gates and completion-state updates, not all physical file writers.); Offered human reflection-arena winner/tie selection: known — manual [afforded] (The user-facing branch-selection interface affords explicit approval of selected content; actual chooser identity is not established.); Generic arena choice caller and listener/create-body or backup-restore content control: not-determinable (Choice values, listener requests and restore force flags do not authenticate an explicit human per-content decision; generic caller identity leaves admission agency unresolved.); Mutable constraint overrides, initial/dynamic skill admissions and generic-worker/repair finalization: uninspected (An environment authorization flag is not proof of human control; complete source routes for these alternatives are not recorded.); Hosted and attached repository content admission: not-determinable (Service acceptance and content authorship/control are external or unresolved; sync software alone does not establish who admits each memory assertion.)	partial coverage (Both automatic context supply and requested retrieval exist locally. Operations within a chain are separated: fulfilling a Memory/Skill/search request is pull, not an additional push because code transports its result. Hosted alternatives remain unresolved.); Core bodies/directory pointers and active local history or summary to subsequent turn: known — push [wired] (A new turn automatically selects and supplies retained context without a request for those specific retained parts.); Deferred Memory path, Skill name and transcript/conversation search requests: known — pull [wired] (The consumer/caller requests retained content; queued skill text is fulfillment of that request.); Available retained-skill metadata to later model discovery context: known — push [wired] (Eligible names/descriptions are supplied before the model requests full bodies; procedure loading is a separate pull unit.); Reflection evidence/current memory and successful-progress state at later maintenance invocation: known — push [wired] (Trigger and selector automatically supply transcript slices/checkpoints; upstream manual launch does not make later automatic selection a request for each item.); Hosted non-skill projection, full replay and repair/arena feedback later-consumer alternatives: not-determinable (External delivery is opaque and complete replay/repair/feedback routes are uninspected; local pushes and pulls cannot close those alternatives.)	partial coverage (Signals classify automatic selection only. Requested file paths, skill names and search terms are not identifier/lexical push merely because they select pull results. Hosted selectors and complete maintenance selection remain unresolved.); Turn-triggered committed core/index projection: known — coarse [wired] (Trigger is a local agent/conversation turn; selector inputs are MemFS enablement, committed HEAD and Markdown layout; delivered parts are root bodies and directory pointers.); Agent/conversation identity selects retained history and active summary projection: known — identifier [wired] (Trigger is the next turn; agent/conversation identity selects its ordered retained history and compaction projection, rather than identifying a requested file only.); Automatic eligible-skill metadata advertisement: known — coarse [wired] (Trigger is skill discovery for turn context; source/attachment eligibility and model-invocable filtering select metadata, not full procedure text.); Reflection progress-based transcript slice selection: known — identifier [wired], coarse [wired] (Trigger is reflection launch; reflected-through message ID and per-conversation state identify the next recorded slice, with replay boundaries/thresholds separately limiting coverage.); Complete multi-conversation/reflection normalization and selection machinery: uninspected (Automatic conversation selection is not fully inspected; supplied reasons/priorities do not establish a complete inferred-judgment selection algorithm.); Hosted memory push selectors: not-determinable (Server-side selection and embedding mechanisms are opaque; visible metadata and hosted vector/hybrid API options do not classify actual push signals.)	partial coverage (Compaction and reflection establish wired automatic trace-fed durable outputs with later context consumers. This is not observed improvement or parametric training. No bounded absence records support a system-wide or local no.); Automatic summary production, retention and later continuation: known — yes [wired] (The summary persists as a compaction entry and is later consumed as model context; manual launch does not supply per-summary human admission.); Automatic reflection proposal/integration to future memory or skill consumers: known — yes [wired] (The route can admit transcript-fed edits for core/deferred/skill use; no-change outcomes and unobserved semantic quality do not negate the wired write path.); Initial/caller-authored content and dynamic skills: not-determinable (Provenance and per-content control are unresolved; retaining imported/authored text alone establishes neither qualifying automatic trace-fed learning nor bounded absence.); Generic-worker repairs, arena feedback consumers and migration alternatives: uninspected (Worker update inputs and later feedback use are not fully traced; the recorded choice log is not proof of automatic learning.); Hosted maintenance and provider-retained learning alternatives: not-determinable (External update internals and later consumers are not captured; a wired local yes cannot resolve external learning coverage.)	partial coverage (Authority follows each retained part at its actual consumer. Local context, learning inputs, ranking inputs, policy vetoes and maintenance selectors have distinct warrants; opaque server and feedback consumers prevent completeness.); Core/deferred factual prose at later turn-model context consumers: known — knowledge [wired] (Compiler and tool delivery are wired; model obedience and factual endorsement are not.); Retained preference/direction prose and loaded skill procedure at later turn-model consumers: known — instruction [wired] (This unit names standing directions/procedures, not every fact in the same files.); Retained raw session and previous summary at compaction/reflection update consumers: known — learning [wired] (Traces feed durable context or memory/skill updates, independently of downstream knowledge/instruction authority.); Raw history and retained compaction prose at later turn-model consumer: known — knowledge [wired] (The selected history/summary provides context; no reliable imperative content or enforcement is inferred from a user-role wrapper.); Stored transcript text and conversation summary at local search selector: known — ranking [wired] (Query/term-match selection and ranked returned results establish a ranking input; complete ranking code remains unaudited.); Mutable memory constraints at hook-covered commit admission: known — validation [wired], enforcement [wired] (Budgets/layout/read-only criteria have a program consumer; hook bypass and restore limit the enforcement boundary.); Reflection successful-progress IDs and counters at later maintenance selector: known — routing [wired] (Progress is consumed symbolically rather than as model advice.); Attachment membership at shared-skill discovery: known — routing [wired] (This unit covers local eligibility, not the external attachment enforcement contract.); Arena choice log feedback consumer: uninspected (Choice controls current integration, but no later learning or selection consumer of its persisted log is established.); Hosted non-skill consumers and generic-worker repair/update effects: not-determinable (Server consumers are opaque and precise repair behavior is uninspected; no global authority inventory follows.)	AAS-2026-10-12-letta-code-be76afeab9b7-01
lhtb	2	Use-accumulated or revised Terminus-2 handoff/overflow context, provider conversation chains, retained workspace products, trajectories, outcomes/configuration, checkpoint snapshots, optional saved task-quality assessments, trial/job analysis caches, recovery/sweep selectors, viewer/export/upload routes and afforded adopted runtime replacements. Includes unresolved SDK/alternate-agent retention and remote downstream consumers. Excludes static shipped task/skill/MCP/prompt/rubric definitions as memory objects, fixed model weights not updated by inspected routes, task-local RAG/memory internals as general harness capabilities, independent Harbor training applications and external deployment internals beyond their exposed retention boundary. Initial provenance/control and remaining task-product details stay unresolved rather than being inferred from an inspected example.	partial coverage (Files and process retention are resolved locally; the service boundary is known without its internals. SDK, alternate agents and remote platform internals prevent complete coverage.); Terminus-2 derived handoff text in local Chat and associated retained logs: known — in-memory [wired], files [wired] (Covers local handoff state and its logged representation, not the remote chain.); Host trajectory, checkpoint, saved result/configuration and trial/job analysis files: known — files [wired] (File-backed objects are separately identified by the accepted inventory.); Task workspace retained between commands/phases and optional saved task-quality output: known — files [wired] (No task-local database or memory-store substrate is inferred from the workspace container.); Terminus Responses conversation chain behind previous_response_id: partial — service-object [wired] (Service-object describes the exposed retention boundary; underlying representation, storage and lifetime remain opaque.); Analysis SDK context, alternate-agent retention and remote upload platform internals: uninspected (SDK/CLI state and remote persistence implementation are not inspected; no concrete backend database or vector substrate is licensed.); Version-specific replacement runtime module when maintained and adopted: known — files [afforded] (This is a separate adoption path, not static shipped source counted as accumulated memory.)	partial coverage (Known automatic admission survives unresolved initialization and maintenance control. The reports do not support conflating caller initiation with manual curation.); Handoff and fallback generated-content admission: known — automatic [wired] (Launching a task is not manual admission of later handoff text.); Workspace edits and host trajectory/result/checkpoint recording: known — automatic [wired] (Model authorship and physical file writes are not separate human admission decisions.); Trial/job analysis and optional task-check result admission or overwrite: known — automatic [wired] (Structural validation does not make the admission manual.); Configured trial-end streaming upload admission: known — automatic [wired] (The hook establishes this admission mechanism independently of generic explicit uploader callers.); Initial content, explicit export/upload caller control and installed runtime replacement admission: not-determinable (Generic caller or maintenance-actor identity does not establish an explicit human content decision. No manual finding is licensed.); Provider/SDK and alternate-agent internal writes: uninspected (Internal admission and mutation policies remain uninspected.)	partial coverage (Pull and push coexist at distinct request/selection operations. Opaque alternatives and retained objects with no established later reader prevent complete direction coverage.); Automatic retained context selection and supply on the handoff/recovery path: known — push [wired] (Full replacement and prompts submitted through unwound Chat remain distinct forms of recovery.); Model-selected later file reads of workspace or analysis evidence: known — pull [wired] (The route is wired; actual analysis file choices and complete evidence reading are unobserved.); Job recovery, outcome-driven sweep selection and cache-to-aggregate construction: known — push [wired] (An upstream request to run/resume/analyze does not itself request every automatically selected retained part.); Explicit viewer/export/upload access versus configured streaming upload: known — pull [wired], push [wired] (Requested export traversal is a suboperation of fulfilling that request, not independently counted as unsolicited delivery.); Version-specific installation followed by a later runtime import: known — pull [afforded] (The maintenance replacement decision is not reclassified as a read.); Hidden checkpoint file reload, optional task-check output reader, provider/SDK and alternate-agent later consumption, remote platform consumers: uninspected (Retention or access paths alone do not establish an additional later read-back operation for these branches.)	partial coverage (Positive selectors are attached only to actual push operations, with triggers and selected parts. Unknown alternatives prevent complete signal coverage.); Handoff/recovery automatic selection of retained history and generated continuation payload: known — coarse [wired], inferred-judgment [wired] (Trigger, code selection and model content choice are distinct; no identifier or embedding targeting is established.); Automatic completed-result recovery selection: known — identifier [wired] (This is an actual retained-result identity comparison, not an identifier on a requested file alone.); Automatic sweep successor selection from saved outcomes: known — identifier [wired] (The >0 reward condition accompanies identity selection; singular reward-field compatibility remains unresolved.); Automatic cache/field supply to aggregation: known — coarse [wired] (Trial names in aggregation input are not evidence of identity-targeted semantic retrieval.); Configured END-hook upload of a completed trial: known — coarse [wired] (No identity match selects particular delivered content; duplicate-upload veto is not an identifier-based retrieval witness.); Explicit file/viewer/export requests and later replacement-module import: inapplicable (These named operations are pull; their paths and continuation references do not independently establish a push signal.); Provider/SDK and alternate-agent selection, remote downstream supply and unestablished checkpoint/task-check readers: uninspected (Unknown direction/selector facts cannot be converted to inapplicability from the known pull routes.)	partial coverage (Two positive automatic trace-fed derivation chains survive unresolved routes. There is no system-wide or local no finding without a bounded absence warrant; the union retains yes without asserting measured learning.); Automatic handoff and generated short-summary continuation write to later foreground model: known — yes [wired] (The profile definition includes this within-run durable handoff without requiring new knowledge, weight change or improved performance.); Automatic persisted trial analysis and aggregate narrative chain: known — yes [wired] (This is an interpretation route, not demonstrated improvement of the task solver or a training pipeline.); Optional task-quality assessment from static inputs: inapplicable (This named route consumes task files/rubric rather than recorded activity; no qualifying trace-fed update is established by its static-input scope.); Raw trajectory/outcome/checkpoint retention and exports as further learning routes: not-determinable (Raw recording, replay/scheduling and packaging alone do not identify another durable behavior-shaping update. No bounded absence record licenses a local no.); Arbitrary workspace products, adopted maintenance changes, provider/SDK, alternate agents and remote downstream updates: uninspected (Generic edits/imports and opaque internal retention do not resolve whether these branches implement qualifying trace-fed updates.)	partial coverage (The inventory separates update-consumer learning from downstream knowledge/instruction and scheduler control. Opaque and unestablished consumers remain included; no ranking finding follows merely from numerical rewards.); Interaction history at automatic handoff update consumer: known — learning [wired] (Learning authority at the update consumer does not assert improved capacity.); Handoff findings and continuation directions at later solver: known — knowledge [wired], instruction [wired] (Guidance is interpretive, not a symbolic compliance guarantee.); Saved configurations/outcomes at job recovery scheduler: known — routing [wired], validation [wired], enforcement [wired] (These forces belong to retained scheduler references, not a model answer oracle.); Saved result reward at sweep successor selection: known — routing [wired] (Code-level route is wired; singular reward-key compatibility and actual filtering remain unresolved.); Trial evidence at the automatic analysis update consumer: known — learning [wired] (This update-consumer force does not establish improved capacity.); Retained analysis summaries/check explanations at the aggregate model and operator consumers: known — knowledge [wired] (No scored reward withdrawal, solver adoption or epistemic certification follows.); Analysis cache state at later analyzer: known — routing [wired] (The static schema is not counted as retained learned validation policy.); Trajectory narrative at the file-inspecting model/operator: known — knowledge [wired] (This interpretive evidence path is distinct from structural export traversal.); Retained trajectory continuation references at the exporter: known — routing [wired] (No training authority is inferred from export formatting.); Inspected task product at later verifier execution: not-determinable (The records establish execution of retained workload source and checking its outputs, but do not establish that this retained part supplies a permission veto, selection rule, criterion or other controlled authority at that consumer. Executable form alone does not establish enforcement.); Adopted replacement module at future Harbor control consumer: known — enforcement [afforded], routing [afforded] (This authority is afforded after installation, not observed deployment.); Hidden checkpoint file reload, optional saved task-quality output, opaque provider/SDK and alternate-agent consumers, remote platform consumers: uninspected (Checkpoint live-return control is not file read-back. These later retained-payload consumers are unresolved, preventing authority classifications for them.)	AAS-2026-10-15-lhtb-a0548ba4edbf-01
link	2	Includes accumulated memory claims/procedures and their controls/context; wiki knowledge and links; raw/imported capture/proposal material; dismissal history; handoffs; wiki/FTS/vector access state; usage metadata; ingest ownership manifests; operation journals/checkpoints, Git/team copies and retained exports; their admission, maintenance and later consumers, including uncharacterized digest/auxiliary alternatives. Excludes unchanged shipped instructions, runtime code, external host state/compaction, provider parameters and model caches not updated through Link use, temporary evaluation corpora and unretained current-query packets. Known vector derivation retains unresolved external encoding provenance.	partial coverage (The named content, control, cache and checkpoint stores are classified. Detailed digest and auxiliary retained outputs are not inventoried by accepted records; no whole-runtime storage union is asserted.); Memory claim, lifecycle/provenance fields, retrieval context and non-memory wiki pages: known — files [wired] (Markdown content and frontmatter share a filesystem store but remain distinct operative parts.); Raw capture notes/proposals, dismissal history and handoff packets: known — files [wired] (Markdown captures/handoffs and JSON dismissal history are explicit retained objects.); Wiki index, backlinks, persistent parsed text/FTS and process maps: known — files [wired], sqlite [wired], in-memory [wired] (The record names each cache substrate; cached payload is not another authored corpus.); Derived semantic vectors and usage event ring: known — files [wired] (Both are explicit local JSON stores.); Structured ingest source/output ownership manifest: known — files [wired] (Ownership hashes persist under .link-ingest and are consumed by later apply.); Operation markers, pre-write snapshots and backup archives: known — files [wired] (Recovery retains filesystem checkpoints with distinct metadata and byte payloads.); Git-synchronized workspace and team copies: known — repo [wired], files [wired] (Git history is operative retained recovery/synchronization state, distinct from the Markdown copies.); Exported HTML, manifest and retained audit JSON: known — files [wired] (The durable exports are distinct from a live permalink, which does not create another retained store.); Detailed digest and auxiliary output inventory: uninspected (The memory coverage account leaves digest and auxiliary internals uninspected; their complete retained-object inventory cannot be classified.)	partial coverage (Automatic software/model admission is not made manual by workflow initiation. Explicit human decisions are supported at the protocol layer; generic caller identity, file adoption and replacement control remain unresolved locally.); Automatic session hook capture, index/vector refresh, usage recording and journal snapshot production: known — automatic [wired] (Software admits these generated records without a per-content human decision.); Project seed content production and software Git/team reconciliation: known — automatic [wired] (Caller initiation/options do not approve each generated/imported passage.); Structured rendered-page admission after caller apply and staged code gate: known — automatic [wired] (Applying a workflow is not a per-claim human approval; code admits generated staged content on its finite conditions.); Agent-authored wiki ingest within workflow authorization: known — automatic [claimed] (Shipped procedure delegates claim extraction/authoring to the external agent rather than requiring per-claim assent.); Human-supplied/approved remember, selected capture, review and edit protocol: known — manual [claimed] (The documented human decision qualifies manual at this layer only; core calls do not authenticate its occurrence.); Actual ordinary writer, capture acceptance, lifecycle and handoff caller control: not-determinable (Generic callable caller identity and confirmation flags do not establish a human decision or its absence for each admitted part.); Dismissal/cleanup, direct Markdown adoption and explicit restore replacement control: not-determinable (Software I/O and caller selection/confirm alone do not identify whether a human supplied, approved or replaced this retained content.); Detailed digest and auxiliary admission mechanisms: uninspected (Admission control for these included alternatives is uninspected.)	partial coverage (Explicit request and independent automatic supply are distinct operations. Known push/pull routes do not complete the uncharacterized auxiliary route inventory.); CLI/MCP/HTTP focused and generic recall, explain, full context and direct requested files: known — pull [wired] (The consumer requests retained material; ranking and delivery inside that request do not add a second push.); Inbox/acceptance, maintenance, sync and explicit recovery consumers: known — pull [wired] (Requested operations consume prior notes, controls, checkpoints and files.); Startup hooks and first suitable MCP response autobrief: known — push [wired] (Independent memory/handoff/review supply is not an explicit request for these retained parts.); Per-prompt saved-constraint reminder: known — push [wired] (Prompt submission triggers an automatic constraint selector, not an explicit recall request.); Wiki/vector selection state and retained cooldown/dismissal controls at later software invocation: known — push [wired] (Code loads access/control state automatically to carry out a later request or host event; this is distinct from the requested content return.); Documented offline export and raw-import reader routes: known — pull [afforded] (Their reader role is afforded; actual later external consumption is unobserved.); Detailed digest and auxiliary later-consumer routes: uninspected (The accepted inventory does not characterize these routes completely.)	partial coverage (Signals describe actual push selection, not query/file identifiers on requested returns. Startup eligibility/budget and project matches, guard lexical overlap and control/index supply remain distinct; auxiliary pushes are unresolved.); Startup/autobrief selection of memory, capture-review and newest handoff substance: known — coarse [wired], identifier [wired] (Trigger is session start or first eligible JSON response. Active/review/recency/newest-pending eligibility and clipping determine the brief.); Prompt-guard retained constraint selection: known — inferred-lexical [wired], identifier [wired], coarse [wired] (Trigger is per-prompt hook/query wrapper; input is current prompt/project plus saved constraint text, with event cooldown.); Automatic supply of cached indexes/vectors and control ledgers to software consumers: known — coarse [wired] (The later selector loads the available cache/history; hashes/model keys and ring/window eligibility govern reuse. Query similarity is downstream ranking, not selection of this cache supply.); Explicit requested recall, recovery, acceptance and offline content returns: inapplicable (These are pull operations; a request identifier or ranked requested response is not a push signal.); Detailed digest and auxiliary push-selection routes: uninspected (Uncharacterized alternatives prevent a complete signal inventory; known pull routes cannot make them inapplicable.)	partial coverage (Automatic trace extraction and metadata-feedback writes have durable later-consumer paths. This is the profile write-route sense of learning, not demonstrated improved capacity. Generic caller/seed/history and auxiliary alternatives prevent a system-wide negative or complete positive inventory.); Automatic session-end retained notes/proposals to later review context and accepted-memory consumers: known — yes [wired] (The hook automatically extracts traces into durable notes/candidates; later startup review supply consumes these and selected accepted claims feed future recall. Approval of final memory is a separate uncertain mechanism.); Automatic usage events to later guard repetition decisions: known — yes [wired] (Returned identity events are retained automatically and later cooldown changes emit/silence decisions; popularity ranking is a separate optional path.); Automatic operation journal/status and snapshot production to later recovery: known — yes [wired] (Operation activity automatically produces durable recovery records; later confirmed recovery consumes their status/path information. Restored byte content alone is not a learned claim.); Dismissal history and caller-authored accepted claims/handoffs: not-determinable (The retained rejection experience shapes proposal vetoes, but actual per-content human control and mixed caller origins are unresolved. These paths are not upgraded by the automatic hook witness.); Conditional semantic-cache derivation from trace-origin versus non-trace-origin memory text: not-determinable (Automatic encoding and later ranking are known; the accepted records do not identify the original trace-fed provenance of every included vector input.); Optional Git-history project seed and imported/manual note alternatives: not-determinable (The routes include mixed static/history/session inputs, but records do not fully establish every original trace kind and qualifying automatic durable update.); Detailed digest, auxiliary maintenance and checkpoint alternatives: uninspected (Uncharacterized transformations and replacement inputs prevent a bounded complete absence of qualifying trace-fed writes.)	partial coverage (Findings name actual consumers/effects, including ranking by retained text and controls. External compliance/benefit is not observed; unknown auxiliary consumers prevent complete inventory coverage.); Memory claims and wiki content at later answering agents/human readers: known — knowledge [wired] (Focused recall and independent generic wiki selection supply context, not truth acceptance.); Retained procedures, handoff next steps and saved-constraint reminders at host agents: known — instruction [wired] (Natural-language force is distinct from mechanical permission controls.); Claim/title/TLDR/body/context, dense vectors, graph/search features and reviewed metadata at rankers: known — ranking [wired] (Ranking includes retained content scored by fixed code; no learned ranking policy is required.); Lifecycle/project fields and dismissal membership at focused selection/proposal admission: known — enforcement [wired], validation [wired] (These controls can veto their covered code paths; generic wiki delivery bypasses focused lifecycle gates.); Ownership manifests at later structured apply: known — validation [wired], enforcement [wired] (Retained source/output hashes determine managed replacement versus conflict.); Usage event names/times at guard cooldown and optional salience consumer: known — routing [wired] (Live cooldown and optional popularity ranking are distinct consumers.); Usage event counts at optional library salience ranking: known — ranking [afforded] (The optional boost can be supplied to ranking; inspected ordinary adapters do not wire it.); Capture notes at accept-time re-mining and retained-memory production: known — learning [wired], knowledge [wired] (Retained conversational experience feeds extraction into durable future-context claims; caller acceptance remains separately controlled.); Journal markers and snapshot manifests at explicit recovery: known — routing [wired] (Recovery uses retained paths/status to choose restoration work; it does not treat markers as agent factual knowledge.); Exported summaries/articles at documented offline readers: known — knowledge [afforded] (Production of files is wired; the later external reader role is documented rather than deployed.); Detailed digest and auxiliary retained-part consumers: uninspected (Unknown payload and consumer paths prevent a complete authority inventory.)	AAS-2026-10-14-link-595f6f9cba36-01
llm-wiki	2	Human-admitted raw originals; maintained source, entity/concept, synthesis, contradiction and answer pages; changed catalog, page access metadata and operation history; instructions and templates only as changed through use; their write, maintenance and later-consumer routes, including optional opaque qmd retained index state. Excludes unchanged shipped scaffolds, transient computed graphs/reports and watcher suppression, external host context/compaction and provider model weights.	partial coverage (Files are supported across the retained project inventory. The included alternative qmd store prevents complete substrate coverage. Optional Git initialization alone does not establish retained memory in a repository.); source summary prose: known — files [wired] (Retained Markdown file; implemented corpus enumeration and reload expose these files, while semantic production remains prescribed.); entity and concept prose: known — files [wired] (Retained Markdown file; implemented corpus enumeration and reload expose these files, while semantic production remains prescribed.); revised synthesis prose: known — files [wired] (Retained Markdown file; implemented corpus enumeration and reload expose these files, while semantic production remains prescribed.); contradiction claims and resolution prose: known — files [wired] (Retained Markdown file; implemented corpus enumeration and reload expose these files, while semantic production remains prescribed.); filed answer prose: known — files [wired] (Retained Markdown file; implemented corpus enumeration and reload expose these files, while semantic production remains prescribed.); maintained catalog prose: known — files [wired] (Retained Markdown file; implemented corpus enumeration and reload expose these files, while semantic production remains prescribed.); appended operation history prose: known — files [wired] (Retained Markdown file; implemented corpus enumeration and reload expose these files, while semantic production remains prescribed.); human-curated raw originals: known — files [claimed] (Original documents persist in project raw storage under the prescribed acquisition protocol.); modified maintainer instructions and project templates: known — files [wired] (Sync and template consumption use filesystem text; only versions changed through use are included.); optional qmd retained collection and numerical/index state: uninspected (External storage implementation is not recorded; a vector or database substrate cannot be inferred from embedding setup.)	partial coverage (Human raw curation and prescribed model admission have separate warrants. Filing, generic scaffold callers, guidance changes and external index admission retain control uncertainty.); human admission of raw originals: known — manual [claimed] (An explicit human source-curation decision supplies the retained originals.); agent semantic admission during ingest and prescribed immediate lint corrections: known — automatic [claimed] (Agent determines individual writes; an initiating human request or discussion is not per-content approval. Batch ingest supplies a countercase to universal discussion.); optional answer filing admission: not-determinable (An offer to file is recorded, but no explicit acceptance handshake establishes whether the human or agent admits each retained answer.); page creation and forced replacement admission: not-determinable (Generic caller chooses title/type/force; caller identity does not establish a human decision supplying or replacing retained content.); changed instruction and template admission: not-determinable (Human co-evolution is requested, but final per-change control and template editing decisions are unspecified. Blind copying is physical propagation, not proof of admission agency.); qmd collection admission and subsequent refresh: not-determinable (Manual setup is afforded, but records do not establish control of all retained content or refresh writes.)	partial coverage (Tool and protocol file-request routes support pull. Host loading of changed guidance remains unresolved. Watch notifications expose pending filenames, not a wired unsolicited knowledge-context delivery; no push is inferred from requested search returns.); question-driven catalog/search and page/section expansion: known — pull [wired] (Consumer/agent requests retained material; ranked selection and code delivery fulfill that request rather than create an unsolicited push.); requested metadata, log, status and lint inspection: known — pull [wired] (Later inspection requests read current accumulated content or associations and return selected data.); requested qmd alternative search: known — pull [wired] (Inspected adapter invokes qmd for a consumer query and returns adapted results; internal state and freshness remain opaque.); direct agent file reads prescribed for later maintenance: known — pull [afforded] (The external maintainer is instructed to read affected retained pages.); later loading of changed instructions and templates by external host: uninspected (Later guidance consumption is afforded, but host loading trigger and request-versus-unsolicited delivery are not established.)	uninspected coverage (No positive push selector is established. Pull-only tool units are inapplicable locally; unresolved guidance loading prevents axis-wide inapplicability.); fully inspected requested CLI/MCP knowledge and inspection returns: inapplicable (These are pull operations; identifiers and BM25 relevance in a requested response do not establish push signals.); external host loading of changed guidance: uninspected (Without an established direction, trigger or selector input, neither a push signal nor pull-only inapplicability can be assigned.)	not-determinable coverage (No accepted record establishes a qualifying automatic trace-fed write or a complete bounded absence across the included update inventory. Host session compaction and provider weight updates are excluded, not classified as no.); semantic ingest, answer filing and authored operation history: not-determinable (Recorded inputs are documents and questions plus selected knowledge, not an established automatic trace-to-durable-update chain. An authored history alone cannot establish trace learning or a bounded system-wide no.); instruction convention changes and synchronization: not-determinable (No automatic log-to-guidance update is recorded; absence of a wired connection in this account is not a dedicated bounded absence warrant for every qualifying update.); optional qmd retained updates: uninspected (External update inputs and later consumers are not inspected; embedding setup alone establishes no qualifying trace-fed write.)	partial coverage (Knowledge and instruction concern different external consumers; ranking, structural validation and notification routing concern inspected code. Opaque qmd effects prevent a complete authority inventory. No automatic log-fed learning consumer is established.); retained subject-matter prose at later answering and maintenance agents: known — knowledge [afforded] (Requested later reload supplies evidence, interpretation, prior analysis and discovery advice; compliance and benefit remain unobserved.); changed maintainer guidance at external maintainer: known — instruction [afforded] (Reloaded changed conventions provide standing directions, conditional on host loading.); retained prose and frontmatter at built-in BM25 scoring: known — ranking [wired] (Current file text and boosted headings determine relative retrieval scores under fixed lexical code.); raw_file, sources, paths and wikilinks at structural inspectors: known — validation [wired] (Retained associations and references supply criteria for structural conformity, not source truth or pre-write semantic acceptance.); source association metadata at coverage classifier and pending notification selector: known — routing [wired] (The retained raw-file mapping changes pending/ingested classification and which filenames are notified. Process-local suppression state is excluded.); qmd retained state at external ranking and selection consumers: uninspected (The adapter returns scored results, but the consumed retained parts and their internal effects are not inspected.)	AAS-2026-10-14-llm-wiki-b437987c5802-01
llm-wiki	2	Retained raw session renderings; derived source/entity/concept/overview and saved-answer wiki content; navigation, graph, quality, mtime and deferred-task metadata; candidate archives/reasons; queue and special-memory alternatives; approved method/steering/skill revisions; reader/search/export products and their write, maintenance and later-consumer routes. Includes unresolved initialization, special-cache wiring and unread repair/vault/editor variants. Excludes static shipped instructions as accumulated memory, transient unsaved lint reports, native private stores, external provider weights and host windows.	partial coverage (Known retained objects use files. Uninspected included variants prevent claiming complete substrate coverage. External provider weights and live host windows are excluded, not additional accumulated memory stores.); Retained raw sessions and derived wiki page bodies/frontmatter: known — files [wired] (Covers the recorded raw and wiki products, not provider parameters.); Navigation/folder context, quality fields, candidate archives/reasons and JSON work state: known — files [wired] (Includes initialized scaffolds and the stored queue even though their proposed automatic use is unresolved.); Retained graph JSON and reader/search/export products: known — files [wired] (Graph encoding does not establish a graph storage service; transient NetworkX structures are not a separate accumulated store.); Retained method/steering/skill revisions: known — files [afforded] (The instruction-mediated route is afforded, not an observed deployment.); Included vault/editor retained variants and repair-produced representations: uninspected (The supplied memory coverage leaves overlay/write-helper and repair variants uninspected; their entire retained-object inventory cannot be classified.)	partial coverage (Automatic and manual admissions coexist. Candidate caller identity and initialization remain unresolved; maintenance status disagreement is retained as distinct warranted findings, not silently upgraded.); Converter admission of selected raw files and mtime state: known — automatic [wired] (Scope is converter admission, not authorship of native session material.); Backend page/state/index writes and deferred prompt/completion admission: known — automatic [wired] (Foreground initiation does not make the code/model admission manual.); Routine foreground semantic ingest and requested page-update admission: partial — automatic [afforded], automatic [claimed] (Keep both maintenance evidence bases: reconciliation leaves afforded versus claimed unresolved. No single stronger implementation assessment is selected.); Saving substantial query answers, saving/editing after reflection, and self-learn method revisions: known — manual [afforded], manual [claimed] (Starting the workflow alone is not this approval; the explicit later content decision establishes manual admission.); Graph export and regenerated reader/publication admission: known — automatic [wired] (Opting into the command is not approval of each resulting content item.); Candidate creation and caller-selected promotion/merge/discard: not-determinable (Transition code is wired, but generic CLI caller identity does not establish a human decision or autonomous admission. Candidate initialization is separately unresolved.); Customized navigation, quality-field updates, queue callers, special caches and unread repair/vault variants: not-determinable (Who supplies or approves each retained update is unresolved or uninspected; known automatic page writes cannot decide these admissions.)	partial coverage (Requested read-back is established on several independent routes. Unresolved automatic surfaces prevent a completely assessed pull-only boundary; no supported push finding is invented.); Host question/page requests, MCP query/read/search/quality/export and CLI graph query: known — pull [wired], pull [afforded] (Automatic ranking within a requested return is not a separate push operation.); Synthesis/ingest and deferred-task pickup on explicit invocation: known — pull [wired], pull [afforded] (The backend does not launch a worker or schedule unsolicited delivery; UUID association is not evidence of push.); Reader/export consumption after requested publication: known — pull [wired] (No deployment or actual later reading is observed.); SessionStart/queue, hot-refresh and special-memory injection: not-determinable (Missing hook and unresolved automatic selectors/loading prevent establishing the direction of proposed delivery. Ordinary readability is not proof of push.)	not-determinable coverage (No push-signal controlled value is established. Pull-only assessed units are locally inapplicable, but unresolved proposed push alternatives prevent making the axis inapplicable or absent.); Completely assessed requested query/read/graph/quality and deferred pickup operations: inapplicable (These individual operations fulfill consumer requests, so their lexical, identity or judgment selectors are not push signals.); Proposed SessionStart/queue and hot/special-file delivery: not-determinable (No automatic push selector, actual selected retained parts or channel is established. Special filenames and queue paths alone cannot support identifier; scaffold availability cannot support coarse.)	partial coverage (Qualifying automatic trace-fed writes support yes locally. Explicitly human-gated admission mechanisms are locally inapplicable, not bounded absence findings. Unknown alternative admissions and consumers prevent complete trace-learning coverage.); Automatic session-to-source-page backend pipeline and deferred rendered-task chain: known — yes [wired] (This is an architectural trace-fed write classification, not proof of model truth, completed agent work or exercised learning benefit.); Routine foreground session ingestion: known — yes [afforded] (The host instruction route is afforded, not observed.); Optional metadata-fed reader overview: known — yes [wired] (Separate reader memory route; no claim that this output enters ordinary wiki query context or improves an agent.); Human-approved self-learn method revision and permission-gated saved answers/reflection: inapplicable (These explicitly human-approved admission mechanisms are outside the automatic-write condition of trace learning. This is not an evidenced absence of other automatic writes; the accepted set supplies no bounded absence record for a local no. Self-learn can still carry learning authority at its human-gated update consumer.); Mixed non-session wiki sources, candidate initialization, quality writers and hot/Auto Dream/queue alternatives: not-determinable (Original source/control and later consumer facts are unresolved. Raw-log retention or a cache label alone cannot decide whether an automatic durable behavior-shaping write exists.)	partial coverage (Knowledge, learning, ranking, routing, instruction, validation and enforcement have separate local witnesses. Unconsumed metadata/reasons do not gain authority merely by retention; unresolved consumers prevent complete coverage.); Retained raw and wiki prose at synthesis and answer consumers: known — knowledge [wired], learning [wired] (Knowledge at delivery and learning at the trace-fed update consumer are separate paths.); Wiki text, graph labels/connectivity and quality fields at retrieval selectors: known — ranking [wired], routing [wired] (Relevance/metadata selection does not establish epistemic acceptance or a stale-read veto.); Mtime state, pending identities and queue paths at coordination consumers: partial — routing [wired], validation [wired], enforcement [wired] (Queue retention is implemented but its actual later consumer is unresolved; no queue routing force is inferred from its docstring.); Navigation descriptions and pending task guidance at the foreground host: known — routing [afforded], instruction [afforded], knowledge [afforded] (No automatic injection or binding host policy follows from these file-mediated procedures.); Trace-derived method revisions at update and later host consumers: known — learning [afforded], instruction [afforded] (Human-gated learning authority is not automatic trace learning. Deployment-specific later loading is uninspected.); Retained reader prose at site readers and archived wiki prose at ordinary queries: known — knowledge [wired] (Archival therefore is not general enforcement at read consumers; no dedicated later reason-file consumer is established.); Quality update consumers, hot/Auto Dream/special-memory routes and unread maintenance variants: not-determinable (Production quality-field writers and special-file selectors/loading are unresolved; downstream forces for those routes cannot be enumerated.)	AAS-2026-10-16-llm-wiki-6786c5985cdc-01
llm-wiki	2	Accumulated raw/normalized sources, extraction sidecars, wiki bodies and frontmatter, newsletters, catalogs and operation/research notes, generated paper drafts, monitor admission rows, evaluation fixtures/manifest and operational metrics/experiments; their acquisition, editing, maintenance, access and later-consumer routes. Includes unresolved initial curated-table admission, generic API caller control and metric producer provenance. Excludes static shipped prompts, prompt-registry constants and governance, current-call messages and transient queues, and external provider weights, hidden provider retention and coding-agent host memory. Git history is an externally populated historical-read alternative, not automatic save storage.	known coverage (The recorded inventory resolves substrate within the scoped client memory boundary. Provider internals and host memory are excluded, not classified from readable client payloads.); Raw/normalized source text, JSON sidecars and their access fields: known — files [wired] (Markdown and JSON are encodings on the same file substrate, not graph storage.); Wiki bodies/frontmatter, curated table, newsletters, catalog, operation/research notes and paper outputs: known — files [wired] (Covers the retained file outputs and their metadata, not transient BM25 objects or model context copies.); Monitor admission state and operational metric/experiment rows: known — sqlite [wired] (Local SQL storage, distinct from human-readable experiment notes.); Evaluation fixture inputs, generated outputs and manifest: known — files [wired] (No independent expected-answer store is implied.); Requested historical versions backed by externally existing commits: known — repo [afforded] (API save does not create a commit; this alternative is not promoted to automatic checkpoint wiring.)	partial coverage (Automatic processing and manual browser edits coexist; untyped callers and initial admission prevent complete coverage.); Acquisition/normalization, extraction/integration, monitor rows, metrics/experiment logging, fixture construction, lint/validator and synthesis/draft admissions: known — automatic [wired] (Known subsequent admissions, distinct from unresolved initialization of their inputs.); React editor save of human-supplied article content: known — manual [wired] (This separately evidenced browser path does not establish the identity of every API client.); Generic API saves/uploads, initial curated-table and shipped fixture admission, external historical commits and RSS reset caller: not-determinable (Caller invocation and editable content do not by themselves distinguish human per-content control from software admission. External commit and initial-entry decisions are not reconstructed.)	known coverage (The scoped inventory contains both requested returns and independent downstream supply. Unresolved write provenance does not obscure direction on these recorded consumer routes.); Requested CLI/API query, article, backlink, log and historical retrieval: known — pull [wired] (Current-file retrieval is wired; historical payload availability depends on external commits.); Historical-content delivery if externally created Git commits exist: known — pull [afforded] (Kept separate from live-file reads to preserve its conditional warrant.); Write-triggered old-body reuse, paper/newsletter context, evaluation inputs, admission/maintenance/optimizer selectors and curated polling: known — push [wired] (An upstream CLI/operator choice supplies trigger or targeting inputs, not a separate request by the receiving model.)	known coverage (Every recorded push family has an identified selector. Requested BM25 search remains pull, so its lexical scoring does not independently add a push finding.); Integration target and identity-based monitor/sidecar/manifest selection after driver triggers: known — identifier [wired] (These identities select consumed retained parts. An identifier merely displayed in an article request is not counted.); Newsletter, deep lint, curated polling, retrospective/research scans and optimizer time-window supply: known — coarse [wired] (Count/prefix budgets and broad eligibility remain coarse rather than semantic targeting.); Paper topic-to-page matching after paper-generation trigger: known — inferred-lexical [wired], coarse [wired] (Optional model topic choice is upstream; the retained-page selector itself is lexical, not an embedding or model relevance judgment.)	partial coverage (A positive local automatic trace-fed feedback route coexists with unresolved producer coverage. Neither missing prompt installation nor missing benefit erases that route; neither makes all logs qualifying.); Optimizer execution-feedback to retained experiment/hypothesis/decision state and later proposal selection: known — yes [wired] (The qualifying result is retained optimizer state, not the proposed prompt: candidate wording is neither installed nor saved as active guidance. No improvement is asserted.); Complete metric-row producer coverage and other retained operational/research logs: not-determinable (All original producers and qualifying durable-update/later-consumer chains are not established. Raw logging or possible topic-matched log reuse is not enough to classify every branch as learning.); Ordinary external-source knowledge processing, direct article edits and page-to-page synthesis/metadata maintenance: inapplicable (These recorded inputs are source documents, page claims or caller text, not identified activity traces feeding a trace-learning write. This does not establish a system-wide no or classify unknown input histories.)	partial coverage (Positive forces attach to actual consumers; unresolved history does not erase them or justify complete authority enumeration. Static prompt instruction authority is outside accumulated memory scope.); Source/fixture text to extraction, retained old claims and wiki/newsletter/notes to later models or requested readers: known — knowledge [wired] (Explicit evidence fields omitted by integration gain no invented downstream authority.); Retained wiki text as BM25 token-scoring input: known — ranking [wired] (Ranking influence requires neither learned weights nor a persistent search index.); Admission rows and sidecar-existence checkpoints at acquisition/extraction gates: known — enforcement [wired], validation [wired] (The same retained checkpoint is a criterion and an enforced veto at its actual consumer.); Frontmatter/mtime, fixture paths, curated query table and metric/experiment state at code selectors: known — routing [wired] (Confidence/status displays are not generalized content enforcement; notably conflict state does not quarantine reads.); Retained operational metrics consumed in automatic durable experiment-feedback updates: partial — learning [wired] (This local update-consumer force does not imply learned parameters, improved capacity or installed candidate prompts; complete metric-producer history is unresolved.); Authority of externally committed historical replacements and unaudited initial curated/fixture content beyond recorded consumers: not-determinable (Recorded historical reads and known selectors do not establish all uses or effects of externally initialized content. No retained standing-instruction update is inferred from a generic text save.)	AAS-2026-10-17-llm-wiki-507fade8a56e-01
llm-wiki	2	Retained raw/article bodies and metadata, navigation indexes, operational session state/digests/feedback, research provenance, topic guides/proposals, personal method/allowlist and adapter registry changes, inventory/Ideas, dataset manifests and external payload interfaces, project rationale/deliverables, maintenance findings, checkpoint bundles and benchmark retained artifacts/later consumers. Includes unresolved native-data/media/private-output and external continuation alternatives at their interfaces. Excludes static shipped instructions as accumulated memory, ordinary transient tool returns, host inference weights and provider training internals, and release administration.	partial coverage (Local retained bodies, access metadata and control artifacts are file-backed. External dataset and adapter payloads and host continuation state prevent complete substrate coverage; no repository, graph or vector substrate is inferred from links or Git distribution.); Local raw/articles, indexes/frontmatter, sessions/feedback/provenance, guides, inventory, dataset notes, project text, reports, registries, methods and checkpoint bundles: known — files [wired], files [claimed] (The records enumerate filesystem Markdown, JSON, JSONL and colocated artifacts; this classifies storage rather than their mixed encodings.); External dataset bytes, private adapter outputs and external host continuation storage: uninspected (Locations and receipts do not establish operative external storage substrates; no provider-weight memory is inferred.)	partial coverage (Ordinary model/software admission is automatic even after workflow initiation. Explicit human content choices establish manual paths; generic CLI caller identity does not.); Ordinary source/article writes, index/lint fixes, hook capture/digest/feedback and ordinary lesson appends: known — automatic [wired], automatic [claimed] (No per-content human decision is required on these paths.); User-selected refresh action, adopted topic guide, Idea approval, lesson dry-run save and individually approved instruction proposals: known — manual [claimed] (These contracts specify an explicit content/adoption decision, not merely starting work.); CLI-only digest/feedback promotion, specialist enablement, adapter approval and withdrawal apply; initial/external admission: not-determinable (Explicit invocation alone does not authenticate a human content decision. Initial controls and untraced dataset/media/private branches remain unresolved.)	partial coverage (Requested retrieval and independent hook supply coexist. Untraced dataset/media/private consumers and host continuation alternatives prevent complete direction coverage.); User queries, explicit rehydrate/show/promotion, resume, audit reuse, checkpoint verification/import and benchmark fixture requests: known — pull [claimed], pull [wired] (Requested delivery is pull even when code performs selection and output.); Configured SessionStart/UserPromptSubmit automatic context delivery: known — push [wired] (The hook supplies context without a model request for those retained states.); Native-data/media/private-output read routes and external host continuation delivery: uninspected (A known query pull does not settle these included alternatives.)	partial coverage (The inspected hook push uses exact cwd eligibility and recency/count filtering. Explicit file paths on pull routes are not push signals; opaque alternatives prevent complete coverage.); Automatic start/prompt hook: matching cwd states with digest paths, newest last_seen_at first, default limit three: known — identifier [wired], coarse [wired] (Trigger, selector input and delivered part are explicit: hook event, cwd/state metadata, state summary plus path rather than digest body.); Model-requested query/files, explicit session reads and requested fixture delivery: inapplicable (These named operations are pull; relevance judgment within a requested retrieval is not an independent push selector.); Untraced native-data/private-output and external continuation alternatives: uninspected (Neither availability of manifests nor a known local push establishes selector inputs on these branches.)	partial coverage (Automatic trace-fed durable context/candidate and lesson paths support yes locally. This is not demonstrated capacity improvement; unresolved write alternatives prevent a complete system-wide classification or negative.); Hook metadata/manual-summary state to retained digest/summary and later host context; user turns to retained feedback and later digest/review: known — yes [wired] (Automatic extraction/template writes have a retained output and later context/review consumer; full transcript preservation is unnecessary.); Ordinary conversation-to-lesson raw note/article Rule extraction: known — yes [claimed] (User invocation does not make subsequent model admission manual; individually approved instruction proposals are a separate path.); Research provenance/checkpoint updates, native dataset/media/private writes and external host compaction: not-determinable (Raw provenance retention alone does not establish a behavior-shaping automatic derivation; opaque inputs/write controls/later consumers prevent a bounded no or further yes.)	partial coverage (Authority is local to retained part and actual consumer; advisory evidence is not enforcement. External native-data/private-output consumers and actual personal method contents remain unresolved.); Raw/article/index/project/report prose at later answering, compilation, audit and checkpoint consumers: known — knowledge [claimed] (These paths supply reference and evidence without a universal truth gate.); Topic guide and enabled specialist method at planning/review models: known — instruction [claimed] (Retained standing directions request behavior but do not grant tools.); Trace-derived state at digest/feedback/lesson update consumers, and derived summaries at later host invocations: known — learning [wired], learning [claimed], knowledge [wired] (Update-consumer learning force and output context force are separate; neither demonstrates improved capacity.); State identities/dates and research checkpoints selecting continuation or context: known — routing [wired], routing [claimed] (Routing is operation/window selection, not knowledge endorsement.); Retained date, confidence, score and navigation metadata at query/maintenance prioritization consumers: known — ranking [claimed] (Priority influence is separate from fixed-code validity checks.); Adapter registry/manifest identity, specialist allowlist and checkpoint hashes/overrides at admission consumers: known — validation [wired], enforcement [wired] (Retained symbolic controls have bounded executable force, not universal filesystem or remote safety guarantees.); Native dataset/project payload consumers and private adapter outputs or host continuation consumers: uninspected (Opaque content and untraced consumer routes prevent complete authority coverage.)	AAS-2026-10-17-llm-wiki-a93f86f698d7-01
llm-wiki-coordination	2	Included: downstream thread context/OQ, metadata and orientation estimates, contribution/closure/reopen and historical entries, canon-page content and revision-specific endorsements, persona manifests, retained audit reports/global-log summaries and cemetery tombstones; their admission, maintenance and later participant/page-review/audit consumption. Excluded: unchanged shipped protocols and maintainer source, audit-instance caches/indexes/findings, external provider parameters and unshipped host context management.	known coverage (Known for the complete prescribed retained-object inventory, not observed deployment. Process memory and external model weights are outside this memory boundary.); Thread context, meta/zhat, entries, canon pages, endorsements, manifests, reports/log summaries and tombstones: known — files [claimed] (All recorded retained surfaces are file objects, including unspecified report and tombstone layouts; unknown schema does not imply unknown substrate.)	partial coverage (Known agent and user mechanisms coexist with unresolved profile/import/deletion control.); Agent creation/appending of context, metadata, contributions, evaluations, closing summaries/checkpoints, page edits and individual endorsements; report/log authorship: known — automatic [claimed] (Automatic is a control classification of the prescribed agent path, not implemented scheduling or physical persistence.); Closure recognition, full/soft canonical promotion, manifest recognition and mutually acknowledged modification: known — manual [claimed] (Human admission is separate from earlier agent authorship and endorsement writes.); Orientation-estimate replacement, historical import and tombstone/deletion admission: not-determinable (The profile writer is unspecified; generic adopter/deletion operator identity does not resolve per-content human control for the other paths.)	partial coverage (Requested agent reads and a requested/automatic audit chain coexist with unresolved later delivery of other retained surfaces.); Participant requests for thread context, preceding prose and chain metadata; page work requests for content and consensus: known — pull [claimed] (Both prescribed later-consumer routes are classified without claiming wired model delivery.); Caller-requested wiki audit followed by automatic filesystem selection for the checker: known — pull [wired], push [wired] (The automatic supply reaches deterministic checks, not a model-context injector.); Profile later guidance, manifest loading, report/log and cemetery later readers: not-determinable (Intended guidance/availability does not specify a delivery mechanism; these alternatives cannot be removed from direction coverage.)	partial coverage (Coarse automatic audit supply is supported; referenced-path existence checks add no targeted retained-content delivery. Unresolved external delivery alternatives prevent complete signal coverage.); Audit-triggered automatic selection of wiki Markdown and fixed dialogue metadata/entry paths: known — coarse [wired] (File-type/root coverage is coarse, not an identity-matched model-memory injection.); Conditional cross-protocol audit checks of canon pointers already supplied in retained metadata: inapplicable (L3 checks referenced-path existence and field presence; it does not select and deliver retained canon-page parts. The consumed pointer arrives within the coarse-selected metadata covered above, so this existence predicate supplies no additional push signal.); Thread/page requested reads: inapplicable (These direction units are requested pull reads, so their filenames do not establish push signals.); Profile, manifest and maintenance-history delivery alternatives: not-determinable (Unknown delivery prevents determining whether push exists and which selector would supply it.)	partial coverage (A prescribed qualifying agent transformation is positive; no shipped automatic writer establishes wiring, and other admission/consumer gaps remain local uncertainty rather than a system-wide no.); Agent evaluation and closure-summary/checkpoint writes from retained discussion to later participants: known — yes [claimed] (The qualifying path is derived assessments/continuation context, not mere storage of raw contribution prose.); Orientation refinement and initial preferred-direction content: not-determinable (Past-evaluation input and intended later guidance are specified, but automatic admission, writer and actual durable replacement are unresolved.); Canon production/admission, historical import, manifest maintenance, retained reports and cemetery history: not-determinable (Human canonical/identity recognition is not automatic learning. Earlier canon generation has unspecified transformation and later substantive use; import control and maintenance later consumers are unknown. No bounded record proves a negative across these alternatives.)	partial coverage (Local consumers support several forces; unspecified maintenance readers and manifest/profile delivery prevent complete authority coverage.); Context, preceding contribution and canon prose read by later participants/reviewers: known — knowledge [claimed] (No concrete adopted canon page supports an additional standing-instruction classification.); Contribution prose and peer scores at evaluation, closing summary and next-direction consumers: known — learning [claimed], ranking [claimed], routing [claimed] (All these effects are prescribed external-agent consumers, not dispatched or observed behavior.); Past evaluation vectors at orientation refinement and resulting zhat at later participants: partial — learning [afforded], routing [afforded] (Producer, actual replacement and delivery are unresolved, and the records disagree about doctrine status.); Consensus/revision metadata at page reviewers and canon approvers: known — instruction [claimed], validation [claimed], enforcement [claimed] (Normative veto and criterion are separately classified without attributing implemented enforcement.); Thread migration/status metadata at deterministic audit dispatch: known — routing [wired] (This is an implemented effect on a later checker, not a provider context selector.); Manifest/persona retained identity at recognition and active-agent qualification: partial — knowledge [claimed], validation [claimed] (No loader or reconstruction algorithm establishes the retained manifest's additional downstream behavioral effects.); Retained reports/global-log summaries and cemetery tombstones: not-determinable (No specified later consumer establishes authority for these retained outputs; current stdout advisory force cannot be transferred to them.)	AAS-2026-10-14-llm-wiki-coordination-a12f3c7007f2-01
LLM-WIKI-MCP	2	Included: use-produced or changed wiki pages and raw ingest copies, sidecars, SQLite indexes/provenance/decisions, process caches, exact and compressed ask histories, CLI session/command history, runtime config, metrics/journal records, generated documentation and their write/maintenance/later-consumer routes. Additional editable guidance/maintenance paths and any accumulated provider state remain unresolved included alternatives. Excluded: static shipped prompts/developer instructions, fixed model weights, transient current-call context before retention, external host responsibilities and mere hypothetical external learning.	partial coverage (Known storage routes establish files, SQLite and process memory, not vector, repository or provider-service memory. Additional guidance/maintenance coverage and any accumulated provider state remain unresolved; positive local findings do not close the inventory.); Wiki, sidecars, live and compressed history, configuration, diagnostics and generated documentation files: known — files [wired] (These retained objects use local filesystem files. JSON/JSONL/Markdown are encodings, not separate substrates.); Page index, ingest provenance and decision rows: known — sqlite [wired] (The SQLite-backed access/update structures and decision rows are separate from the readable files.); Search/context caches and CLI session fallback: known — in-memory [wired] (Process-local material survives across requests within a shell/process, not across process exit.); Additional editable guidance and incompletely traced maintenance alternatives: uninspected (The accepted corpus and broad write-route records do not enumerate every editable guidance part or its complete maintenance/consumer chain. No controlled value is inferred for those additional parts; an exhaustive classification is prevented.); Opaque provider state, if accumulated or changed through use: uninspected (The endpoint component is recorded, but its retained mutable state and internal update/consumer paths are uninspected. Shipped or fixed model weights alone are excluded from accumulated memory. No provider-memory value or negative is inferred.)	partial coverage (Automatic derivation is supported; no manual value is inferred merely from a CLI, caller or apply flag. Additional guidance/maintenance coverage and any accumulated provider state remain unresolved; positive local findings do not close the inventory.); Automatic ask/metrics admission and CLI visible-turn guard: known — automatic [wired] (A user initiating an ask does not approve each retained answer/record.); Compaction, file ingestion, sidecar iteration, structural repairs, indexing and document regrouping: known — automatic [wired] (Batch apply or command invocation can authorize work without deciding each produced entry.); Transcript clipping and readable decision-log derivation: known — automatic [wired] (Software transforms supplied transcript/decision fields into retained derivatives; this does not resolve who supplied the originals.); Direct body edits, explicit record tools, config changes and caller-selected rename/delete: not-determinable (Generic caller identity does not establish an explicit human content decision or fully resolve software agency for the original proposal.); Additional editable guidance and incompletely traced maintenance alternatives: uninspected (The accepted corpus and broad write-route records do not enumerate every editable guidance part or its complete maintenance/consumer chain. No controlled value is inferred for those additional parts; an exhaustive classification is prevented.); Opaque provider state, if accumulated or changed through use: uninspected (The endpoint component is recorded, but its retained mutable state and internal update/consumer paths are uninspected. Shipped or fixed model weights alone are excluded from accumulated memory. No provider-memory value or negative is inferred.)	partial coverage (Both requested returns and automatic retained-input supply exist; external host reinjection after a tool return is not established. Additional guidance/maintenance coverage and any accumulated provider state remain unresolved; positive local findings do not close the inventory.); Explicit page/search/retrieve/history/recursive/diagnostic requests and CLI history display: known — pull [wired] (Fulfilling a request for retained content is pull even when code ranks or formats the returned material.); Recent history automatically supplied to intent resolution and an answer prompt: known — push [wired] (A request for an answer is not a separate request for prior turns. Code selects them automatically; final prompt assembly does not guarantee full delivery.); Configuration and provenance automatically loaded during later work: known — push [wired] (Code supplies retained control/update metadata to its decision consumers without a separate request for that retained material.); Archive/recent insertion within an explicitly requested recursive context: known — pull [wired] (The requested return is pull, not an additional push merely because the helper composes text.); Additional editable guidance and incompletely traced maintenance alternatives: uninspected (The accepted corpus and broad write-route records do not enumerate every editable guidance part or its complete maintenance/consumer chain. No controlled value is inferred for those additional parts; an exhaustive classification is prevented.); Opaque provider state, if accumulated or changed through use: uninspected (The endpoint component is recorded, but its retained mutable state and internal update/consumer paths are uninspected. Shipped or fixed model weights alone are excluded from accumulated memory. No provider-memory value or negative is inferred.)	partial coverage (Signals describe automatic supply only. Lexical ranking remains a positive authority finding without being misclassified as push selection. Additional guidance/maintenance coverage and any accumulated provider state remain unresolved; positive local findings do not close the inventory.); Ask-triggered recent-history selection for intent and prompt sections: known — coarse [wired] (Trigger: later ask. Input: available live rows and recency/count/budget rules. Selected part: latest question or bounded recent rows, with separate sanitized tool-row serialization.); Automatic configuration supply at later dispatch/recording/context construction: known — coarse [wired] (Trigger: later operation. Input: current config/default fields. Selected part: active settings as a broad control bundle, not targeted memory matching.); Automatic ingest-status lookup: known — identifier [wired] (Trigger: re-ingest. Input: source-file identity/path plus recorded hash and datestamp. Selected part: matching source-status row used to skip or update.); Explicit query-ranked page/note retrieval: inapplicable (Lexical scoring on these requested retained-content returns does not establish an inferred-lexical push signal. No positive push finding is assigned to this pull operation.); Additional editable guidance and incompletely traced maintenance alternatives: uninspected (The accepted corpus and broad write-route records do not enumerate every editable guidance part or its complete maintenance/consumer chain. No controlled value is inferred for those additional parts; an exhaustive classification is prevented.); Opaque provider state, if accumulated or changed through use: uninspected (The endpoint component is recorded, but its retained mutable state and internal update/consumer paths are uninspected. Shipped or fixed model weights alone are excluded from accumulated memory. No provider-memory value or negative is inferred.)	partial coverage (Qualifying automatic trace-to-durable-context routes support yes locally, not observed learning benefit or a system-wide negative elsewhere. Additional guidance/maintenance coverage and any accumulated provider state remain unresolved; positive local findings do not close the inventory.); Automatic older-turn compaction to persistent archive with recursive-context consumer: known — yes [wired] (The archive is a durable behavior-shaping context artifact with an implemented later context-return route. External model delivery is afforded, not observed; improvement is not required for this axis.); Automatic transcript clipping to indexed page for later answers: known — yes [wired] (The transformation retains reduced conversational material and makes it available to ordinary later context retrieval. Actual source authenticity and model activation remain unobserved.); Raw ask records, metrics, journal helpers and CLI copies apart from qualifying reductions: not-determinable (Raw logging alone is not a qualifying update; these records do not provide a bounded exhaustive absence warrant for all downstream transformations. No local no is asserted.); Sidecar updates over pages whose initial provenance may include transcripts: not-determinable (Known note rewriting does not resolve which original experience sources feed a particular page or whether all included paths are trace-fed. No trace-learning positive or negative is inferred for this alternative.); Additional editable guidance and incompletely traced maintenance alternatives: uninspected (The accepted corpus and broad write-route records do not enumerate every editable guidance part or its complete maintenance/consumer chain. No controlled value is inferred for those additional parts; an exhaustive classification is prevented.); Opaque provider state, if accumulated or changed through use: uninspected (The endpoint component is recorded, but its retained mutable state and internal update/consumer paths are uninspected. Shipped or fixed model weights alone are excluded from accumulated memory. No provider-memory value or negative is inferred.)	partial coverage (Authority is attached to actual local consumers or an explicitly afforded external role. Static answer rules do not make every retained page an instruction; no common validation register is inferred. Additional guidance/maintenance coverage and any accumulated provider state remain unresolved; positive local findings do not close the inventory.); Wiki, sidecar and recent-history text delivered as answering evidence or follow-up context: known — knowledge [wired] (These parts supply interpreted evidence/context, not executable permission or verified truth.); Archive and recent history returned by recursive context: known — knowledge [afforded] (Local composition and context return are wired; actual external model use is not inspected.); Page/title/note text and FTS/cache content at retrieval scorers: known — ranking [wired] (Lexical fixed-code ranking gives retained content influence before any semantic acceptance.); Prior questions at intent resolution, provenance at ingest planning, and config at later dispatch: known — routing [wired] (Consumers select a resolved query, skip/update work, endpoint/model/mode or input budgets.); Configured history enablement and budget controls at recorder/pack code: known — enforcement [wired] (These retained controls can suppress append or constrain delivered material. The effect is code-enforced, not model instruction.); Prior turns at automatic compaction and transcript records at clipped-page production: known — learning [wired] (The update consumer uses experience to produce a durable later-context artifact. This is distinct from downstream knowledge authority and from improved capacity.); Generated sidecar suggestions/actions at later human or agent editor: known — instruction [afforded] (The emitted note supplies directions for review/promotion; actual editorial compliance is unobserved.); Diagnostic metrics and CLI/session history at explicit diagnostic/display consumers: known — knowledge [wired] (Reports and history displays inform callers; no automatic score-driven policy update is established.); Additional editable guidance and incompletely traced maintenance alternatives: uninspected (The accepted corpus and broad write-route records do not enumerate every editable guidance part or its complete maintenance/consumer chain. No controlled value is inferred for those additional parts; an exhaustive classification is prevented.); Opaque provider state, if accumulated or changed through use: uninspected (The endpoint component is recorded, but its retained mutable state and internal update/consumer paths are uninspected. Shipped or fixed model weights alone are excluded from accumulated memory. No provider-memory value or negative is inferred.)	AAS-2026-10-15-llm-wiki-mcp-5dff0dd0576a-01
LLM-Wiki-v3	2	Retained raw bytes and metadata, pending/admitted page bodies and fields, ingest manifests, lexical/graph/vector access derivatives, audit records, evaluation queries and reports, and their shipped write, maintenance and later-consumer routes. Optional Git preservation of admitted pages is included as an affordance. Excludes static shipped extraction instructions/schema, model weights and provider-internal retention, transient chunks/query scores and in-invocation baseline counters, and unshipped MCP/answer-context compilation. Missing raw/index artifacts limit observation, not the implemented storage routes.	known coverage (The recorded inventory resolves shipped memory substrates. Unknown external model artifacts do not change these local storage classifications.); Raw source and sidecar, pending/admitted pages, manifests, audit, datasets and reports: known — files [wired] (Classifies implemented stores, not a replay of unavailable raw artifacts.); Lexical, graph and optional dense retained access derivatives: known — sqlite [wired], graph [wired], vector [wired] (Optional dependency availability does not remove the implemented dense branch.); Optional Git preservation of admitted page history: known — repo [afforded] (This is a history/recovery affordance, not a transactional storage guarantee.)	partial coverage (Supported automatic and manual mechanisms coexist with unresolved callable/initialization control.); Raw retention, proposal staging, derivative/diagnostic/report construction, seed execution and bootstrap admission: known — automatic [wired] (Human initiation is not manual approval of each resulting page or row.); Interactive review content edits and accept/reject decisions: known — manual [wired] (Manual admission belongs to this route, not every callable write.); Direct callable acceptance and initial active dataset admission: not-determinable (A generic callable supplies no proof of human content decision; active dataset initialization history is unknown. Neither is classified by physical I/O alone.)	known coverage (The scoped inventory contains requested pull routes and bootstrap's automatic retained-body push to the generation model. Unshipped MCP and excluded answer construction add no route.); Raw reload/checking, pending review, requested rebuild/query/evaluation and upstream bootstrap generation request: known — pull [wired] (Query ranking fulfills a retrieval request. The upstream bootstrap command requests generated queries, not a retained-body read by the later model consumer.); Bootstrap automatic enumeration and retained-body delivery to each fresh model invocation: known — push [wired] (This automatic input selection and delivery is distinct from both the upstream generation request and the requested query return; it does not establish answering-agent integration.); Optional later operator browsing of stored diagnostics/history/reports: known — pull [afforded] (No automatic report-to-policy consumer is projected from retention.)	known coverage (The included push operation has known coarse selection. Other request-led operations have no push signal to classify; lexical/vector query retrieval and requested-file identifiers do not add push signals.); Bootstrap automatic enumeration and retained-body delivery to each fresh model invocation: known — coarse [wired] (There is no explicit input-body truncation budget; the response cap is not a retained-body selector. Page stems supply output labels, not identity-matched push selection.); Request-led reload/review/rebuild/query/evaluation, upstream bootstrap request and optional inspection operations: inapplicable (This axis classifies push selection only. Query lexical/vector/graph selectors fulfill requests, and identifiers on requested raw files are not identifier-based push.)	not-determinable coverage (No controlled finding is warranted from these records: missing qualifying provenance/consumer facts cannot become either a positive or an unrecorded system-wide negative. This says nothing about provider-internal learning.); Document/seed-to-page and page-to-evaluation-query automatic writes: not-determinable (Known writes consume document prose or page bodies, not established session/tool/event/trajectory inputs. No accepted bounded absence record licenses a local no, and no qualifying trace-fed chain licenses yes.); Audit, ingest diagnostics and evaluation-result retention and possible later adoption: not-determinable (Process records persist, but no recorded automatic write turns them into a durable behavior-shaping artifact with a later consumer. Raw trace retention alone is insufficient; the set declares no evidenced-absence record.)	known coverage (The records bound actual consumers and nonconsumed outputs. Static system instructions and external model parameters are excluded; no retained instruction or learning authority is inferred from them.); Raw/source/page prose and provenance at checking, browsing and bootstrap consumers: known — knowledge [wired] (No downstream answering prompt is implied; query returns metadata only.); Page search content and lexical/graph/dense derivatives at retrieval and baseline ranking consumers: known — ranking [wired] (Direct baseline loading and index-backed component evaluation are distinct ranking routes.); Retained type/currentness fields and raw metadata at destination/load/storage consumers: known — routing [wired], enforcement [wired] (These are concrete local effects, not a system-wide authorization boundary.); Expected IDs and retained source pointers at evaluation/check consumers: known — validation [wired] (Software checks and metric criteria are distinct from human truth acceptance.); Manifests, audit records and saved reports without an automatic decision consumer: known — knowledge [afforded] (Afforded human inspection remains separate from wired automatic authority.)	AAS-2026-10-18-llm-wiki-v3-0c553a3118eb-01
llmwiki-marimo	2	Accumulated source/extracted text; wiki prose, collection/index metadata and retained warnings; dataset files; generated aliases and mutable workspace controls including configured prompt text; stored page diagnostics; caller-held history reused by later invocations; optional trace artifacts; Git/index recovery and frozen developer reference copies. Includes their known writes, transformations and later consumers plus locally unresolved initialization, maintenance and evaluator alternatives. Excludes shipped static prompts, unretained current-call results, external provider state/weights and independently unfrozen dependency implementations.	known coverage (The recorded inventory covers file-backed knowledge and controls, SQLite indexes/diagnostics, caller-held history, inspection artifacts and recovery/developer copies. External provider internals are excluded, not classified as accumulated memory.); Source files/cache and extracted source rows/chunks: known — files [wired], sqlite [wired] (Original files and indexed extraction are separate retained parts.); Wiki pages, datasets and configuration controls: known — files [wired] (Markdown/TOML parts are retained files; transient dataset parsing does not add a database store.); Search/reference state and stored verification: known — sqlite [wired] (The indexed content, access metadata and diagnostic scores are SQLite state.); Caller-held conversation and reduced later-turn messages: known — in-memory [wired] (Later turns reuse application-held messages; cross-session persistence is not established.); Trace inspection artifacts, history, snapshots and golden fixtures: known — files [wired], repo [wired], sqlite [wired] (The recorded persistence mechanisms are classified, not universal availability or lossless restoration.)	partial coverage (Human content control and software admission both exist. Initial dataset/workspace admission and unaudited mutation paths remain unresolved.); Ingestion, aliases, diagnostics, repair mutations and trace recording: known — automatic [wired] (A user may start the job; code admits the resulting retained content without per-content approval.); Web submitted content, manual page/control edits and requested historical replacement: known — manual [wired] (The explicit human decision supplies or replaces the retained content at these mechanisms.); Marimo/convenience generated page admission: known — automatic [wired] (Starting save does not make the subsequently generated draft human-approved.); Golden fixture replacement: known — manual [wired] (The operator explicitly chooses the staging snapshot for replacement; code performs the copy.); Initial/external dataset replacement and pre-existing workspace initialization: not-determinable (Generic workspace/file origin does not establish who decided the retained content.); Remaining unaudited repair/evaluation artifact admissions: not-determinable (Known automatic transformations do not resolve all uninspected mutation paths.)	partial coverage (Requested reads and automatic context selection both exist within chains. The request for an answer is not a request for each selected retained part; code selection is a distinct push operation. Remaining maintenance/evaluator internals prevent complete direction coverage.); Model-requested page/search/data tools and operator inspection/restore: known — pull [wired] (Requested material returns are pull; mere code delivery is not another push.); Pre-retrieval evidence and shaped history delivery to answer/condenser models: known — push [wired] (Code selects retained material before the answer invocation, rather than waiting for its retrieval request.); Same-slug existing-page input and retained configuration at later invocations: known — push [wired] (Code selects earlier retained content for generation or config consumers.); Stored diagnostic and frozen fixture reads: known — pull [wired] (Page/lint callers load diagnostic state; development callers load a chosen reference baseline.); Other repair, evaluation packet and feedback read paths: not-determinable (Detailed consumers are not completely inspected; no pull-only boundary or complete union is asserted.)	partial coverage (Signals describe push selectors, not identifiers merely carried by requested files. Coarse and targeted selection remain distinct; unresolved paths are not made inapplicable by known pulls.); Pre-retrieval wiki/source/dataset selection: known — inferred-lexical [wired], identifier [wired], coarse [wired] (Trigger: requested QA turn. Inputs: question or accepted rewrite, aliases/roster, open-page identity and collection intent. Selected parts: chunks, page bodies, dataset rows or key names.); Later-turn history and configured-context delivery: known — coarse [wired] (Trigger: later model/agent invocation. Inputs: available history, refusal/table predicates, tail count and loaded workspace settings. Parts: filtered recent text and configured guidance.); Existing-page selection for ingestion and chat structuring: known — identifier [wired] (Trigger: concept update or same-title draft. Input: generated/selected slug. Part: matching existing page body.); Condenser-mediated pre-retrieval choice: known — inferred-judgment [wired] (Trigger: uncovered follow-up with history. Input: history and question. Part: retained evidence selected after a model rewrite is rechecked.); Remaining maintenance/evaluator automatic selection: not-determinable (Incomplete selector inspection prevents a complete signal inventory; pull file identifiers alone do not supply a push signal.)	partial coverage (Software-admitted chat-to-page writes establish a qualifying trace-fed durable artifact and later consumer. Other included routes do not have bounded absence records; no local or system-wide no is asserted.); Marimo/convenience conversation-to-wiki admission: known — yes [wired] (Recorded conversational turns feed automatic structuring/write and the resulting durable page enters later retrieval or merge.); Human-reviewed web admission and direct human edits: not-determinable (Human content approval is established here, not an automatic trace-fed admission; no evidenced-absence record closes all alternatives.); History reduction and condenser rewrite: not-determinable (Later-invocation context reuse is established, but no durable conversation-summary checkpoint is recorded; this does not support a qualifying positive or bounded negative.); Source ingestion, datasets, aliases and diagnostics/repair: not-determinable (Ordinary source content is not automatically a trace. Its complete origin and all repair/update alternatives are not established as trace-fed or bounded absent.); Raw spans, evaluator outputs, snapshots and frozen fixtures: not-determinable (Raw retention/recovery/reference use alone is not a qualifying automatic durable behavior update. Uninspected feedback and evaluator internals prevent an absence conclusion.)	partial coverage (Authority is consumer-relative. The supported forces below do not establish beneficial activation; uninspected maintenance/evaluator consumers prevent a complete inventory.); Source/wiki/dataset prose and retained conversation at later model/human consumers: known — knowledge [wired] (Inspection rendering is reference use; QA context is not an executable instruction.); Existing page at update models and saved conversation at structuring models: known — learning [wired] (Prior content shapes a durable page update, separately from its downstream evidence role.); FTS content and dataset values at retrieval/projection ranking consumers: known — ranking [wired] (Fixed algorithms can rank using retained operands without learned ranking weights.); Retained configured prompt at agent construction: known — instruction [wired] (Only the user-configured retained prompt part is included; shipped static templates are excluded.); Aliases, fingerprints, blacklist/settings and dataset attributes at scope/grant selectors: known — routing [wired], enforcement [wired], validation [wired] (Each effect is bounded to its inspected consumer.); Verification scores/hashes and retained threshold at diagnostic consumers: known — validation [wired], knowledge [wired] (These classify diagnostic reuse/support, not page truth or universal chat admissibility.); Warning/TODO text and recovery state at maintenance/restore consumers: known — knowledge [wired], routing [wired] (Readable warnings remain advisory while markers and identities drive code decisions.); Frozen golden fixtures at regression/evaluation consumers: known — validation [wired] (The reference baseline checks conformity, not universal factual truth.); Remaining repair/evaluator consumption paths: not-determinable (Every repair, packet/rubric consumer and evaluation feedback path was not audited; no complete authority union follows.)	AAS-2026-10-18-llmwiki-marimo-58054ccd9ae7-01
Mem0	2	Accumulated Python and TypeScript memory text and access structures; Python recent messages/history; coding-agent evidence/checkpoint/retrieval state; service-held memories as exposed by clients; installed extraction-guidance overrides; inspected OpenClaw, Claude Code, Pi and Hermes write/maintenance/later-consumer paths. Includes unresolved hosted internals, optional stores, other coding adapters, async/provider variants and CLI/framework/dashboard consumers named by the accepted inventory. Excludes external provider weights/implementations, native host compaction, unavailable evaluator contents and shipped static prompts/protocols as memory objects; those may describe included mechanisms.	partial coverage (Known local stores coexist with an exposed service-object boundary and unresolved physical/provider alternatives. Model weights and shipped repository prompts are production dependencies, not accumulated memory substrates here.); Python fact text, access payload, dense vectors and entity collection: known — vector [wired] (These operative parts reside in configured vector collections; entity links do not establish graph storage.); Python recent messages and mutation history: known — sqlite [wired] (Separate SQLite tables retain raw context and old/new text plus event metadata.); Claude Code raw evidence, flush and retrieval bookkeeping: known — sqlite [wired] (Local evidence.sqlite3 stores event payloads and retrieval copies; it is not the remote derived store.); Persisted server configuration overrides including extraction guidance: known — rdbms [wired] (Settings rows retain override JSON; database persistence is attempted and can fail.); Pi session seen-ID state: known — in-memory [wired] (Session-local selection state accumulates through recalls and resets at session start; this does not imply durable fact storage.); Hosted retained memories as exposed service objects: partial — service-object [afforded] (Clients afford retained objects retrievable by search/get; backend substrate and actual persistence remain opaque. This is not a vector-store inference.); Included unresolved alternatives: hosted internals; optional TypeScript history stores; other coding-agent adapters, framework/CLI/dashboard consumers, async and provider variants: uninspected (Alternative history substrates, non-default adapters and remote physical persistence are not classified by accepted records; no exhaustive substrate union is established.)	partial coverage (Automatic mechanisms are positively wired; no manual value is established merely from a generic caller or admin role.); Inferred extraction, procedural summary, raw message-tail save and index/history writes: known — automatic [wired] (Software/model admission lacks per-content human approval. Triggering add does not make its subsequent admissions manual.); OpenClaw foreground-agent fact admission and legacy capture: known — automatic [wired] (Agent fact choice and callback filtering/admission are computational; host scheduling is not a human decision approving each fact.); Coding checkpoint assembly, Pi capture and Hermes trace sync: known — automatic [wired] (Adapters select/redact and submit traces without mandatory per-memory human review; remote internal selection remains opaque.); Generic direct/raw insertion, explicit replacement, CLI file selection and server override installation: not-determinable (Caller/admin initiation and permission do not reveal whether a human supplied or approved the retained content. Known software I/O is not enough to classify these content-control decisions.); Included unresolved alternatives: hosted internals; optional TypeScript history stores; other coding-agent adapters, framework/CLI/dashboard consumers, async and provider variants: uninspected (Other adapters, explicit Pi tools and remote admission decisions lack complete recorded control facts; automatic witnesses do not establish a universal admission policy.)	partial coverage (Pull requests and automatic host supply coexist at different operations in the same chain; neither eliminates the other.); README search, extraction context request and explicit ID/history/search tools: known — pull [wired] (Consumers request retained material. The backend return fulfilling a hook search is also pull at that backend boundary, not a second automatic push.); Automatic OpenClaw/Pi/coding first-prompt/Hermes prompt-time supply: known — push [wired] (Host events/current prompts trigger selectors that supply recalled facts without an explicit memory request from the receiving model; external host invocation remains conditional.); Coding sidekick start delivery of selected parent recall snapshot: known — push [wired] (Start code automatically supplies stored injected-memory text for the tracked parent session, not the whole parent conversation.); Hosted API retained-object retrieval contract: partial — pull [afforded] (Later service search/get is afforded; wrappers do not establish remote persistence or all later consumers.); Included unresolved alternatives: hosted internals; optional TypeScript history stores; other coding-agent adapters, framework/CLI/dashboard consumers, async and provider variants: uninspected (Unread consumers and backend alternatives prevent a complete direction inventory. The proxy recall branch is incompatible and is not counted as successful delivered read-back.)	partial coverage (Push selection is not inapplicable because concrete push routes exist. Backend query affordances do not resolve exact algorithms; identifier suppression is distinct from coarse budgets and identity-matched snapshot delivery.); OpenClaw automatic prompt-build recall: partial — coarse [wired] (Interactive non-bootstrap prompt triggers provider semantic search followed by scope/category/importance/relevance filtering and approximate budget over retained fact text.); Coding first-prompt, Pi before-agent-start and Hermes query prefetch: partial — coarse [wired] (Current prompt/query triggers scoped backend search and automatic context supply. Accepted records establish query/scope/top-k/seen-ID/format limits, not the opaque remote relevance algorithm.); Pi and coding recall already-shown ID exclusion: known — identifier [wired] (Automatic prompt recall matches candidate memory IDs against accumulated seen/retrieval state to select unseen delivered facts; IDs on ordinary requested gets are not counted.); Sidekick start selection of parent injected-memory snapshots: known — identifier [wired] (At sidekick start, tracked parent session/repository identity selects its stored injected-memory text for supply; this is an actual identity-matched delivery route.); Included unresolved alternatives: hosted internals; optional TypeScript history stores; other coding-agent adapters, framework/CLI/dashboard consumers, async and provider variants: uninspected (Opaque remote relevance selectors and unread push branches prevent complete signal coverage; no inferred-lexical or inferred-judgment push is inferred from Python public search or extraction model admission.)	partial coverage (Positive routes suffice for existence, not complete coverage or benefit. Raw trace storage alone and static shipped protocols are not treated as trace learning.); Python inferred conversational facts and procedural continuation summaries with later extraction/answer use: known — yes [wired] (Automatic conversation-fed writes produce retained context/summary text available to later consumers. Durable behavior-shaping here means later model input, not learned weights or demonstrated improvement.); Coding checkpoints, Pi capture, Hermes sync and OpenClaw legacy capture to future recall: known — yes [wired] (Adapters wire automatic trace submission and subsequent fact delivery. Remote extraction/admission and host activation remain opaque/unobserved; the positive is the implemented route, not service compliance.); Generic raw/direct replacement, CLI import and configuration submissions: not-determinable (Content can be static or trace-derived and human or computationally selected. These paths do not establish either qualifying trace learning or its bounded absence.); Included unresolved alternatives: hosted internals; optional TypeScript history stores; other coding-agent adapters, framework/CLI/dashboard consumers, async and provider variants: uninspected (Other adapters and hosted internal write/consumer routes remain uninspected; no system-wide no or exhaustive yes-route inventory is justified.)	partial coverage (Each force is local to an actual consumer. Retained experience at an update consumer and derived fact text at an answer consumer have distinct learning and knowledge paths; no blanket equivalence is asserted.); Fact and procedural text at README answer model and host recall consumers: known — knowledge [wired] (Selected text supplies reference/context to later models; system-message placement alone does not make each fact a standing instruction.); Accumulated custom extraction instructions at extraction model: known — instruction [wired] (Installed guidance shapes subsequent extraction requests; generated proposals have no installed force until separately submitted.); Retained prior messages and existing fact context at extraction update consumer: known — learning [wired] (Experience/context influences a later automatic fact-producing write. Learning authority here does not establish improved capacity or weight updates.); Stored fact text, embeddings, keyword terms and entity links at search scoring: known — ranking [wired] (Retained content and features affect relative priority under fixed similarity/lexical/entity scoring; retained ranking policy is unnecessary.); Scope fields and local checkpoint/status state at selection and flush assembly: known — routing [wired] (Stored scopes and event ranges/status choose retained inputs and submission/retry windows.); Retained hashes at candidate admission: known — validation [wired], enforcement [wired] (Existing hashes supply both a conformance criterion and a veto; this checks bounded identity, not truth.); Expiry metadata and accumulated seen IDs at delivery filters: known — enforcement [wired] (Expiry excludes default public search results; seen IDs veto re-delivery. Direct ID and extraction counterpaths remain; this is not universal suppression.); Already-shown memory ID set at recall eligibility check: known — validation [wired] (Membership determines eligibility for another injection as well as enforcing its exclusion; it is not content-truth validation.); Included unresolved alternatives: hosted internals; optional TypeScript history stores; other coding-agent adapters, framework/CLI/dashboard consumers, async and provider variants: uninspected (Hosted internal consumers and unread integration/admission routes have unresolved forces; metadata with no recorded consumer is assigned none, not assumed powerless everywhere.)	AAS-2026-10-14-mem0-5b642a1731b4-01
memex	2	Includes raw archive, model-changeable wiki pages/schema/index/activity log, retained queue payload/results/audit events, mutable per-wiki instruction/tool/model/MCP configuration and unresolved use-accumulated external auxiliary/checkpoint alternatives. Includes their write, maintenance, replay, requested file/client read and configuration delivery routes. Excludes bundled static prompts, current-run transient context, credentials as access metadata, and model weights with no established Memex use-fed update; external provider internals are not attributed to Memex.	partial coverage (The inventory covers wiki products, archive, operational histories and mutable configuration. Local stores are classified; included opaque external retained alternatives prevent complete substrate coverage.); Raw archive, derived pages, schema, index and activity log: known — files [wired] (All named local products persist as files; model-created instances are not supplied.); Queue payload/results and daemon audit rows: known — sqlite [wired] (These distinct retained objects share SQLite persistence.); Mutable custom guidance, allowed-tool lines and MCP configuration: known — files [wired] (Configuration files are distinct from bundled static policy.); Per-wiki model selection retained in registry: known — sqlite [wired] (The model string is registry state, not retained model weights.); Included external Claude auxiliary/checkpoint state and configured MCP alternatives: storage: uninspected (The accepted annotations leave provider-managed directory state, internal context retention and configured MCP behavior uninspected. No controlled value or exhaustive route inventory follows.)	partial coverage (Model/software admission has positive automatic witnesses. Owner calls and editor access do not resolve human versus automation control for supplied content.); Delegated page/schema/index/log writes during ingest/query/lint: known — automatic [wired] (Starting a job is not per-content approval; prompt partitions do not add an approval gate.); Queue/result and coarse audit admission: known — automatic [wired] (Physical persistence follows fixed runtime rules, separate from authorship.); Scaffold seeds: known — automatic [wired] (Defaults are admitted by creation code, not individually approved content.); Owner raw upload, job/history submission, administrative withdrawal and configuration/editor replacement: not-determinable (The recorded owner/caller can be human or automation. An editor command and UID authorization do not establish an explicit human content decision.); Included external Claude auxiliary/checkpoint state and configured MCP alternatives: admission control: uninspected (The accepted annotations leave provider-managed directory state, internal context retention and configured MCP behavior uninspected. No controlled value or exhaustive route inventory follows.)	partial coverage (Model-directed file requests and client requests are pull; fixed replay/configuration supply is push. Automatic lint scheduling alone is not another push of wiki content. External host read-back remains unresolved.); Ingest/query/lint requested filesystem reads: known — pull [claimed] (Request, model selection and file delivery are one pull chain; path mismatch limits realized access.); File browsing and later stored result/audit retrieval: known — pull [wired] (Exact-path reads/status/log calls fulfill consumer requests rather than automatically targeting memory.); Restart pending-payload consumption: known — push [wired] (Scheduler automatically reintroduces stored work into a fresh model task without a new request for that retained payload.); Per-job retained tool/model/MCP configuration resolution: known — push [wired] (A job request is not a request for these retained configuration contents; the runner automatically resolves them.); Included external Claude auxiliary/checkpoint state and configured MCP alternatives: read-back direction, including custom instruction discovery: uninspected (The accepted annotations leave provider-managed directory state, internal context retention and configured MCP behavior uninspected. No controlled value or exhaustive route inventory follows.)	partial coverage (Selectors of wired configuration/replay pushes are retained separately from requested file names. No embedding or lexical push selector is established; external discovery remains opaque.); Runner configuration push: job trigger, wiki identity, selected per-wiki files/model row: known — identifier [wired] (Wiki identity selects the configuration actually supplied to the next invocation.); MCP path push: per-job presence check: known — coarse [wired] (File availability is an additional coarse selector, not proof of payload interpretation.); Replay payload push: startup/pending-drain trigger, status/order and wiki match, selected queued payload: known — coarse [wired], identifier [wired] (Status/order filtering and identity selection have separate warrants.); Requested file/result/audit delivery: inapplicable (These pull operations have no independent push selector; exact requested paths are not identifier-push evidence.); Included external Claude auxiliary/checkpoint state and configured MCP alternatives: automatic selection signals: uninspected (The accepted annotations leave provider-managed directory state, internal context retention and configured MCP behavior uninspected. No controlled value or exhaustive route inventory follows.)	partial coverage (Schema adaptation and retained activity summaries supply requested automatic experience-to-later-context paths, not observed learning or benefit. Opaque sources/host updates prevent a complete trace-learning verdict.); Current filing experience to revised schema to future ingest/lint guidance: known — yes [claimed] (The requested automatic policy update and later read are qualifying at the claim layer; no historical log transformer or instrumented execution is established.); Completed job activity to durable log summary to later maintenance/reader context: known — yes [claimed] (A generated continuation/history summary can qualify without new knowledge; logging alone would not suffice, so retain the named later read routes.); Audit/job histories, mixed supplied inputs and source-driven page updates: not-determinable (Raw retention is not a qualifying learned artifact. Prior results are not automatically fetched by the model, while source files/history can have unspecified trace origins; records do not supply a bounded absence warrant.); Included external Claude auxiliary/checkpoint state and configured MCP alternatives: trace-fed updates: uninspected (The accepted annotations leave provider-managed directory state, internal context retention and configured MCP behavior uninspected. No controlled value or exhaustive route inventory follows.)	partial coverage (Authority is assigned at each recorded consumer, not inferred from file names or retention alone. Opaque MCP/custom-discovery effects and auxiliary state prevent complete coverage.); Raw sources and wiki pages at ingest/query/reader consumers: known — knowledge [claimed], knowledge [wired] (Content supplies reference/evidence; grounding is not an enforced truth gate.); Schema at future filing consumer: known — instruction [claimed] (Standing conventions are requested through file reads, not injected by the daemon.); Schema at lint comparison consumer: known — validation [claimed] (The model compares file patterns against retained conventions; no fixed validator enforces their truth.); Index at model discovery and relevance-selection consumer: known — knowledge [claimed], routing [claimed], ranking [claimed] (Semantic summaries guide which pages are selected, without a daemon scorer.); Activity log at maintenance/file-reader consumer: known — knowledge [claimed], knowledge [wired] (Historical descriptions are context, not approval records.); Queue payload/control state at scheduler and replay consumer: known — routing [wired] (Use-accumulated pending work chooses continuation/input rather than warrant.); Returned results and audit details at later client reader: known — knowledge [wired] (Completion and audit metadata establish operational history, not accepted facts.); Allowed-tool and model configuration at next runner: known — routing [wired] (Actual tool enforcement belongs to external Claude; Memex controls invocation selection.); Mutable .claude.md at intended external discovery consumer: known — instruction [afforded] (Intended guidance is classified only at its documented affordance; activation is uninspected.); MCP consumed payload and provider-managed auxiliary/checkpoint state: not-determinable (A supplied configuration path or credential directory does not reveal its later consumer/effects.)	AAS-2026-10-14-memex-42ecde52cc05-01
Memori	2	Included: accumulated conversation rows, derived facts, embeddings, mention/access metadata, summaries, graph triples, process attributes, hosted conversation/agent/resume memory, benchmark index/metadata/report parts and unresolved feedback retention; their admission, maintenance and later-consumer routes, including native mock/fallback and uninspected adapter/framework alternatives. Excluded: transient provider request envelopes, static shipped skills/configuration, inference/embedding model weights and external host machinery. Hosted retained state is included at its client boundary, with private implementation unresolved.	partial coverage (Local rows, access structures, evaluator files and hosted interfaces are covered; private storage and cross-adapter implementation remain unresolved.); Inspected SQLite conversations, facts, summaries, process attributes and access metadata: known — sqlite [wired], rdbms [wired] (These retained parts are database rows/columns, not transient request messages.); SQLite semantic graph and fact-vector access structures: known — sqlite [wired], rdbms [wired], graph [wired], vector [wired] (Logical graph/vector stores coexist with their physical SQLite substrate.); Benchmark FAISS index, metadata and evaluation JSON: known — files [wired], vector [wired] (Notebooks write and later reload index/metadata files; reports are saved locally.); Hosted conversations, agent memory, resume material and cloud SDK memory: partial — service-object [wired] (Client contracts establish service-held state, not its private physical stores or ontology.); Other configured drivers and bridges: uninspected (Configured database alternatives remain included; their physical substrate semantics were not independently audited.); Feedback retention behind submission endpoint: not-determinable (Transmission is wired; whether feedback is retained and in what substrate is unavailable.)	partial coverage (Automatic client admission is known. Caller identity is not a human decision; private initialization, fallback content control and feedback handling remain unresolved.); Completed-turn conversation admission and generic-stream exhaustion writes: known — automatic [wired] (Software decides to persist formatted turns; a human starting a call is not per-content approval.); Mapped service response, graph/process/summary writes and access updates: known — automatic [wired] (Client mapping, nonempty checks and driver rules admit writes automatically, independently of opaque proposal policy.); Host capture under Hermes/OpenClaw/Claude and MCP submission policy: partial — automatic [wired], automatic [claimed] (Capture wiring is inspected for shipped clients; MCP agent admission is documented, not server implementation.); Benchmark index, metadata, answers and report admission: known — automatic [wired] (Running a notebook does not constitute explicit approval of each retained output.); Native mock/fallback content control, private hosted initialization and feedback handling: not-determinable (Generic caller or submission identity does not decide human versus computational admission. No manual classification is established.); Explicit entity withdrawal caller decision: not-determinable (The API caller chooses destructive maintenance, but the records do not identify an explicit human content decision.)	partial coverage (Requested retrieval and automatic within-invocation selection coexist. Internal SDK requests are pull operations; unrequested context injection within the chain is push, not independent background supply.); SDK recall backend requests and automatic fact/history delivery: known — pull [wired], push [wired] (The wrapper requests retained data; its automatic selector enriches the provider request without a consumer request for those memory items.); Prior-summary read into Python augmentation: known — push [wired] (On augmentation, software selects retained summary and messages without a separate request from the update consumer for those parts.); Intentional host recall, summary and compaction: known — pull [wired] (These inspected clients return requested service material; JSON delivery is not another push.); Documented MCP retrieval: known — pull [claimed] (Classification concerns the documented route, not verified server operation.); Benchmark reload and automatic context selection: known — pull [wired], push [wired] (Notebook reload requests saved artifacts; automatic hybrid retrieval supplies the answer model selected memories.); Independent graph/process, feedback/report read-back and framework-specific alternatives: uninspected (No independent later consumer is established for these stored outputs, and framework-specific branches are not exhaustively inspected.)	partial coverage (Signals describe automatic selectors only. Requested host project/session/category filters are not additional push signals; private SDK selection and uninspected alternatives remain unresolved.); Invocation-triggered local fact selection: known — inferred-embedding [wired], inferred-lexical [wired], identifier [wired], coarse [wired] (Trigger: registered model call. Inputs: query text/embedding and entity; selected parts: facts and mention-linked summaries, after bounded candidate ordering and threshold.); Conversation replay and prior-summary augmentation selection: known — identifier [wired], coarse [wired] (Triggers: later model call or augmentation. Inputs: current resolved conversation identity and summary presence; selected parts: matching history rows or summary/latest pair.); Benchmark question-triggered memory selection: partial — inferred-embedding [wired], inferred-lexical [wired], coarse [wired] (Trigger: benchmark question. Push selector inputs: question text/embedding and limits; selected parts: saved metadata memories and unique session summaries. Conversation ID belongs to requested per-conversation index access, not an established identity match selecting delivered memory parts. Dense, lexical and coarse selection are supported; records do not establish a separate identifier-based push selector or warrant a complete signal inventory.); Intentional host/MCP return operations only: inapplicable (These requested returns are pull; push-signal classification does not apply to this unit.); Cloud SDK internal selection and other uninspected consumer branches: not-determinable (Client query/filter transmission does not establish the service selector; no inferred-judgment or complete signal inventory is established.)	partial coverage (Automatic trace-fed durable fact/summary production has a later context consumer. That architectural channel is not observed improvement; opaque hosted and native alternatives cannot inherit complete classification.); Ordinary SDK conversation-to-derived-facts/summaries-to-later-provider chain: known — yes [wired] (Completed interaction messages feed automatic retained outputs and later wired model-context delivery. Server semantics remain opaque without negating this client-mediated chain.); Python recursive summary update chain: known — yes [wired] (Existing summary plus latest turns feed subsequent durable memory; Rust summary=None is a distinct route.); Hosted host-agent memory and resume generation: partial — yes [claimed] (Trace submission and requested delivery are wired in clients, but private derivation/retention and actual resume fields are unavailable.); Native mock/fallback, graph/process independent consumer routes, benchmark import ancestry and feedback adaptation: not-determinable (Mock/content origin, separate later consumers, upstream benchmark extraction and feedback updates are unresolved. Raw persistence/report creation alone does not establish qualifying learning; no system-wide no is inferred.)	partial coverage (Authority is assigned to actual part/consumer/effect paths, not all outputs of augmentation. Opaque hosted outputs and uninspected graph/process/feedback consumers prevent completeness.); Replayed conversation, recalled facts and joined summaries at provider input: known — knowledge [wired] (Delivery gives contextual force at the next inference, not enforced truth or reliable obedience.); Retained fact text, vectors and access counters at recall selectors: known — ranking [wired] (Text lexical scores, vector similarity and candidate recency/count ordering influence relative priority.); Prior Python conversation summary at next augmentation: known — learning [wired], routing [wired] (Summary text feeds subsequent durable derivation; its presence also selects the latest user/assistant pair. Rust sends no prior summary.); Benchmark retained metadata/index at retrieval and answer model: known — ranking [wired], knowledge [wired] (Index and prose have distinct consumers and effects.); Hosted recall/resume results at requesting foreground agents: partial — knowledge [wired], instruction [claimed] (Result delivery is wired; advertised standing-order fields and their actual contents are not inspected. Static shipped skills are excluded as accumulated memory.); Graph rows, process attributes, retained evaluation reports and feedback downstream: uninspected (Storage/report output and feedback transmission are established, but later readers or adaptation consumers are not inspected; no enforcement, validation or learning authority is inferred.)	AAS-2026-10-17-memori-fd4a0418c09b-01
MemoryOS	2	Accumulated recent dialogue, sessions/pages and summaries/overviews; user profiles and user/assistant facts; their dense indices and symbolic maintenance/access state; ChromaDB collections and separately persisted metadata; bundled MCP and Playground admission/read/clear routes; evaluation sample memories and retained answer/reference ledger. Includes initial-loaded/direct-update provenance uncertainty, uninspected ChromaDB index internals and remaining shipped driver/helper consumer coverage. Excludes shipped static prompts/source as memory, transient personality-display output, external provider/encoder weights and caches, and external MCP-host internal storage/behavior.	partial coverage (The recorded operative inventory establishes files, process collections and vector access. Dependency internals remain unresolved; external model weights and static source repositories are not accumulated memory substrates here.); JSON recent dialogue deque and session/page, profile, user/assistant knowledge stores, including MCP and Playground bundled JSON paths: known — files [wired], in-memory [wired] (These operative objects are retained as JSON files and reloaded into process collections; files and process state are separate substrate findings.); JSON access structures: stored vectors, counters, flags, rebuilt FAISS indices and heat heap: known — files [wired], in-memory [wired], vector [wired] (Dense encoding is not by itself a vector substrate; the operative FAISS index supplies that access substrate.); ChromaDB collections and separately saved short-term/profile/session metadata: known — vector [wired], files [wired], in-memory [wired] (Persistent vector collections and process metadata with JSON close/exit persistence are distinguished. No dependency-internal SQLite or other physical substrate is inferred.); Evaluation sample memory, selector state and retained answer/reference results: known — files [wired], in-memory [wired] (Sample JSON memories and result files are distinct from product stores; operative lists and access state exist in process.); ChromaDB dependency-internal retained index details: uninspected (Vector collection existence is resolved above; internal organization was not inspected and cannot establish additional substrates.)	partial coverage (Automatic admissions are supported on shipped transformations. Initial-content and alternative-caller control remain unresolved; no manual finding is inferred from an API call or a workflow start.); Generated dialogue, overflow migration and profile/fact proposals under orchestrator admission: known — automatic [wired] (Initiating the request or forcing analysis is not per-content approval.); MCP add and Playground import admission after supplied-content checks: known — automatic [wired] (Software admits supplied pairs; the records do not identify the upstream supplier as a human approving each retained item.); ChromaDB generated content, metadata and surviving-backup derivation: known — automatic [wired] (Code/model admission is implemented despite separate durability windows and the new-session backup defect.); Evaluation import, derivation and result-file writes: known — automatic [wired] (Configured benchmark execution does not make each derived item manually approved.); Access metadata changes following retrieval and analysis: known — automatic [wired] (Counters, heat and analyzed markers change under fixed code.); Direct setters/alternative updater and initial JSON/ChromaDB contents: not-determinable (Caller identity and initial filesystem provenance do not identify explicit human content decisions; neither manual nor automatic initialization control is established.)	partial coverage (Every established response/tool/display/scorer route is requested pull. Remaining included demonstration/helper coverage prevents a whole-boundary pull-only conclusion; automatic formatting alone adds no push.); Requested JSON response context retrieval and delivery: known — pull [wired] (A foreground response query requests retained context as part of generation; selection and formatting are fulfillment, not independent unsolicited supply.); Requested ChromaDB response retrieval and attempted backup enrichment: known — pull [wired] (Missing initial-page backups limit enrichment, not direction; surviving vector text is returned on request.); MCP retrieve/profile tool return and Playground state/profile display: known — pull [wired] (The host/browser explicitly asks for retained material; actual host model use is separate and unresolved.); Evaluation question answering and offline result scoring: known — pull [wired] (Question requests retrieve sample memory; the separate scorer explicitly reads the retained result file.); Additional consumers in remaining demonstration drivers and uninspected included helpers: uninspected (Records establish the named routes but not an exhaustive inventory of every shipped demonstration/helper consumer. No additional push can be asserted or ruled out.)	uninspected coverage (No positive push signal is established. The inspected request chains are inapplicable locally; remaining consumer coverage is uninspected, so no system-wide inapplicability or evidenced absence is asserted.); Assessed pull-only JSON, ChromaDB, MCP, Playground and evaluation request chains: inapplicable (Signals classify push selectors only. Query embeddings and IDs on these requested reads do not establish push signals.); Remaining included demonstration/helper consumer alternatives: uninspected (The accepted inventory does not inspect every such consumer; direction and any push selector remain uninspected. A known pull witness cannot make this branch inapplicable.)	partial coverage (Qualifying automatic conversational derivation is wired. The result ledger lacks a qualifying positive or canonical bounded-absence warrant; unknown initial/direct/extra-driver production remains included. No local or system-wide no is asserted.); JSON dialogue-fed summary/continuity production and profile/fact updates: known — yes [wired] (Automatic writes persist behavior-shaping summaries/profiles/facts read by later selection, update or response calls; this does not establish improved capacity.); ChromaDB dialogue-derived summaries and surviving merged-backup profile/fact updates: known — yes [wired] (New-session initial pages cannot feed the backup-only extractor. Summary selection and later merged-page derivation still qualify; no intact initial-backup chain is implied.); Evaluation imported conversation-to-summary/profile/fact route: known — yes [wired] (Imported dialogue retains its session origin; benchmark answers/results are a separate ledger rather than feedback adaptation.); Benchmark answer/reference result ledger to offline scorer: not-determinable (The recorded main/scorer route saves and compares answers, but supplies no canonical bounded-absence record resolving trace learning for this ledger. It establishes no qualifying behavior-shaping update and cannot establish controlled no.); Initial loaded state, alternative direct-update producer and uninspected extra driver routes: not-determinable (Upstream producer/trace input and complete extra-driver wiring are not established; these cannot inherit the known positive derivation route.)	partial coverage (Authority is localized to actual consumers, including update learning, downstream guidance, ranking and control eligibility. Host reliance remains unresolved; no static prompt/schema guard is classified as accumulated-memory authority.); Recent/history pages, overviews and user facts at response model: known — knowledge [wired] (These are evidence/background context, not local truth checks or enforced commands.); User profile and assistant statements at response model: known — knowledge [wired], instruction [wired] (Trait/preference text supplies standing style/persona guidance as well as reference content; message position does not establish compliance.); Retained dialogue and old profile at summary/profile/fact update consumers: known — learning [wired] (The update consumer and the later answering consumer are distinct authority paths.); Retained summaries, page/fact content and embeddings at similarity ranking consumers: known — ranking [wired] (Fixed ranking code does not remove the ranking authority of retained inputs.); Access/heat/analyzed state at LFU, heat selection and analysis eligibility consumers: known — ranking [wired], routing [wired], enforcement [wired], validation [wired] (Counters choose survival and next analysis input; analyzed flags exclude previously processed pages. The flags supply eligibility criteria, not factual validation.); ChromaDB dialogue, profile and fact text at response consumers: known — knowledge [wired], instruction [wired] (Vector dialogue and retained profile/facts supply context and standing guidance at the response model, independently of maintenance eligibility.); ChromaDB vectors, summary keywords and access state at retrieval and retention priority consumers: known — ranking [wired] (Retained representations and access state affect relative retrieval and retention priority under fixed code.); ChromaDB retained heat and surviving-backup analyzed flags at maintenance input selection and eligibility consumers: known — routing [wired], validation [wired], enforcement [wired] (Surviving merged backups retain analyzed flags consumed by the extraction filter. These flags supply both an admissibility criterion and an enforced exclusion, not factual validation. New-session backup loss does not remove this surviving branch.); ChromaDB surviving backup dialogue and old profile at durable profile/fact update consumers: known — learning [wired] (Retained dialogue and prior profile feed durable behavior-shaping updates when backup inputs survive. This update authority is separate from the output's response-context authority and from the flags' eligibility veto.); Evaluation sample memory at answer and update consumers: known — knowledge [wired], instruction [wired], ranking [wired], routing [wired], learning [wired] (Benchmark memory is not the product algorithm. Full profile/all assistant content and query-selected user/history material have distinct delivery paths.); Retained benchmark answer/reference ledger at offline scorer: known — validation [wired] (References define the lexical conformity target and predictions supply the candidate; the score does not veto memory writes.); MCP return at external host and browser display at operator: not-determinable (Delivery is wired, but downstream host channels and operator action are not established; a lossy display cannot classify the consumed profile as an enforced instruction.)	AAS-2026-10-16-memoryos-bb72a1b70e03-01
MemPalace	2	Included: use-accumulated drawer text/access metadata/embeddings, closet locators, diary/checkpoint content and ingest state, editable identity, setup entity/origin state, temporal KG rows, tunnel/hallway navigation and proposal state, media references and original externally retained bytes, logstream tasks/events/artifacts, write audit state and backup/replication/recovery alternatives, plus unresolved provider request retention. Excluded: static shipped skills/prompts/configuration and pretrained model weights as computation, transient queues/search tags/audit reports, benchmark-only reporting without an established operational memory consumer, and the external host/model implementation. Shipped host adapters and their later-context routes remain included.	partial coverage (The inventory includes drawer/index collections, sidecar and identity files, temporal and coordination rows, proposal/WAL state, and unresolved backend, recovery and provider-retention alternatives. Known stores do not resolve those alternatives.); Drawer text, metadata, diary specializations, closets and their dense indexes in the inspected collection route: known — vector [wired] (The Chroma collection route retains document/access parts in a vector store; dense vectors are not model weights.); Temporal relationship rows and logstream task, patch and acknowledgment retention: known — sqlite [wired] (These independent retained objects have explicit SQLite storage.); Identity, setup origin/entity files, original daily files and ingest state, tunnel sidecar, proposal plan and write audit log: known — files [wired] (The records identify editable text, JSON/sidecar/plan and JSONL retention rather than merely in-process values.); Identity Layer0 instance cache and host wake-context cache retained across later prompt assemblies: known — in-memory [wired] (These are later-consumed cached copies of retained material, not durable filesystem replacements or transient worker queues.); Hallway sidecar lifecycle, non-Chroma backend variants, native index persistence and backup/export/migration/replicated copies: uninspected (Adapter names and exposed repair/replication capabilities do not establish complete physical retention or replacement semantics; remote/native and recovery internals remain incompletely inspected.); External provider request retention or other use-accumulated provider state: not-determinable (The clients supply content to providers, but request retention is uninspected. Pretrained provider weights are excluded static computation, not classified as accumulated memory.)	partial coverage (Admission is classified per mechanism, not physical I/O or authorship. Known automatic writes coexist with human-reviewed setup and claimed approval policies. Generic caller/initial-file control and alternate maintenance/capture mechanisms remain unresolved.); Hermes turns, hook transcript filing, corpus segment admission and media indexing: known — automatic [wired] (A host trigger or operator starting import does not approve each retained item.); Silent checkpoint, heuristic/LLM closet generation, automatic setup topics and tunnel traversal counters: known — automatic [wired] (These exact branches admit derived material without a per-content human decision.); Interactive people/project entity confirmation: known — manual [wired] (This branch permits human editing/approval of the detected retained entity content.); Noninteractive --yes setup admission: known — automatic [wired] (Bypassing item review does not convert subsequent classifier output admission into human content approval.); User-edited tunnel proposal followed by explicit application: known — manual [wired] (The proposal orchestration exposes retained rows for deletion/editing before operator adoption; underlying generation/prune algorithms remain outside this unit.); Identity authoring, audit restructuring and exact normalized task approval under shipped operating doctrine: known — manual [claimed] (These policies require a human content decision but are not enforced by the file reader/write handlers/task constructor.); Explicit drawer/diary/KG writes and replacements, task/patch/ack submissions, initial daily-file/identity content: not-determinable (Software structural checks are known, but generic client identity does not determine whether a human supplied/approved each retained content item. Automatic physical persistence alone cannot settle admission control.); All legacy/prompted capture variants, graph prune/generation controls, backup/migration/replication and provider-retention alternatives: uninspected (Known branches do not establish every alternate admission mechanism; prompted agent summaries remain requests, not demonstrated writes.)	partial coverage (Requested recall and independent startup supply coexist. Each request/selection/delivery chain is classified once, not every return as push. Legacy host, recovery/provider and additional coordination consumers remain incompletely inspected.); MCP/CLI search, host-requested Hermes prefetch and direct logical drawer GET: known — pull [wired] (The later consumer requests retained material; ranking/hydration and delivery fulfill that request.); Requested diary continuity, entity/time KG query, tunnel follow, media query and coordination event/artifact reads: known — pull [wired] (These are distinct request-driven later-consumer routes; documented worker reload is not proof of execution.); DSH independently triggered session-start wake-up prompt providers and Hermes startup context cache: known — push [wired] (Startup code selects and supplies retained identity/snippets without a consumer request for those specific stored contents; an explicit CLI wake request is separately pull.); Explicit CLI wake-up: known — pull [wired] (Requested wake content is not a second push because code delivers it.); Exposed Layer2 filtered API without an established independent later consumer: not-determinable (The filtered-read capability is known, but its unspecified caller does not establish an additional read-back route or direction.); Other legacy wake/autosave host variants, diagnostic/recovery reloads, opaque provider retention and uninspected event/artifact recipient workflows: uninspected (The accepted records do not establish all later consumer routes or their request boundaries. Known push/pull paths do not close this inventory.)	partial coverage (Only unsolicited startup selection is positively classified. Semantic and lexical queries are pull, not inferred push signals. Opaque/legacy startup alternatives remain unresolved and cannot be made inapplicable by inspected requested readers.); Startup Layer0/Layer1 context selection supplied by DSH/Hermes: known — coarse [wired] (Trigger is session startup/cache refresh; inputs are identity availability, optional wing, a recent candidate window, metadata weights/filing time and size budget; selected parts are full identity plus up to fifteen truncated drawer snippets.); Inspected requested search, prefetch, diary, KG, tunnel, media and coordination reads: inapplicable (These units fulfill later consumer requests. Their embeddings, lexical matches, file IDs or correlation IDs do not establish push-selection signals.); Additional legacy host wake/recall adapters and provider/recovery or other recipient supply alternatives: uninspected (The records leave those routes incompletely inspected; their unsolicited selectors and delivered parts cannot be inferred from a known startup branch.)	partial coverage (Automatic durable continuity and trace-fed locator writes qualify without observed improvement. Raw transcript retention alone does not. Counters with only navigation/audit display and unresolved caller/provider/host/graph alternatives remain local uncertainties, not a system-wide no.); Silent hook transcript-prefix checkpoint to retained diary to later continuity/search consumer: known — yes [wired] (The automatic write produces a durable compact continuation artifact from session content; later diary/search consumption is recorded.); General session extraction and heuristic/LLM locators on session-derived drawers to later recall/ranking: partial — yes [wired] (This unit includes the trace-fed branch, not every source group: imported project files can be non-traces. New knowledge, learned weights and benefit are not required.); Raw-ish Hermes exchange and transcript retention without derived continuation/update: inapplicable (This isolated raw-retention part supplies future information access but no qualifying derived behavior-shaping update. General extraction and checkpoint/locator branches are classified separately.); Traversal counters/strength with later audit/navigation display: not-determinable (Automatic event-fed persistence is known, but the records do not establish those counters changing downstream selection or durable behavioral guidance. Mere trace retention/display does not decide a qualifying learning route.); Prompted or direct summaries, caller KG/content revisions, outcome drawers and initial daily summaries: not-determinable (Caller control, actual trace inputs and/or completion of the derived write are unresolved. A requested agent summary or prescribed outcome drawer cannot be upgraded to a wired qualifying write.); Full hallway/tunnel graph maintenance, legacy host, recovery/native/backend and opaque provider update alternatives: uninspected (The accepted records do not enumerate every trace-fed write and later behavior-shaping consumer; static embedding weights are excluded, not a negative trace-learning witness.)	partial coverage (Each finding names retained part, consumer and effect. Context delivery is knowledge or guidance, not observed compliance. Index/update consumers differ from answering consumers. Uninspected graph, provider, recovery and coordination alternatives prevent complete coverage.); Drawer/diary text, KG assertions and media descriptors at later recall/fact/reference consumers: known — knowledge [wired] (The consumers receive selected content or references, not independently accepted truths or necessarily full payloads.); Identity text at startup host guidance: known — instruction [wired] (The host-facing identity channel is implemented; actual host insertion and compliance remain unobserved.); Retained task directions at the later worker named by the task protocol: known — instruction [afforded] (The stored body has a documented later instructional consumer; the pointer handoff does not establish completed body reload.); Drawer and closet words, stored vectors and weighted/dated access metadata at retrieval/startup selection: known — ranking [wired] (Fixed software scoring still gives its retained inputs ranking authority; no learned scoring policy is required.); Drawer source/chunk fields, entity registry, tunnel endpoints and task/event routing fields at fixed selectors: known — routing [wired] (These retained fields choose filing scope, reconstructed input, navigation destinations and coordination recipients.); Retained IDs/freshness metadata and drawer/vector content at duplicate/admission checks: known — validation [wired], enforcement [wired] (Retained state supplies structural/idempotency criteria or approximate similarity, not a truth oracle.); KG validity intervals at explicit as_of/current-category consumers: known — routing [wired], enforcement [wired] (Interval fields affect eligibility or temporal grouping while history survives; reused successor visibility remains conditional.); Session trace material at automatic checkpoint and trace-fed locator update consumers: known — learning [wired] (This is the update-side force of retained experience, separate from output knowledge/ranking. No improvement is inferred.); Tunnel strength/decay beyond displayed history, opaque media payload use, additional provider/backend/recovery and event/artifact consumers: uninspected (Records establish counters and references but not default-ranking influence, payload interpretation, full recovery effects or all coordination consumers. WAL diagnostics are afforded, not an automatic model-memory path.)	AAS-2026-10-14-mempalace-ed71875aea5a-01
MemWiki	2	Includes mutable hot/index/domain/decision/log wiki files accumulated or changed through use; human-supplied raw originals, including unresolved payloads; current and legacy starter admission; foreground-agent work/raw/repository writes; lint proposal-to-edit uncertainty; fold; and later session/source/maintenance readers. Excludes static shipped protocol/hooks as memory content, external host parameters/context/compaction, npm/dependency internals and untracked pages merely claimed in the log. Static instructions establish the included routes but are not themselves accumulated memory.	known coverage (All included retained parts are project files. External host context/parameters and static hook/protocol installations are excluded, not alternative internal memory stores.); Mutable hot cache, index, stack/patterns/bugs, decisions and work log in the project wiki: known — files [wired], files [afforded] (The complete wiki inventory shares filesystem retention; neither Markdown encoding nor links establish a database or graph substrate.); Human-supplied raw originals and their directory: known — files [wired], files [claimed] (Payload encoding and initial deployed content are unresolved, but their declared retention substrate is files.)	partial coverage (All specified write paths are retained, including the unresolved lint-to-edit transition. No arbitrary manual wiki editing is added merely because files are editable.); Seeding absent wiki pages through either scaffold executable: known — automatic [wired] (Operator launch and cwd selection do not turn fixed-template admission into manual content control.); Human delivery of raw original files: known — manual [claimed] (This is source admission, not approval of the subsequent wiki sentences.); Agent discovery/raw synthesis, repository ingestion, hot refresh and work-log append: known — automatic [claimed] (Launching ingestion is not approving its results. Actual execution and external host vetoes remain unobserved.); Older-history replacement by fold: known — automatic [claimed] (The admission mechanism is separate from the read selecting older entries.); Admission of lint proposals as retained corrections: not-determinable (Protocol says propose; README says fix. No required edit transition or approving principal is established, so neither automatic nor manual correction admission can be concluded.)	known coverage (All included later-consumer routes are requested file reads. Static host-hook discovery is not independent memory push; unknown external host behavior is not an included alternative.); Session-entry hot reads and conditional index/domain/history reads: known — pull [claimed] (The recorded route covers hot-first and conditional wiki access; external host injection/compaction is excluded.); Raw original rereading for source use and synthesis: known — pull [claimed] (There is no recorded watcher or unsolicited original-payload supply in the included mechanism.); Retained older log input to fold and wiki input to lint: known — pull [claimed] (Requested material delivery is not another push operation. Unknown lint-result admission does not create an alternative input direction.)	inapplicable coverage (Inapplicability is bounded to the completely classified pull-only MemWiki file-read routes, not to external host context machinery.); All scoped session, source-rereading and maintenance read-back operations: inapplicable (Every included direction unit is requested pull. File identifiers and index entries select requested reads, not unsolicited delivery; therefore no push selector needs a signal classification.)	partial coverage (Claimed work-derived continuation and fold routes establish positive existence. Raw payload and lint-admission uncertainties remain local. Repository-reference acquisition is a nonqualifying branch; no local or system-wide no is asserted without a bounded absence warrant.); Work/task experience to hot continuation context and timestamped work summaries, then later coding readers: known — yes [claimed] (This qualifies at claimed route level even without an automatic MemWiki model call or demonstrated improvement. Actual trace production and read-back remain unobserved.); Retained summarized work history to folded summary, then conditional later history reader: known — yes [claimed] (Already-derived experience can feed another qualifying update; no raw transcript is asserted.); Pure repository-source ingestion into project reference pages: inapplicable (This unit acquires/compiles repository reference material rather than a qualifying trace-fed write. Inapplicability concerns that source-defined branch, not bounded absence of learning elsewhere.); Human-supplied raw originals to wiki and later project readers: not-determinable (No concrete payload determines whether imported material is static reference or recorded experience such as meeting activity. A qualifying trace-fed alternative cannot be ruled in or out.); Lint-generated corrections, if admitted: not-determinable (No recorded input/evaluation experience, correction admission and durable later-consumer chain resolve a qualifying trace update.)	known coverage (The included consumers and effects are recorded across the complete wiki/raw inventory. Static protocol criteria and external host controls are outside this memory scope; no ranking, validation or enforcement force is inferred from priority language or lint proposals.); Hot state and next steps at the later foreground coding agent: known — knowledge [claimed], instruction [claimed] (Context and next-step directions are distinct effects; host compliance remains unobserved.); Wiki index at the agent choosing additional pages: known — routing [claimed] (Navigation does not approve facts or impose a permission check. The outgoing-link count conflict does not change this force.); Stack/patterns/bugs and decisions at the task-dependent coding reader: known — knowledge [claimed], instruction [claimed] (Conventions and advisory rationale do not become executable enforcement. The bug placeholder establishes no actual bug-learning instance.); Work log at later history readers: known — knowledge [claimed] (Reading the log is conditional, unlike the hot-cache obligation.); Retained work history at the fold update consumer: known — learning [claimed] (Update-consumer learning force is distinct from the folded output's downstream contextual force; no improved capacity is demonstrated.); Raw originals at source-reading and synthesis consumers: known — knowledge [claimed] (Unknown payload form and source truth do not prevent classification of this declared consumer role.)	AAS-2026-10-14-memwiki-79dcc4d04bcb-01
MentisDB	2	Retained thought content/envelopes (including Python and evaluator conversation writes), uploaded skill bodies/version/lifecycle state, managed embedding vectors and symbolic retrieval access structures, caller-retained Markdown views and evaluator/interaction diagnostics, with their admission, curation and later-consumer routes. Includes unresolved custom adapters, backend detail, caller-controlled semantic updates and diagnostic/log consumers. Excludes static compiled core/project instructions, product binaries, inference-provider weights, transient extraction/source-window/bundle responses as retained objects, and current append-event delivery as historical read-back. External host/provider internals remain outside the frozen boundary; unknown callback retention is retained as a limitation, not attributed to MentisDB.	partial coverage (Thoughts, uploaded skills and derived access state have positive substrate witnesses. Custom storage, backend detail and diagnostic/log lifecycle keep the included inventory partial. External model weights are inference dependencies, not memory accumulated by these routes.); Canonical thought content and symbolic envelope; built-in binary and JSONL chain persistence: known — files [wired], in-memory [wired] (Built-in persisted and loaded parts are identified; custom adapters are assessed separately.); Uploaded skill bodies, version patches and lifecycle registry: known — files [wired] (This is accumulated uploaded state, not the compiled core skill.); Managed vectors, retrieval indexes and semantic-neighbor overlay: partial — vector [wired], files [wired], in-memory [wired], graph [wired] (Full HNSW/WAL persistence and metadata freshness are incompletely inspected; file and vector witnesses do not resolve every backend.); Caller-saved MEMORY Markdown view: known — files [afforded] (Export itself is transient; file retention is only afforded.); Evaluator diagnostic outputs: partial — files [wired] (Complete output lifecycle and subsequent readers are not established.); Custom canonical storage adapters and optional service interaction-log persistence: uninspected (Custom storage semantics and complete log rotation/reload consumers are not inspected; substrate coverage cannot be closed.)	partial coverage (Automatic message/derivative mechanisms are known. No manual finding is inferred from a CLI/UI or an agent ID; caller-controlled alternatives keep agency partial.); Python message admission and benchmark per-turn ingestion: known — automatic [wired] (Starting the driver or configuring an adapter does not make each admission manual.); Managed vector/index/neighbor updates and optional similarity supersession: known — automatic [wired] (Configuration/maintenance initiation is distinct from per-output admission.); CLI/service/direct caller thought, correction, checkpoint and handoff admissions: not-determinable (Generic caller attribution does not determine whether a human supplied, approved or replaced each retained item.); Skill upload/revision/lifecycle, Markdown import, agent-copy, merge/restore and deletion: not-determinable (Explicit request, dashboard access, signature or structural parsing is not proof of a per-content human decision. Automation versus human content approval is unresolved on these caller-controlled paths.)	partial coverage (Established accumulated-memory return is pull. Current-event webhook dispatch is excluded as a read-back finding, but its unknown external later consumer and diagnostic/log alternatives prevent a complete pull-only inventory.); Requested thought recent/ranked/flat/federated/addressed/history access, bundles and Markdown export: known — pull [wired] (Incomplete exact/history filtering prevents a uniform visibility claim, not classification of these request-driven routes.); Requested uploaded skill/version read and derived search-state use: known — pull [wired] (An identifier in a request is addressing, not identifier-based push.); Later consumption of optional interaction logs, diagnostic outputs and any callback-owned historical memory: uninspected (Complete later file-reading consumers and external callback retention are uninspected. The wired webhook only pushes the current append event, not selected accumulated history.)	uninspected coverage (No controlled push-selection value is established. Unresolved later-consumer alternatives prevent system-wide inapplicability; observed current-event webhook selection cannot classify a retained historical payload.); Inspected requested thought/skill returns and request-driven derived retrieval: inapplicable (These completely characterized direction units are pull operations; ranking/query selectors are not push signals.); Potential later diagnostic/log or callback-owned historical-memory supply: uninspected (There is no established push of accumulated memory in these records and no complete later consumer inspection. Current-event chain/type matching is not a historical read-back selector.)	partial coverage (A wired trace-fed numerical derivative witness coexists with unresolved semantic and diagnostic update routes. No local or system-wide no is asserted without bounded absence evidence.); Managed embeddings and semantic-neighbor update from retained raw conversation, then later search: partial — yes [wired] (The qualifying witness is durable behavior-shaping retrieval machinery, not raw-log storage alone. Other nontrace input text and metadata are included, and their origins are unresolved. No semantic lesson, weight update or improved task capacity is implied.); Caller-authored lesson/summary/checkpoint and uploaded instruction revision: not-determinable (Durable behavior-shaping content and later return are afforded/wired, but per-content human versus automatic admission and qualifying original trace inputs are unresolved.); Diagnostic/log-driven later updates and callback consumers: uninspected (Inspected diagnostics have no traced automatic successor-policy consumer; no bounded absence record closes every included route, and complete later consumers are uninspected.)	partial coverage (Knowledge/instruction delivery, ranking, routing and structural vetoes have local consumers. The trace-fed numerical update has a distinct learning consumer. Downstream compliance and unresolved diagnostics are not upgraded by those witnesses.); Returned thought content and handoffs at human/model context readers: known — knowledge [wired], instruction [wired] (Wired delivery supports available forces at the interpreter, not observed obedience or task improvement.); Uploaded skill bodies at downstream interpreters: known — instruction [wired] (Returned revoked content is not itself an enforced execution prohibition.); Thought text/metadata and numerical vectors at ranking code: known — ranking [wired] (Ranking is a force of retained inputs under fixed code, not a learned policy claim.); Retained invalidation-ID set at default retrieval predicate: known — enforcement [wired] (Audit, direct lookup and unfiltered renderers bound the veto.); Retained thought sequence/reference identities and canonical hashes at append/open checks: known — validation [wired], enforcement [wired] (Structural force does not license content truth, authorship or crash-proof storage.); Retained Summarizes targets at source-window selection: known — routing [wired] (Target membership routes proposed work; it does not evaluate summary adequacy.); Automatically compiled embeddings from retained conversation content at the update consumer: partial — learning [wired] (This learning force is the numerical update path, not semantic lesson generation, parameter training or demonstrated capacity improvement. Actual inputs may instead be nontrace caller text.); Unspecified live skill/status interpretation and later consumers of diagnostic/log files: not-determinable (No complete subsequent diagnostic/log reader or live host status-handling path is established; these parts cannot acquire additional forces by inference.)	AAS-2026-10-14-mentisdb-a8b5903ae3e2-01
MerchantBench	2	Includes mutable run-local latest Markdown and its version archive, ReAct accumulated message history and maintenance, persisted agent traces and human inspection, and visible Hermes profile/session storage with unresolved external memory/update/consumer alternatives. Excludes static shipped scenario guidance and tool grants, ordinary simulator/business state, unchanged recovery/idempotency checkpoints, terminal evaluation reports without an agent feedback route, and remote provider weights not established as through-use memory. Hermes external algorithms remain included as unresolved alternatives, not inferred from displayed summaries or counters.	partial coverage (Document, archive, foreground history, persisted traces and visible Hermes storage are classified. External Hermes native stores and opaque consumed parts remain unresolved.); Latest document and full-version archive: known — files [wired] (Both retained bodies are UTF-8 filesystem files.); ReAct accumulated message history: known — in-memory [wired] (The inspected baseline keeps a Python list, not a restored disk history.); MerchantBench per-step trace artifacts: known — files [wired] (JSON encoding does not change the file substrate.); Hermes run-local home and externally authored session database visible to MerchantBench: partial — files [wired], sqlite [wired] (Files and the read-only SQLite session adapter are visible; they do not enumerate external native memory stores.); Other external Hermes retained parts: uninspected (Native memory storage and opaque alternative payloads are not inspected; known profile and display storage cannot resolve them.)	partial coverage (Admission differs across foreground software, explicit human edits, raw capture and external initialization. Generic API caller identity and external updates remain unresolved.); Foreground model document replacement and archive append: known — automatic [wired] (No human approves each model-selected body; optional invocation does not make software admission manual.); Human editor save of the same document: known — manual [wired] (The save button supplies the textarea content as an explicit per-content human decision.); Other generic document callers and initialization: not-determinable (The generic API and an unobserved initial payload do not establish whether a human supplies or approves each content change.); Foreground history admission and server raw trace capture: known — automatic [wired] (These mechanisms record messages without a per-message human admission decision.); Hermes initial profile copy and configuration preparation: known — automatic [wired] (Operator installation provides the source files; run-specific admission is a software copy, not per-skill human approval.); External Hermes native-memory, summary and skill updates: uninspected (Actual external admission controls and checkpoint replacement routes are not inspected.)	partial coverage (Requested document/trace delivery and automatic foreground history selection are distinct operations. External model-context read-back remains uninspected.); Explicit latest-document requests, including human editor reads: known — pull [wired] (Identity addressing fulfills the request; the tool return is not an independent unsolicited push.); Human historical trace requests and visible Hermes trace adapter: known — pull [wired] (A chosen step/all-trace request returns retained material to a viewer, not automatically to an agent.); ReAct selected history at automatic subsequent request construction: known — push [wired] (This is the automatic reuse of retained context, not a second direction assigned to the requested document return itself.); Archive markers at later version append: known — pull [wired] (The later writer explicitly reads the archive to count markers; no historical-body reload is established.); External Hermes actual model-context and skill read-back: uninspected (Active/compacted display filtering is not evidence of what its runtime delivers to a model.)	partial coverage (The known push uses recency and budget, not task relevance or an identity match. Pull routes do not contribute push selector values; external push alternatives remain unresolved.); Foreground historical suffix automatically supplied on a later model request: known — coarse [wired] (Trigger: request construction/threshold maintenance. Inputs: message order, roles and estimated or reported size. Selected part: newest surviving history, with leading orphan tool messages removed.); Explicit document and trace requests, plus archive marker read: inapplicable (These units fulfill requested reads. Run/agent/step identifiers address those requests; they do not establish identifier-based push.); External Hermes automatic retained-content selectors: uninspected (The external selectors and their inputs are not inspected. Display filters cannot classify model-context push signals.)	partial coverage (The local trace-to-document-to-later-model chain is afforded, not exercised or mandatory. Raw capture alone is not trace learning, and external writes remain unresolved.); Foreground experience saved as durable document and later explicitly recalled: known — yes [afforded] (Model/software admission is automatic if the foreground agent takes the offered save route. Persistence and a later context consumer exist, but successful save/recall in a real run is unobserved.); Raw history/trace capture, archive duplication and suffix eviction considered alone: not-determinable (These records establish raw retention, copying and filtering, not a separate qualifying durable derived artifact. They do not supply a formal bounded-absence record for a local no finding; the qualifying document route is assessed separately.); External Hermes summary, skill and native-memory writes: uninspected (Configuration and evolution/compaction counts do not establish or exclude an automatic trace-fed durable behavior-shaping write.)	partial coverage (Knowledge and conditional learning paths are supported locally. No instruction, permission, ranking or validation force is inferred merely from a memory filename or from static runtime controls.); Latest document at foreground model and human editor: known — knowledge [wired] (The text supplies notes/planning advice through a requested tool result or textarea, not standing system guidance or an executable grant.); Retained foreground observations/actions/results at later model calls: known — knowledge [wired] (The ordinary consumer receives selected historical context; no expected-answer oracle is created.); Foreground trace material at optional durable-note update: known — learning [afforded] (This is the afforded update consumer, separate from the resulting document's downstream advisory force.); Persisted trace text at human inspection: known — knowledge [wired] (Viewing raw recorded material supplies inspection evidence, not automatic strategy acceptance.); Archive version markers at the later archive writer: not-determinable (The established consumer counts markers to construct the next version number. This metadata computation establishes no controlled authority value; it selects no operation, destination, scope or input window. No historical-body restore consumer is inferred.); External Hermes retained skills, summaries and native memory at external consumers: uninspected (Launch settings and displayed session text do not establish actual external authority paths. The static configuration itself is not learned memory.)	AAS-2026-10-15-merchantbench-d2f4d78dbe37-01
meta-harness	2	Use-derived classification examples/checkpoints/logs/local completion replay; reference results/frontiers, separately assessed summary result and rationale parts, requested reports and retained candidate code; KIRA task-local continuation summaries/fallbacks and opaque inherited/provider alternatives; pilot source snapshots and evaluation history with separate model-feedback and final-score consumers; afforded domain-specification handoff. Includes unresolved generated-state and arbitrary initial-checkpoint control. Excludes external model weights, static datasets/seed prompts/skills/suite/verifier code as memory objects, ordinary untransformed conversation/tool accumulation, and unfrozen external run artifacts. Provider cache state is retained as an opaque included alternative, not equated with local response files.	partial coverage (The inventory includes live state, local durable files, afforded documents and opaque alternatives. Positive local file storage does not enumerate provider or generated-candidate internals.); Live FewShotMemory examples and epoch best_state: known — in-memory [wired] (Live list and best-state capture are distinguished from disk transport.); Classification checkpoints, logs, local response cache, reference histories, result files and candidate code: known — files [wired] (These named local artifacts are files; JSON and Python describe encoding, not an additional substrate.); Pilot history and source/result snapshots: known — in-memory [wired], files [wired] (Tuple accumulation and persistent snapshots/report are separate operative parts.); KIRA handoff chat and trajectory structures: partial — in-memory [wired] (Local chat/trajectory retention is inspectable; broader Harbor persistence is external.); Requested report and onboarding documents: known — files [afforded] (Classification reports and domain specification are afforded file artifacts, not observed instances.); Included generated-candidate internal state and opaque provider/host retained parts: not-determinable (Generated payload internals and provider-cache substrate/lifetime are unresolved; disk checkpoint transport does not determine every internal substrate.)	partial coverage (Admission control is separate from authorship and physical copying. Known automatic transformations do not settle arbitrary checkpoint initialization.); Batch admission, best-epoch restore, checkpoint/log/cache writes, frontier/summary admission and pilot history: known — automatic [wired] (Starting the workflow is not per-content human approval.); Candidate-code admission and KIRA handoff production: known — automatic [wired] (Reference import gates, pilot source gate and inherited summarizer invocation do not ask a human to approve each retained change.); Requested report authorship: known — automatic [afforded] (Procedure requests model-produced guidance; no report instance is supplied.); Consequential operator decisions retained in domain specification: known — manual [afforded] (This is content-level operator choice, not merely launch authorization.); Direct set_state replacement and unspecified initial/generated/provider admission: not-determinable (Generic caller identity does not establish manual versus automatic control of supplied checkpoint content; provider admission internals are opaque.)	partial coverage (Both requested retrieval and automatic context injection exist. Pilot ordinary tool-result accumulation is excluded as a new read-back mechanism; durable history scoring is included.); Run-keyed checkpoint load, cache replay and deterministic history/frontier reads: known — pull [wired] (These consumers request retained material; fulfilling those reads is not another unsolicited push.); Proposer/analysis history reads and onboarding document load: known — pull [afforded] (Instructions afford later requested file reads; completion and reliance remain unobserved.); Demonstration selection and delivery to solver: known — push [wired] (The solver asks for a task prediction, not for those particular retained examples; candidate code selects them automatically.); KIRA continuation handoff: partial — push [wired] (Visible proactive/overflow retry supplies handoff automatically; inherited selection/delivery alternatives remain opaque.); Generated-candidate alternative read-back and provider-side reuse: not-determinable (The candidate interface and cache decoration do not enumerate actual alternate delivery operations.)	partial coverage (Coarse selectors have positive witnesses, but opaque push alternatives prevent completeness. Cache digests and requested paths identify pull requests, not identifier-based push.); Few-shot automatic demonstration injection: known — coarse [wired] (Trigger: prediction. Selector input: input-derived random seed, list availability, max_examples and character budget. Selected part: formatted examples.); KIRA fallback injection: known — coarse [wired] (Trigger: failed summary after context pressure. Selector input: original instruction and current screen. Selected part: instruction plus last 1000 screen characters alongside remaining chat.); Inherited KIRA summary selection and generated/provider pushes: not-determinable (The inherited selector inputs and generated alternative selectors are unknown; model-generated display text cannot establish inferred-judgment or embedding selection.); Checkpoint/cache/history/document request fulfillment: inapplicable (These named operations are pull reads, so the push-selector axis does not apply to them. The unresolved push alternatives remain included above.)	partial coverage (Positive automatic continuation and feedback-to-artifact routes coexist with unresolved training/generated/provider alternatives. Trace learning here does not assert learning benefit, accepted explanations or successful self-improvement.); KIRA summary and fallback continuation: partial — yes [wired] (Task traces automatically yield retained continuation text delivered to later inference; exact inherited input selection and rationale retention remain unresolved.); Reference trajectory/history-to-candidate and requested report route: known — yes [afforded] (The shipped procedure requests experience-fed durable code/guidance; it does not prove compliance.); Pilot evaluation-feedback-to-candidate edit route: known — yes [wired] (Feedback reaches a model able to edit persistent source subsequently evaluated by children. No actual evidence-responsive edit is observed.); Supervised batch trace qualification, generic checkpoints and provider transformations: not-determinable (Few-shot stores source input/target pairs and drops prediction/reason feedback; offline batches are static labels. These records do not justify a universal trace-fed classification for all training modes or generated candidates, nor a bounded system-wide no.)	partial coverage (Authority is classified at each actual consumer: experience can feed updating while derived content supplies knowledge or routing. Static proposer skills and verifier code are excluded as use-derived objects, not misclassified as memory controls.); Selected demonstrations at classification solver and replayed completion text at requesting candidate: known — knowledge [wired] (These content paths supply answer context or a returned answer, not a human-approved rule.); Reference histories, rationale, traces and requested reports at proposer/analysis reader: known — knowledge [afforded], learning [afforded] (Model file reads and interpretation are afforded by instructions, not automatic full-content injection.); Summary iteration counters at resume and frontier identities/metrics at final selection: known — routing [wired], ranking [wired] (Different retained fields select continuation versus relative candidate priority.); Supervised batch experience at candidate update: partial — learning [wired] (The update is wired; generated candidates may transform feedback differently and no improvement follows.); Retained candidate code at task-harness execution: partial — routing [wired] (Inspected imports grant executable selection rules. Arbitrary generated replacements prevent exhaustive enumeration of their effects.); KIRA handoff at later inference and pilot projected history at editing opportunity: known — knowledge [wired], learning [wired] (Handoff text and readable feedback are advisory; pilot ordinary message accumulation is not itself a derived memory write.); Pilot accepted flags and rewards at historical scorer: known — validation [wired], enforcement [wired], ranking [wired] (History fields control eligibility and maximum historical reward, not final-source restoration.); Domain specification at later implementation agent: known — instruction [afforded] (The afforded handoff conveys operator decisions and explicitly bounded implementation direction.); Opaque generated-state and inherited/provider consumers: not-determinable (Their internal consumer rules and effects cannot be enumerated from transport or display fields.)	AAS-2026-10-14-meta-harness-5e480e270fee-01
Meta-n	2	Included: accumulated evaluated traces; retained preprocessing, Python/bash helpers and rationale; frozen winner maps; checkpoint/admission/access metadata; within-task continuation summaries; later proposal, solving, resume, test and diagnostic consumers. Included but unresolved: SDK conversation state, copied transcript payloads, adapter/evaluator caches and remaining diagnostic transformations, including optional analytical embeddings. Excluded: shipped static prompts/instructions, task datasets as static inputs, ordinary unsummarized invocation state, provider weights without an established in-scope update, and upstream SDK/provider implementations outside the frozen source.	partial coverage (Files and in-memory objects are supported for the named archive and summary inventory. SDK/cache/diagnostic storage remains uninspected; no repo or vector substrate is inferred from repository ownership or an embedding mention.); Evaluated traces and persisted trace/source copies: known — in-memory [wired], files [wired] (Both live and disk substrates are identified for this object.); Retained preprocessing, Python/bash helpers, rationale and frozen maps: known — in-memory [wired], files [wired] (Source objects and their disk copies are covered; file encoding is not a database substrate.); Search checkpoint and reconstructed selection state: known — files [wired], in-memory [wired] (Checkpoint and candidate files are complementary retained objects.); Within-task derived prior-work summary: known — in-memory [wired] (Task-local retention is established; no disk continuation of this summary is established.); Included SDK conversation state, copied log payloads, adapter/evaluator caches and remaining diagnostic transformations: uninspected (These included alternatives lack inspected storage substrate facts. Readable archive traces and diagnostic pointers do not classify their opaque payloads or internal mechanisms.)	partial coverage (Automatic generation, repair, compaction and state capture coexist with manual seed supply. Opaque SDK/cache/diagnostic admissions remain unresolved.); Trace capture, checkpoint/index writes and candidate admission: known — automatic [wired] (Starting or resuming search does not make these writes manual.); Ω generation and within/downstream repair admission: known — automatic [wired] (A non-improving downstream variant can still be archived; no per-change human approval is required.); Summary replacement and frozen winner assembly: known — automatic [wired] (A provider return suffices for summary replacement; neither route requires per-artifact human approval.); CLI seed helper admission: known — manual [wired] (This classifies the content-supply decision, not authorship or function-by-function review; static software checks are additional constraints.); Included SDK conversation state, copied log payloads, adapter/evaluator caches and remaining diagnostic transformations: uninspected (These included alternatives lack inspected write/admission control facts. Readable archive traces and diagnostic pointers do not classify their opaque payloads or internal mechanisms.)	partial coverage (Both requested archive/file delivery and automatic chain/summary/frozen-source supply are supported. Uninspected SDK conversation and diagnostic alternatives prevent a complete direction inventory.); Requested Ω archive view, resume, test and diagnostic reads: known — pull [wired] (Ω sampling is selected requested delivery, not an additional unsolicited push merely because software formats the return.); Reused chain guidance supplied to native/agentic/external solves: known — push [wired] (Automatic context supply is distinct from a consumer asking for an archive view.); Summary supply on later foreground turns: known — push [wired] (Foreground does not request that summary; compaction code selects and supplies it.); Frozen matching source supplied for native task execution: known — push [wired] (The native shortcut is a targeted supply operation; explicit terminal test reads remain pull.); Included SDK conversation state, copied log payloads, adapter/evaluator caches and remaining diagnostic transformations: uninspected (These included alternatives lack inspected read-back direction facts. Readable archive traces and diagnostic pointers do not classify their opaque payloads or internal mechanisms.)	partial coverage (Coarse chain/summary supply and identifier-matched frozen dispatch are wired. Pull-only local units are inapplicable, but unresolved alternatives prevent an axis-wide inapplicable assessment.); Automatic chain guidance push: known — coarse [wired] (No recorded semantic query/embedding selector determines which retained guidance is supplied. Upstream candidate choice is not itself an identifier-match push.); Automatic summary push: known — coarse [wired] (Age/position and budget govern source selection; the summary model authors compression rather than performing an evidenced relevance selector.); Native frozen-source push: known — identifier [wired] (This is an actual identity match selecting source, not merely an identifier on a requested file.); Requested Ω/resume/test/diagnostic material: inapplicable (These named operations are pull; the signal axis characterizes push selectors, not request identifiers.); Included SDK conversation state, copied log payloads, adapter/evaluator caches and remaining diagnostic transformations: uninspected (These included alternatives lack inspected push-selection signal facts. Readable archive traces and diagnostic pointers do not classify their opaque payloads or internal mechanisms.)	partial coverage (Positive trace-learning witnesses remain local. Manual import does not erase its unresolved upstream provenance; raw log retention, absent training calls and missing execution evidence do not establish a system-wide no.); Trace-fed Ω source/rationale generation and repairs: known — yes [wired] (This is wired artifact learning, not demonstrated benefit or model-weight training.); Automatic within-task continuation summary: known — yes [wired] (Task-local retention qualifies; disk persistence and new knowledge are not required.); Automatic frozen winner-map assembly: known — yes [wired] (Deterministic compilation can shape durable behavior without a model proposal or fresh evaluation.); Seed helper upstream trace-fed production: not-determinable (Import itself is a manual supply decision; the record explicitly leaves prior experimentation and trace provenance unresolved. It cannot establish an upstream automatic trace-fed write or a local no finding.); Included SDK conversation state, copied log payloads, adapter/evaluator caches and remaining diagnostic transformations: uninspected (These included alternatives lack inspected automatic trace-fed write and later-consumer chain facts. Readable archive traces and diagnostic pointers do not classify their opaque payloads or internal mechanisms.)	partial coverage (Known consumers establish learning, knowledge, ranking, validation, enforcement, routing and instruction locally. Arbitrary executable contents and uninspected diagnostic/SDK consumers prevent complete authority classification.); Evaluated trace text and prior injections at Ω update consumer: known — learning [wired], knowledge [wired] (Learning force is local to the intervention update; no improved capacity or checked explanatory warrant is inferred.); Stored scores/success and check metadata at archive selection: known — ranking [wired] (Whole-candidate ranking and per-task-best bars are distinct consumers; bars do not remove history.); Retained best-index bars at resume and archive admission: known — validation [wired], enforcement [wired] (The veto concerns promotion, not all execution or truth endorsement; helper stamps remain conditional on trusted JSON.); Retained checkpoint/control state at resumed scheduler: known — routing [wired] (Deterministic scheduling authority is not a content truth criterion.); Frozen task-ID/source map at native solver dispatch: known — routing [wired] (Native dispatch is wired; external parity and fresh score reproduction are not established.); Retained preprocessing rules and helper descriptions at solver/SDK guidance consumer: known — instruction [wired] (Standing chain guidance is model-interpreted; delivery does not establish compliance.); Prior-work summary at subsequent foreground model: known — knowledge [wired] (Unchecked context authority is not validation or a guaranteed factual account.); Executable helpers and trace scripts at execution consumer: not-determinable (Execution determines effects, but source availability or execution alone does not identify a controlled authority value for every arbitrary generated program. Permit/refuse, ranking or routing semantics of unsupplied helper bodies are unresolved.); Retained summaries/source at inspected diagnostic loaders: known — knowledge [wired] (Advisory diagnostic consumption does not establish feedback into search.); Included SDK conversation state, copied log payloads, adapter/evaluator caches and remaining diagnostic transformations: uninspected (These included alternatives lack inspected behavioral authority facts. Readable archive traces and diagnostic pointers do not classify their opaque payloads or internal mechanisms.)	AAS-2026-10-14-meta-n-18c23b6f8bb4-01
MiroShark	2	Retained or changed-through-use round/action narratives, cross-platform tails, belief/checkpoint/trajectory state, document/activity graph parts and access vectors, communities, generated persona/configuration reuse, director assumptions, cached briefings/research, report files/traces and appraisals, plus their admission, maintenance and later-consumer routes. Includes unresolved market/media, dedicated-driver, auxiliary consumer and opaque provider/host alternatives where they can retain or deliver this memory. Excludes shipped static prompts, inference weights as such, ordinary current-call observation filtering without retained results, capability registries/API permissions as such, and external service internals not captured by the boundary.	partial coverage (Known local stores coexist with uninspected provider/host and auxiliary parts. Model weights used for inference are not accumulated memory merely because they are loaded.); Resident round narratives, cross-platform action tails, belief objects and pending tool-result buffers: known — in-memory [wired] (The cited records establish the named inspected part, not successful operation.); Belief checkpoints/trajectories, profiles/configuration, director files and retained reports/appraisals: known — files [wired] (The cited records establish the named inspected part, not successful operation.); Episode/entity/relation/community and optional report-trace graph parts: known — graph [wired], vector [wired] (The cited records establish the named inspected part, not successful operation.); Platform action/recommendation stores read by belief and interview paths: known — sqlite [wired] (The cited records establish the named inspected part, not successful operation.); Ask and research caches: known — in-memory [wired] (The cited records establish the named inspected part, not successful operation.); Opaque provider/host state and incompletely inventoried market/media and auxiliary retained objects: uninspected (Included market/media transforms, dedicated-driver reset/resume alternatives, external host context management, and auxiliary browser/export/cache consumers are not fully characterized in accepted records. These parts remain included uncertainties; inspected routes do not establish an exhaustive inventory.)	partial coverage (Known automatic transformations coexist with unresolved initialization/replacement control; physical file I/O is not the admission agency.); Software/model admission of round, belief, graph, community, persona, report, cache and diagnostic outputs: known — automatic [wired] (The cited records establish the named inspected part, not successful operation.); Caller document/ontology/config initialization, director/counterfactual inputs and appraisal replacement: not-determinable (Generic caller or authenticated administrator identity does not establish an explicit human content decision. No manual finding is licensed for these admissions.); Dedicated-script and opaque host/remote admission mechanisms: uninspected (Included market/media transforms, dedicated-driver reset/resume alternatives, external host context management, and auxiliary browser/export/cache consumers are not fully characterized in accepted records. These parts remain included uncertainties; inspected routes do not establish an exhaustive inventory.)	partial coverage (Both requested retrieval and automatic prompt delivery exist. Known directions do not close auxiliary or opaque host alternatives.); Round scheduling to history, stance and cross-platform context: known — push [wired] (The cited records establish the named inspected part, not successful operation.); Pending external tool-result delivery: known — pull [wired] (The recorded request-to-result chain establishes pull within this delivery unit; no separate unsolicited result-selection operation is established.); Graph, overview, trace, report, cache and appraisal requests: known — pull [wired] (The cited records establish the named inspected part, not successful operation.); Profile/config/checkpoint reload within startup chain: known — pull [wired] (The cited records establish the named inspected part, not successful operation.); Provider history and incompletely traced auxiliary consumers: uninspected (Included market/media transforms, dedicated-driver reset/resume alternatives, external host context management, and auxiliary browser/export/cache consumers are not fully characterized in accepted records. These parts remain included uncertainties; inspected routes do not establish an exhaustive inventory.)	partial coverage (Only actual push selectors are classified here. Embedding/lexical graph ranking belongs to requested pull retrieval and does not establish a push signal.); Common round-history push: known — coarse [wired] (The cited records establish the named inspected part, not successful operation.); Per-agent stance and other-platform digest push: known — identifier [wired], coarse [wired] (The cited records establish the named inspected part, not successful operation.); Pending RPC requested-result delivery: inapplicable (This bounded delivery chain fulfills the agent's tool request and is classified as pull. Agent-ID recipient matching does not establish an independent push selector, so a push signal is inapplicable here. Other unresolved host/remote routes remain separate.); Market/media and opaque provider/host context selectors: uninspected (Included market/media transforms, dedicated-driver reset/resume alternatives, external host context management, and auxiliary browser/export/cache consumers are not fully characterized in accepted records. These parts remain included uncertainties; inspected routes do not establish an exhaustive inventory.)	partial coverage (Positive qualifying routes do not establish universal coverage, observed improvement or parameter training. No system-wide no is inferred from missing evidence.); Round compaction into later history: known — yes [wired] (The cited records establish the named inspected part, not successful operation.); Checkpointed belief feedback: known — yes [wired] (The cited records establish the named inspected part, not successful operation.); Cross-platform bounded digest: known — yes [wired] (The cited records establish the named inspected part, not successful operation.); Optional simulated-activity graph updates: known — yes [wired] (The cited records establish the named inspected part, not successful operation.); Raw report trace retention and external tool-result replay: not-determinable (Inspection/replay is wired but accepted records do not settle every result-specific durable behavior-shaping transformation; raw retention alone does not license a qualifying learning write.); Market/media, dedicated-driver and opaque host alternatives: uninspected (Included market/media transforms, dedicated-driver reset/resume alternatives, external host context management, and auxiliary browser/export/cache consumers are not fully characterized in accepted records. These parts remain included uncertainties; inspected routes do not establish an exhaustive inventory.)	partial coverage (Authority is local to each actual consumer; learning at updates and instruction/knowledge downstream are separate paths. No universal validation or enforcement follows from a schema, static prompt or API guard.); Round/action experience at compaction and belief update consumers: known — learning [wired] (The cited records establish the named inspected part, not successful operation.); Derived history/digest/stance at SocialAgent: known — knowledge [wired], instruction [wired] (The cited records establish the named inspected part, not successful operation.); Persona content and director interventions at agent prompts: known — instruction [wired] (The cited records establish the named inspected part, not successful operation.); Graph text, embeddings and temporal/category controls at search: known — ranking [wired], enforcement [wired] (The cited records establish the named inspected part, not successful operation.); Retrieved graph/community/research and saved reports/traces at requested consumers: known — knowledge [wired] (The cited records establish the named inspected part, not successful operation.); Config controls and matched checkpoint fields at driver/tracker: known — routing [wired] (The cited records establish the named inspected part, not successful operation.); Retained quality and public outcome labels at gallery and signal consumers: known — enforcement [wired], routing [wired], knowledge [wired] (The cited records establish the named inspected part, not successful operation.); Complete retained criteria use, market/media effects and remote payload consumers: uninspected (Included market/media transforms, dedicated-driver reset/resume alternatives, external host context management, and auxiliary browser/export/cache consumers are not fully characterized in accepted records. These parts remain included uncertainties; inspected routes do not establish an exhaustive inventory.)	AAS-2026-10-15-miroshark-b3df0e053a29-01
ModularRSI	2	Includes accumulated raw trials and derived findings/proposals, proposal/disposition/lease state, manifests and journals, evolved executable generations and descriptions, archive/cached bundle access structures, progress/history recovery, and retained same-task summary/scrollback read-back. Includes unresolved historical initialization, additional memory mechanisms inside uninspected released strategy alternatives, unestablished manifest/cross-module receiving paths, and any accumulated provider-side state behind inspected endpoints. Excludes static shipped skill/instruction catalogs, unfrozen external task contents and model weights considered solely as static inference machinery; neither static storage nor API use alone establishes accumulated memory.	partial coverage (Files and process-memory parts are established for the named inventory. Uninspected variants and provider state prevent complete substrate coverage.); Persisted trials, findings/proposals/control ledgers, manifests, journals, archives, generation source, cache and progress/history files: known — files [wired] (These named retained objects have file-backed storage. JSON/Python are encodings, not distinct substrates; repository distribution does not establish a Git-backed memory write mechanism.); Active module registry, folded proposal objects and retained same-task chat/summary and scrollback state: known — in-memory [wired] (These are process-local retained/access parts, distinct from their file backing. Same-task replacement summaries count because another invocation consumes them.); Additional retained objects and update/consumer mechanisms inside uninspected released strategy alternatives: uninspected (Baseline and selected alternatives do not enumerate all executable variant bodies. Their additional accumulated-memory mechanisms, if any, are uninspected; no complete inventory or negative follows.); Accumulated provider-side state, if any, behind configured inference endpoints: uninspected (Clients invoke inference, but provider persistence and parameter changes are inaccessible. No provider-side accumulated-memory value or absence is established. Static weights alone are not classified as a use-acquired memory.)	partial coverage (Automatic live writes are wired. Historical initialization control, uninspected strategy alternatives and provider admission remain unresolved; no manual write is inferred from workflow launch.); Shipped live finding/proposal admission and deterministic state, manifest, archive, journal, progress/history maintenance: known — automatic [wired] (The shipped software/model callers are identified, not inferred from a generic API caller. Operator launch is not per-content approval.); Edited-module deployment through gates and generation promotion: known — automatic [wired] (Reviewer judgment and software guards are computational admission; fail-open and repair qualifications remain attached to the route.); Summary/QA replacement and bundle-cache write on inspected paths: known — automatic [wired] (Workflow configuration initiates these paths but does not decide each retained output.); Historical admission of initial released module content and ancestry: not-determinable (Sanitized initial history does not establish whether content admission involved per-item human decisions. Neither automatic nor manual is assigned to this initialization.); Additional retained objects and update/consumer mechanisms inside uninspected released strategy alternatives: uninspected (Baseline and selected alternatives do not enumerate all executable variant bodies. Their additional accumulated-memory mechanisms, if any, are uninspected; no complete inventory or negative follows.); Accumulated provider-side state, if any, behind configured inference endpoints: uninspected (Clients invoke inference, but provider persistence and parameter changes are inaccessible. No provider-side accumulated-memory value or absence is established. Static weights alone are not classified as a use-acquired memory.)	partial coverage (Both request-driven editor pull and driver/context push are wired. Journal wiring has a local absence; unresolved consumer and variant/provider paths prevent complete direction coverage.); Reflection selection and automatic brief delivery after completed support work: known — push [wired] (The driver selects accumulated evidence for fresh investigator/implementer contexts, rather than answering a solver request for memory.); Later generation load and dynamic description/catalog selection: known — push [wired] (Launching the task is not a foreground model request for each retained strategy or description. Software supplies filtered catalog and selected retained strategy parts.); Summary/QA handoff and retained scrollback delivery on inspected baseline/selected observation paths: known — push [wired] (Budget/overflow and observation capture are automatic selectors of retained material for a later model invocation.); Editor archive, trace and file reads of accumulated content: known — pull [wired] (This unit is the request/return operation: model-selected path/range or archive type/lineage requests are fulfilled on demand. Static skill bodies are excluded from this use-acquired inventory.); Bundle-cache read and forced/K-roll replay into a later solver task: known — push [wired] (The driver-selected mode automatically reconstructs and supplies a stored strategy choice; this is separate from the requested editor-file return.); Compact editor-history and evolution-log read-back in searched shipped callers: absent (Bounded absence is caller wiring in src/ and scripts/, not a denial of the formatter capability or arbitrary external use.); Additional manifest and cross-module handoff consumer routes: not-determinable (Retained helper-accessible records/handoff files do not establish the receiving consumer. Additional direction cannot be assigned.); Additional retained objects and update/consumer mechanisms inside uninspected released strategy alternatives: uninspected (Baseline and selected alternatives do not enumerate all executable variant bodies. Their additional accumulated-memory mechanisms, if any, are uninspected; no complete inventory or negative follows.); Accumulated provider-side state, if any, behind configured inference endpoints: uninspected (Clients invoke inference, but provider persistence and parameter changes are inaccessible. No provider-side accumulated-memory value or absence is established. Static weights alone are not classified as a use-acquired memory.)	partial coverage (Coarse, identifier and model-judgment push selectors are supported at distinct operations. Pull-only local returns are inapplicable locally, not a system-wide signal absence.); Reflection proposal portfolio and task-history scheduling push: known — coarse [wired] (Trigger: completed work or resume. Inputs: retained state, attempts, age, task/window breadth, cooldown and deterministic tie-break. Selected part: eligible proposal/evidence and pending investigation work.); Referenced-finding and earlier-passing-trial push into fresh investigation/implementation: known — identifier [wired] (Trigger: selected proposal or task lacking a current pass. Inputs: proposal finding IDs or exact task identity. Selected part: matched finding bodies or the stored passing-trial referent, subject to display/revalidation limits.); Archive and active-generation filtering/load push: known — coarse [wired], identifier [wired] (Trigger: task setup/composition. Inputs: current generation, archive status/type/niche and explicit pin identities. Selected parts: eligible library entries/catalog and named strategy implementations.); Dynamic model selection of retained strategy catalog entries: known — inferred-judgment [wired] (Trigger: multiple eligible choices/helper catalog. Inputs: task text and retained descriptions. Selected parts: offered variant/helper names and their implementations. No reward or embedding retrieval is established.); Cache replay push selected by task-instruction hash: known — identifier [wired] (Trigger: cache-read mode. Input: instruction SHA-256. Selected part: matching persisted bundle payload, then its named implementations. A real identity lookup occurs, not merely an identifier in a display.); Compression and bounded retained-observation push: known — coarse [wired] (Trigger: low context/overflow or observation capture. Inputs: token budget, tail positions and stale/current output policy; selected parts: surviving chat, summary/handoff and bounded scrollback. Model summary production is a transformation, not itself proof of a semantic memory-retrieval selector.); Requested editor-file and archive-return operations: inapplicable (Signal classifies push selection only. These request/return operations are pull; their path/type arguments are not an additional identifier push finding.); Compact journal read-back in searched callers: inapplicable (No wired delivery/selection route exists in the searched shipped callers, so there is no push selector to classify for that bounded route.); Unestablished manifest and cross-module receiving consumers: not-determinable (Neither a consumer direction nor actual push selector is established; known pull elsewhere cannot make these alternatives inapplicable.); Additional retained objects and update/consumer mechanisms inside uninspected released strategy alternatives: uninspected (Baseline and selected alternatives do not enumerate all executable variant bodies. Their additional accumulated-memory mechanisms, if any, are uninspected; no complete inventory or negative follows.); Accumulated provider-side state, if any, behind configured inference endpoints: uninspected (Clients invoke inference, but provider persistence and parameter changes are inaccessible. No provider-side accumulated-memory value or absence is established. Static weights alone are not classified as a use-acquired memory.)	partial coverage (Three qualifying automatic trace-fed routes support yes. The bounded journal-feedback no remains local. Initial history, additional consumer paths, variant and provider uncertainties retain partial coverage; yes and incomplete coverage each independently prevent a system-wide no.); Primary trace diagnosis through retained proposal/brief to admitted executable module and later solver: known — yes [wired] (Automatic trace-fed production yields a durable behavior-shaping library consumed on later tasks. This controlled yes concerns the write/consumer route, not improved capacity, faithful diagnosis or exercised causal self-improvement.); Automatic session-history compression into later-call continuation: known — yes [wired] (The retained summary/QA result survives replacement and reaches another invocation within the same task; cross-task reuse is not required. Fallback differences do not erase the shared QA witness.); Outcome-fed archive eligibility revision: known — yes [wired] (Recorded task/bundle/reward experience automatically changes durable selection status, which later ordinary composition consumes. Descriptive confirmation is not randomized causal evidence.); Initial released-library training and historical admission: not-determinable (The live update route does not reconstruct whether each initial released entry was automatically learned from traces. Sanitized history prevents a complete initial learning classification.); Compact editor-history and evolution-log feedback through searched shipped callers: known — no [wired] (The absence is bounded to the searched src/ and scripts/ journal-feedback path, not all learning or arbitrary external consumers. Outcome retention and an unused formatter do not close a later-consumer chain.); Manifest and cross-module handoff additional receiving paths: not-determinable (Trace-derived recording alone is insufficient. Helper-accessible records and handoff files do not establish receiving consumers or complete absence of qualifying updates.); Additional retained objects and update/consumer mechanisms inside uninspected released strategy alternatives: uninspected (Baseline and selected alternatives do not enumerate all executable variant bodies. Their additional accumulated-memory mechanisms, if any, are uninspected; no complete inventory or negative follows.); Accumulated provider-side state, if any, behind configured inference endpoints: uninspected (Clients invoke inference, but provider persistence and parameter changes are inaccessible. No provider-side accumulated-memory value or absence is established. Static weights alone are not classified as a use-acquired memory.)	partial coverage (Authority is classified at named consumers, independently of admission agency and benefit. Journal/manifest consumer limits, alternative strategies and provider state prevent full coverage.); Retained findings, historical trial evidence and causal hypotheses at investigator/implementer update consumers: known — knowledge [wired], learning [wired] (Split evidence/reference use from the trace-fed durable-update use. Delivery is wired, but compliance, correct causality and benefit are unobserved.); Retained proposed behavioral delta at the assigned implementer: known — instruction [wired] (The brief specifies the intervention to implement, rather than merely furnishing background evidence.); Proposal eligibility, dispositions and lease state at software selection/recovery consumers: known — routing [wired], ranking [wired], enforcement [wired], validation [wired] (These are fixed-rule effects, not the semantic authority of a rationale stored beside them.); Archive statuses, niche/identity fields and bundle replay at subsequent composition: known — routing [wired], enforcement [wired] (Archive veto is scoped to ordinary dynamic selection. Replay/static/forced paths and import effects are not universally revoked.); Retained module descriptions at the dynamic composer: known — knowledge [wired], routing [wired] (Catalog capability prose is advisory input; software validates returned names but does not validate the descriptions as true.); Retained executable generation at the loader and baseline runtime policy consumers: known — routing [wired], enforcement [wired] (Python is consumed as implementation, not classified as natural-language instruction. These supported effects do not enumerate the force of every uninspected variant.); Checkpoint and historical-pass access state at resumed scheduling and diagnosis: known — routing [wired], knowledge [wired] (Identity/alignment and progress fields have software force; associated summary text separately supplies investigation context.); QA continuation at the next solver invocation: known — knowledge [wired], instruction [wired] (At the next solver invocation the retained handoff supplies context/guidance, not an evidenced parameter or durable-artifact update. Its automatic production separately qualifies for trace learning; that does not give the output learning authority at downstream inference. No activation or improved performance is observed.); Outcome journals, unused compact-history display and manifest consumer details beyond supplied promotion wiring: not-determinable (Journal writes and manifest helper APIs do not establish a consequential later content consumer. Missing consumer wiring prevents assigning additional authority to those payloads.); Additional retained objects and update/consumer mechanisms inside uninspected released strategy alternatives: uninspected (Baseline and selected alternatives do not enumerate all executable variant bodies. Their additional accumulated-memory mechanisms, if any, are uninspected; no complete inventory or negative follows.); Accumulated provider-side state, if any, behind configured inference endpoints: uninspected (Clients invoke inference, but provider persistence and parameter changes are inaccessible. No provider-side accumulated-memory value or absence is established. Static weights alone are not classified as a use-acquired memory.)	AAS-2026-10-18-modularrsi-4218ceff5530-01
nao	2	Retained objects accumulated or changed through use: user-memory content and access controls; conversation/history, compaction summaries and markers; saved files/stories and query evidence caches; usage-derived recommendations, dismissal/status state and proposed edits; changed/imported project context and its later guidance routes; saved evaluation reports with unresolved later consumption. Includes unresolved fork, dynamic-skill, channel, app-state, file-producer/lifecycle, context-provider, feedback and opaque provider-memory alternatives. Excludes unchanged shipped static instructions, initial pretrained weights, ordinary transient grants/credentials and external service internals as evidence. External accumulated provider state remains an unresolved included alternative, not an inferred weight-update route.	partial coverage (Positive retained-object substrates are supported locally. Opaque provider state, file-provider internals, MCP cache and the listed alternative producers prevent a complete substrate inventory; storage is not inferred from encoding.); User-memory rows, including content and access/lineage metadata: known — sqlite [wired], rdbms [wired] (Both supported DB implementations store the same scoped row parts; this is a configured alternative, not evidence that both were deployed.); Conversation transcript and persisted streaming compaction marker: known — rdbms [wired] (Settled messages and successful streamed markers persist in the application DB; DB-engine details for these parts are not separately classified.); Compaction summary retained between model-step invocations: known — in-memory [wired] (Working-array replacement establishes process retention, including generate without the streaming persistence channel.); Retained recommendations and reconciliation metadata: known — rdbms [wired] (The DB transaction stores diagnosis, edits, status and fingerprint parts.); Permanent saved-file payloads: partial — files [afforded], service-object [afforded] (Configured local and S3 alternatives are named, but each provider implementation and its durability are uninspected.); Changed project context and proposed patch branches: known — files [wired], repo [wired] (Live filesystem admission and separate disposable Git proposal branches are inspected; merge and deployment are not inferred.); Cached query-result evidence and retained tool-result parts: partial — in-memory [wired], rdbms [wired] (Process and transcript paths are classified; the direct-MCP cache substrate is not specified sufficiently in the accepted records.); Saved CLI evaluation reports: known — files [wired] (Workflow inspection establishes timestamped JSON report files; later model consumption is not established.); Forked chat/history and copied checkpoints: uninspected (Fork producer and copied payload were located but not traced; Storage cannot be completed for alternative fork initialization.); Dynamic skill loading and channel-specific retained history/memory: uninspected (Dynamic load_skill and messaging-specific retention/consumption were not traced; standard web and explore witnesses do not cover these alternatives.); Story publication, retained app state and blob lifecycle: uninspected (Latest story replay is inspected, but publication, opaque app-state payload and cleanup chains are uninspected; a displayed story does not classify its app state.); Other saved-file byte producers and provider lifecycle: uninspected (Byte/from-disk APIs exist, but their producers, payload interpretation, admission and cleanup are uninspected; converted display text does not classify every stored binary.); Provider-specific CLI/API project-context admission: uninspected (The dispatcher and Git delivery paths are inspected; provider outputs and API-upload admission are not fully traced. Changed project content is included, unchanged shipped instructions are excluded.); Recommendation feedback fallback and app-DB filtering: uninspected (Feedback fallback, coverage postprocessing and complete diagnostic query filtering are uninspected; the known reconciliation branch cannot complete this alternative.); Opaque external provider retention or adaptation: uninspected (Provider retention/training internals are outside the frozen evidence. No local parameter-update step is established. Any provider memory accumulated through use remains unresolved; static pretrained weights are not classified as accumulated memory.)	partial coverage (Known software admission coexists with explicit upload/edit/dismissal surfaces. Initial rows, remote source governance, file producers and alternative caller policies remain unresolved; agency is not authorship or physical I/O.); Extraction and supersession admission: known — automatic [wired] (Software admits normalized model proposals without per-item human approval.); Explicit memory edit/delete/enablement surface: partial — manual [afforded] (The mutation affords an explicit human decision about supplied content or withdrawal; full UI caller wiring is uninspected.); Initial user-memory content admission: not-determinable (Schema/active-row presence does not resolve initial control.); Compaction replacement and streamed checkpoint admission: known — automatic [wired] (Budget trigger and model return lead to software replacement/persistence without semantic approval.); Transcript and query-cache admission: known — automatic [wired] (Settled model messages and query returns are admitted by callbacks/adapters, not human confirmation of every retained item.); Recommendation transaction and configured automatic PR creation: known — automatic [wired] (An operator can start/configure analysis but software selects each retained action and eligible PR.); Manual recommendation dismissal/status admission: partial — manual [afforded] (Human dismissal is described, but the status-mutation UI/policy is uninspected; do not upgrade that gap from the read-side veto.); User-selected file uploads: known — manual [wired] (Explicit choice of uploaded bytes establishes manual content admission for this branch.); Foreground-model file writes: known — automatic [wired] (Model-selected authored text passes path/size gates, with no per-content approval.); Context refresh/sync byte admission: known — automatic [wired] (Operator source selection and workflow initiation are not approval of each transferred/generated file.); Upstream project source authoring/review: not-determinable (The operator selects a source, but per-content human authorship, approval and repository governance are unresolved.); Saved CLI evaluation report admission: known — automatic [wired] (Workflow report saving is automatic after run-result collection.); Forked chat/history and copied checkpoints: uninspected (Fork producer and copied payload were located but not traced; Admission control cannot be completed for alternative fork initialization.); Dynamic skill loading and channel-specific retained history/memory: uninspected (Dynamic load_skill and messaging-specific retention/consumption were not traced; standard web and explore witnesses do not cover these alternatives.); Story publication, retained app state and blob lifecycle: uninspected (Latest story replay is inspected, but publication, opaque app-state payload and cleanup chains are uninspected; a displayed story does not classify its app state.); Other saved-file byte producers and provider lifecycle: uninspected (Byte/from-disk APIs exist, but their producers, payload interpretation, admission and cleanup are uninspected; converted display text does not classify every stored binary.); Provider-specific CLI/API project-context admission: uninspected (The dispatcher and Git delivery paths are inspected; provider outputs and API-upload admission are not fully traced. Changed project content is included, unchanged shipped instructions are excluded.); Recommendation feedback fallback and app-DB filtering: uninspected (Feedback fallback, coverage postprocessing and complete diagnostic query filtering are uninspected; the known reconciliation branch cannot complete this alternative.); Opaque external provider retention or adaptation: uninspected (Provider retention/training internals are outside the frozen evidence. No local parameter-update step is established. Any provider memory accumulated through use remains unresolved; static pretrained weights are not classified as accumulated memory.)	partial coverage (Both requested access and unsolicited assembly are supported, including different operations in one chain. Unknown report consumption and alternate forks/channels/providers prevent a complete direction inventory.); User-memory default prompt injection and extraction-input reload: known — push [wired] (Automatic supply is not a foreground request for these retained rows; full overrides can bypass the default route.); Conversation continuation request: known — pull [wired] (The consumer requests a specific chat; responding with stored history is pull, not another independent push.); Checkpoint and latest-story selection inside continuation: known — push [wired] (Within requested reconstruction, automatic narrowing/replacement supplies retained checkpoint/current story parts without a separate request for them.); Explicit saved-file and cached-query reads: known — pull [wired] (The model/host names a retained path/query and requests its contents; delivery fulfills that request.); Prior recommendation summaries and reconciliation state: known — push [wired] (Initial diagnostic assembly supplies existing summaries and retained fingerprints/status to update consumers.); Admitted evolving project guidance at future prompt/read consumers: known — push [wired], pull [wired] (Prompt supply and explicit context reads are distinct operations; unchanged shipped static instructions are outside memory scope.); Saved evaluation report future consumption: not-determinable (Saving is known, but no later model/behavior-shaping read route is established; this is not a bounded absence.); Forked chat/history and copied checkpoints: uninspected (Fork producer and copied payload were located but not traced; Read-back cannot be completed for alternative fork initialization.); Dynamic skill loading and channel-specific retained history/memory: uninspected (Dynamic load_skill and messaging-specific retention/consumption were not traced; standard web and explore witnesses do not cover these alternatives.); Story publication, retained app state and blob lifecycle: uninspected (Latest story replay is inspected, but publication, opaque app-state payload and cleanup chains are uninspected; a displayed story does not classify its app state.); Other saved-file byte producers and provider lifecycle: uninspected (Byte/from-disk APIs exist, but their producers, payload interpretation, admission and cleanup are uninspected; converted display text does not classify every stored binary.); Provider-specific CLI/API project-context admission: uninspected (The dispatcher and Git delivery paths are inspected; provider outputs and API-upload admission are not fully traced. Changed project content is included, unchanged shipped instructions are excluded.); Recommendation feedback fallback and app-DB filtering: uninspected (Feedback fallback, coverage postprocessing and complete diagnostic query filtering are uninspected; the known reconciliation branch cannot complete this alternative.); Opaque external provider retention or adaptation: uninspected (Provider retention/training internals are outside the frozen evidence. No local parameter-update step is established. Any provider memory accumulated through use remains unresolved; static pretrained weights are not classified as accumulated memory.)	partial coverage (Identifier and coarse selection are established at named automatic operations. Requested path/chat labels do not alone establish identifier push. No inferred embedding, lexical or judgment selector is asserted for uninspected branches.); Default user-memory injection: identity matching: known — identifier [wired] (Trigger is default prompt construction. userId match and current-chat exclusion select delivered active rows; identifiers are actual predicates, not mere labels.); Default user-memory injection: availability/order/budget filtering: known — coarse [wired] (The same trigger selects by enabled/active state, category priority, recency and 1000-token content budget; no semantic relevance search is established.); Existing-memory supply to extractor: known — identifier [wired] (Extraction preparation automatically selects active user rows including IDs for future replacement decisions.); Continuation checkpoint selector: known — coarse [wired] (On reconstruction, latest nonempty data-compaction marker selects summary plus recent tail; requesting chat identity is separately pull.); Current story replay selector: known — identifier [wired] (During preparation, referenced story identity selects current DB body; older duplicates are stubbed.); Prior recommendation summary assembly: known — coarse [wired] (Diagnostic start supplies reconcilable status summaries under the configured period/budget; no task-inferred semantic ranking is established by this initial display.); Recommendation state selection for reconciliation: known — identifier [wired] (Recorded fingerprint matches select existing recommendation state for update/resolve decisions, and dismissed matches select veto state.); Changed project guidance supply: partial — identifier [wired], coarse [wired] (Prompt construction selects project guidance and configured override/provider files. Full provider-specific selectors remain uninspected.); Explicit saved-file and query paging pull routes: inapplicable (These units fulfill named requests, so push-selection signal is inapplicable locally; this is not a system-wide pull-only assertion.); Saved evaluation reports: not-determinable (Later consumption is not traced, preventing determination of either a push route or its signal.); Forked chat/history and copied checkpoints: uninspected (Fork producer and copied payload were located but not traced; Push selection cannot be completed for alternative fork initialization.); Dynamic skill loading and channel-specific retained history/memory: uninspected (Dynamic load_skill and messaging-specific retention/consumption were not traced; standard web and explore witnesses do not cover these alternatives.); Story publication, retained app state and blob lifecycle: uninspected (Latest story replay is inspected, but publication, opaque app-state payload and cleanup chains are uninspected; a displayed story does not classify its app state.); Other saved-file byte producers and provider lifecycle: uninspected (Byte/from-disk APIs exist, but their producers, payload interpretation, admission and cleanup are uninspected; converted display text does not classify every stored binary.); Provider-specific CLI/API project-context admission: uninspected (The dispatcher and Git delivery paths are inspected; provider outputs and API-upload admission are not fully traced. Changed project content is included, unchanged shipped instructions are excluded.); Recommendation feedback fallback and app-DB filtering: uninspected (Feedback fallback, coverage postprocessing and complete diagnostic query filtering are uninspected; the known reconciliation branch cannot complete this alternative.); Opaque external provider retention or adaptation: uninspected (Provider retention/training internals are outside the frozen evidence. No local parameter-update step is established. Any provider memory accumulated through use remains unresolved; static pretrained weights are not classified as accumulated memory.)	partial coverage (Three transformation families establish qualifying automatic trace-fed memory routes, with streaming and nonstreaming summaries kept separate. Unresolved producers/consumers remain; no unsupported no finding is used to close their coverage. This axis does not assert improved capacity, parametric training or exercised self-improvement.); Automatic user-memory extraction to future guidance: known — yes [wired] (Recent recorded turns feed software-admitted durable facts/directives consumed by later prompts/extraction; improvement is not required or observed.); Streaming compaction to persisted continuation summary: known — yes [wired] (Raw/prepared conversation traces produce a summary persisted in a valid streamed marker and supplied on continuation.); Nonstreaming compaction to subsequent in-process model invocation: known — yes [wired] (generate lacks streamed-marker persistence, but the summary is retained in the working array and consumed by later steps. Cross-session persistence is not asserted.); Usage/feedback diagnosis to future diagnostic runs: partial — yes [wired] (Automatically admitted recommendations are resupplied as compact summaries/state; PR merge/live analytics effect is not necessary for this diagnostic-memory route. Fallback inputs remain unresolved.); Raw transcript, query caches, saved files and delivered project revisions beyond qualifying routes: not-determinable (Storage/authoring/import alone does not establish an automatic trace-to-derived-guidance route. Complete producer coverage is missing, so no bounded no finding is made.); Evaluation reports and provider-side adaptation: uninspected (Reports have no established later learning consumer; provider adaptation internals are uninspected. Neither supports a system-wide negative.); Forked chat/history and copied checkpoints: uninspected (Fork producer and copied payload were located but not traced; Trace-fed learning cannot be completed for alternative fork initialization.); Dynamic skill loading and channel-specific retained history/memory: uninspected (Dynamic load_skill and messaging-specific retention/consumption were not traced; standard web and explore witnesses do not cover these alternatives.); Story publication, retained app state and blob lifecycle: uninspected (Latest story replay is inspected, but publication, opaque app-state payload and cleanup chains are uninspected; a displayed story does not classify its app state.); Other saved-file byte producers and provider lifecycle: uninspected (Byte/from-disk APIs exist, but their producers, payload interpretation, admission and cleanup are uninspected; converted display text does not classify every stored binary.); Provider-specific CLI/API project-context admission: uninspected (The dispatcher and Git delivery paths are inspected; provider outputs and API-upload admission are not fully traced. Changed project content is included, unchanged shipped instructions are excluded.); Recommendation feedback fallback and app-DB filtering: uninspected (Feedback fallback, coverage postprocessing and complete diagnostic query filtering are uninspected; the known reconciliation branch cannot complete this alternative.); Opaque external provider retention or adaptation: uninspected (Provider retention/training internals are outside the frozen evidence. No local parameter-update step is established. Any provider memory accumulated through use remains unresolved; static pretrained weights are not classified as accumulated memory.)	partial coverage (The inventory names actual update, downstream guidance and code-control consumers separately. Knowledge delivery is not compliance or epistemic acceptance; untraced payloads/consumers prevent an exhaustive force classification.); Personal facts at default analytics prompt: known — knowledge [wired] (The model receives category/content as previously established but defeasible context, not evidence or IDs.); Global rules at default analytics prompt: known — instruction [wired] (Directions have advisory model force; they do not bypass executable grants.); Active memories and conversation text at extraction update consumer: known — learning [wired] (Existing rows and recent traces feed automatic durable guidance revision; no improved capacity is inferred.); Memory category/timestamp priority at default prompt selector: known — ranking [wired] (Stored category and creation order influence relative delivery priority before token-fit selection.); Memory IDs and supersession metadata at write/read gates: known — validation [wired], enforcement [wired], routing [wired] (Fixed membership and active-set consumers are distinct from prose guidance.); Recorded history at compaction update consumer: known — learning [wired] (Traces feed a retained continuation artifact, with image and budget omissions.); Compaction text and marker at continuation consumers: known — knowledge [wired], routing [wired] (Interpreted context and deterministic window control have different forces.); Recommendation summaries at future diagnostic model: known — knowledge [wired] (Compact titles/status/file/fingerprint summaries, not full stored reasons, inform the next diagnostic pass.); Retained usage/feedback and recommendations at diagnostic update consumer: known — learning [wired] (The model and reconciler update durable diagnostic state consumed by later runs, without establishing truth or live analytics improvement.); Recommendation fingerprint/status/count state at reconciliation and PR selection: known — validation [wired], enforcement [wired], ranking [wired], routing [wired] (Dismissal state is a veto, occurrence/status controls actions, and impact affects priority/admission.); Recommendation guidance/edits at PR or manual-fix surfaces: known — instruction [afforded], knowledge [wired] (PR creation consumes proposed edits. Human execution of manual fix instructions is only afforded; live admission is separate.); Replayed prose, query evidence and saved readable files at later analytics consumers: partial — knowledge [wired] (These establish reference/context supply; binary conversion and complete story/app-state effects remain unresolved.); Changed project rules/reference files at subsequent analytics consumers: partial — instruction [wired], knowledge [wired] (Separate live admission is required. The ordinary write tool restriction is not authority of these retained contents.); Saved CLI report at later consumer: not-determinable (Report saving is inspected, but later report-driven behavior is not; test comparison is not authority of a previously retained report.); Forked chat/history and copied checkpoints: uninspected (Fork producer and copied payload were located but not traced; Authority cannot be completed for alternative fork initialization.); Dynamic skill loading and channel-specific retained history/memory: uninspected (Dynamic load_skill and messaging-specific retention/consumption were not traced; standard web and explore witnesses do not cover these alternatives.); Story publication, retained app state and blob lifecycle: uninspected (Latest story replay is inspected, but publication, opaque app-state payload and cleanup chains are uninspected; a displayed story does not classify its app state.); Other saved-file byte producers and provider lifecycle: uninspected (Byte/from-disk APIs exist, but their producers, payload interpretation, admission and cleanup are uninspected; converted display text does not classify every stored binary.); Provider-specific CLI/API project-context admission: uninspected (The dispatcher and Git delivery paths are inspected; provider outputs and API-upload admission are not fully traced. Changed project content is included, unchanged shipped instructions are excluded.); Recommendation feedback fallback and app-DB filtering: uninspected (Feedback fallback, coverage postprocessing and complete diagnostic query filtering are uninspected; the known reconciliation branch cannot complete this alternative.); Opaque external provider retention or adaptation: uninspected (Provider retention/training internals are outside the frozen evidence. No local parameter-update step is established. Any provider memory accumulated through use remains unresolved; static pretrained weights are not classified as accumulated memory.)	AAS-2026-10-14-nao-f9d36a4b7276-01
Napkin	2	Retained Markdown/daily/folder notes and NAPKIN.md; persisted templates and retrieval/layout configuration; search/overview caches including opaque index payload; base definitions, canvas and bookmarks; benchmark source/summary notes and retained reports; packaged session distill/tend update routes and separately claimed external background update/incremental state. Includes unresolved initial provenance, external context delivery and cursor alternatives. Excludes static shipped skill/method text as a memory object, external provider weights, installed-package changes, process-only vault/template registries, transient SQL/link/graph/query projections and ordinary current-run state. Temporary benchmark notes are included within their later-question-consumer horizon.	partial coverage (The recorded local inventory is file-backed, including the opaque dependency index. External incremental cursor persistence is unresolved; provider weights are excluded as external runtime components, not accumulated vault memory.); Markdown notes, daily notes, folder descriptions, NAPKIN.md and persisted note templates: known — files [wired] (All these retained objects have recorded filesystem locations.); Disk configuration, search-cache payload/metadata and overview-cache result: known — files [wired] (JSON encoding, including an opaque serialized index, does not change the known filesystem substrate.); Base definitions, canvas content and bookmarks: known — files [wired] (The base query database is transient and excluded from the retained substrate inventory.); Benchmark source notes, supplied LoCoMo summaries and retained evaluation reports: known — files [wired] (Temporary files qualify during their later-consumer horizon; current-process result arrays alone are excluded.); External background-distillation incremental cursor or checkpoint: uninspected (The design names new messages since the prior run, but the unavailable extension leaves retained state, format and persistence location uninspected. No substrate is inferred.)	partial coverage (Automatic admission is supported locally; generic callers and initial-content admissions do not identify human control. Starting a workflow or benchmark is not per-content approval.); Search/overview cache admission and benchmark note/result ingestion: known — automatic [wired] (Algorithms and drivers admit these contents without a per-content operator decision.); Packaged distill and ordinary tend semantic admission/repair: known — automatic [claimed] (The instructed external model makes keep/skip, merge, repair and retirement decisions.); Background model/extension note admission: known — automatic [claimed] (The stated admission controller is a separate model/extension, not the human launching pi.); Tend proposed new-template adoption: known — manual [claimed] (This human schema decision is separate from ordinary tend cleanup and unrestricted template APIs.); Generic CRUD, auxiliary edits, template/scaffold/configuration, base/canvas/bookmark admissions and pre-existing content: not-determinable (Caller identity and initial admission control are unspecified. Physical filesystem I/O does not establish automatic admission or human approval.)	partial coverage (Requested core and shipped benchmark routes are established. The always-loaded host-context description supplies a claimed push path, while actual external injection/cursor delivery and graph rendering remain unresolved.); Later overview/search/full-note requests including cached inputs and context bundled with overview: known — pull [wired] (A broad context return fulfilling overview is pull, not another push.); Daily/template/field, base, canvas and bookmark reads or update-consumer requests: known — pull [wired] (A template selected for create/insert is requested retained input at an update consumer.); Benchmark question-agent requested retrieval and evaluator report aggregation: known — pull [wired] (Earlier writes have later question/aggregation consumers; actual exercise remains unobserved.); NAPKIN.md always-loaded host context: partial — push [claimed] (The skill describes automatic host loading, but the host implementation, trigger and actual delivery are unavailable.); External background capture context/templates, incremental cursor delivery and benchmark extension injection: uninspected (Design names inputs and external extensions, but request/selection/delivery operations are not inspected. No additional push wiring is inferred.); Graph human UI continuation: uninspected (The recorded graph snapshot is current-run data, not a distinct retained store; unread renderer continuation prevents classifying later display delivery.)	partial coverage (Only the claimed broad always-loaded context supports a positive push signal. Requested paths, base IDs, bookmark references and query ranking do not establish identifier or lexical push. Unresolved external direction units prevent system-wide inapplicability.); Claimed always-loaded NAPKIN.md context selection: partial — coarse [claimed] (The claim is broad availability-based loading. Exact host trigger, filtering and injection are not recorded.); Requested core/template/structured/benchmark retrieval chains: inapplicable (These recorded chains fulfill explicit requests. Their query/path/view selectors therefore do not classify push signals.); External extension/background cursor selection and graph continuation: uninspected (Missing external direction and selection details prevent concluding whether there is a push signal or assigning targeted identity, lexical, embedding or judgment values.)	partial coverage (Two specified automatic session-to-vault routes support local yes findings at claimed strength. Raw log ingestion, supplied summaries and evaluator reports do not establish an additional qualifying update; unresolved derivation and external state prevent a system-wide no.); Packaged distill session-to-note-to-future-reader chain: known — yes [claimed] (Qualifies by the claimed automatic trace-fed durable content write and later knowledge consumer, not by observed improvement.); Separate-model background conversation-to-vault chain: known — yes [claimed] (The documented write and intended later vault consumer establish a claimed qualifying route, not inspected implementation.); Benchmark raw-dialogue ingestion, supplied summaries and evaluation outputs: not-determinable (Retaining raw logs alone is not a qualifying transformation. Summary production is upstream and uninspected; reports have no inspected score-to-behavior update. No bounded absence record warrants a local no for all these alternatives.); Other generic content edits and external incremental state/host behavior: uninspected (Generic callers do not identify trace-fed admission or a durable behavior-shaping chain; external cursor/host implementation is unavailable.)	partial coverage (Authority is local to consumed parts and effects. Wired routing/ranking coexist with claimed trace-fed learning and external guidance. Unknown injection and partially inspected base/graph consumers prevent complete authority coverage.); Retained note/context/summary and canvas text supplied as caller reference: known — knowledge [wired] (The effect is availability as evidence/context, not compliance or truth certification.); Retained benchmark notes at question-agent evidence consumer: known — knowledge [wired] (The shipped drivers request later question-agent retrieval; exercised dependence remains unobserved.); Note terms, basenames, links/mtimes and serialized index/access metadata at search ranking: known — ranking [wired] (A fixed scoring algorithm still gives retained content ranking authority.); Disk retrieval/layout settings, overview handles, base query definitions and saved access references at selectors or requesting callers: partial — routing [wired], knowledge [wired] (Base filter/formula internals are only partly inspected. Bookmarks expose destinations, not automatically executed queries.); Stored template section guidance and project conventions at external model guidance consumers: partial — instruction [afforded], instruction [claimed] (Template bytes are wired into notes, but host interpretation and automatic convention delivery remain external.); Session experience at packaged distill update consumer: known — learning [claimed] (The local authority is the claimed durable update path, separate from downstream knowledge authority.); Conversation entries at separate-model background update consumer: known — learning [claimed] (This is claimed authority, not verified scheduler or model operation.); Evaluation report text and metrics at aggregation/report consumers: known — knowledge [wired] (Scores affect reports, not knowledge admission or policy successors.); External automatic context injection, incremental checkpoint consumer and graph renderer continuation: uninspected (External delivery/state and partially inspected rendering do not establish additional instruction, enforcement or routing effects. Static packaged skills and core guards are not accumulated memory criteria.)	AAS-2026-10-12-napkin-2b85bef38252-01
nuggets	2	Shipped retained shelf facts (default and named stores), keys/access metadata, extension-local facts and exposed branch snapshots, precompaction task/path artifacts, reconstructed operative HRR state, promoted Markdown and accepted cron controls/prompts, including their admission, maintenance and later-consumer routes. External Pi session/checkpoint payloads are included as unresolved continuation alternatives. Excludes shipped static instructions, transient current-turn RPC/queue state and recall scores, setup/transport controls, provider parameter training and the absent advertised plugin. External Claude loading is an unresolved consumer of included exported memory, not an established host implementation.	partial coverage (Inspected retained objects use files and process RAM. External host session/checkpoint substrates remain unresolved; JSON and HRR are encodings, not additional substrates.); Shelf fact strings, keys, task excerpt and access metadata: known — files [wired], in-memory [wired] (Fact/configuration JSON and the corresponding process arrays are covered; this is not vector-only persistence.); Extension-local facts and exposed tool-result snapshots: known — in-memory [wired] (Local Map storage is inspected; host persistence of the returned details is assessed separately.); Reconstructed HRR banks and vocabulary vectors: known — in-memory [wired] (Numerical encoding is reconstructed in RAM, not serialized or stored in a vector database.); Promoted Markdown artifact: known — files [wired] (The conditional writer establishes a retained filesystem artifact, not actual external reading.); Accepted cron job controls and prompts: known — files [wired], in-memory [wired] (Persisted accepted jobs and loaded job array are inspected, distinct from intermediate proposals.); External Pi-retained session payloads, compaction summaries and alternative checkpoints: uninspected (Gateway continuation and extension branch-detail access are recorded, but external host serialization, checkpoint production and later selection are uninspected; Storage substrate cannot be completely classified.)	partial coverage (Inspected model/code mechanisms are automatic. The records do not support a manual finding merely from migration launch, generic APIs or possible file editing; initial/direct-caller control and host checkpoint admission remain unresolved.); Extension explicit remember/forget and snapshot admission: known — automatic [wired] (Model choices and extension rules admit changes without per-content human approval.); Input/result-hook and precompaction extraction admission: known — automatic [wired] (Human input authorship or workflow initiation is not approval of extracted content.); Branch replay, hydration and local-map replacement: known — automatic [wired] (Lifecycle code admits existing retained content into the later operative Map.); Hit-counter maintenance: known — automatic [wired] (A supplied nonempty session ID enables code-controlled counter updates, with no per-counter human decision.); Legacy migration and conditional Markdown promotion: known — automatic [wired] (Operator launch of migration is distinct from its per-fact code admission; promotion is threshold/code-admitted.); Schedule-tool request and gateway accepted-job mutation: known — automatic [wired] (The user initiates a task, but the model fields are admitted by code without a per-job approval step.); Direct-library and initial-file content admission: not-determinable (Generic caller identity and existing loaded content do not establish explicit human supply/approval or software-only admission.); External Pi-retained session payloads, compaction summaries and alternative checkpoints: uninspected (Gateway continuation and extension branch-detail access are recorded, but external host serialization, checkpoint production and later selection are uninspected; Write agency cannot be completely classified.)	partial coverage (Both requested pull and unsolicited push are wired. Exported Markdown loading and external host continuation mechanisms are unresolved, so the boundary is not completely pull-only or completely classified.); Explicit model recall/list and operator overlay: known — pull [wired] (These operations fulfill consumer requests; automatic decoding inside requested recall does not add another push.); Extension lifecycle selection and per-turn fact injection: known — push [wired] (Branch reconstruction/start hydration supply the local Map, and before-agent code delivers unsolicited retained context.); Persisted cron jobs selected on timer ticks: known — push [wired] (A prior request schedules the job, but the later scheduler independently selects and supplies its retained prompt.); Promoted Markdown at external intended host: uninspected (The output writer and intended reader are recorded, but no loader/selection operation establishes pull versus push.); External Pi-retained session payloads, compaction summaries and alternative checkpoints: uninspected (Gateway continuation and extension branch-detail access are recorded, but external host serialization, checkpoint production and later selection are uninspected; Read-back direction cannot be completely classified.)	partial coverage (Known push operations use identifier and coarse selection. Requested recall lexical/numerical selectors are pull operations, not inferred push signals. Unknown loaders prevent axis inapplicability or complete coverage.); Branch replay selecting nuggets tool-result details: known — identifier [wired] (Lifecycle trigger selects current-branch messages whose tool-result identity is nuggets and whose details contain facts; this selection feeds subsequent unsolicited context.); Start hydration and per-turn local-fact formatting: known — coarse [wired] (Session start enumerates default-kind facts and fills missing keys; before-agent formatting selects prefix groups and a final-ten-file budget, omitting underscore entries.); Cron prompt emission: known — coarse [wired] (Each tick uses enabled state and five cron fields against the clock to select retained prompts; job ID metadata/recipient routing does not itself establish identifier matching of delivered content.); Requested recall/list/overlay operations: inapplicable (Signal classifies push selection only; lexical matching and HRR decoding on a requested return do not populate this unit.); External promoted-Markdown loader: uninspected (No recorded consumer selection resolves its direction, trigger or targeted signal.); External Pi-retained session payloads, compaction summaries and alternative checkpoints: uninspected (Gateway continuation and extension branch-detail access are recorded, but external host serialization, checkpoint production and later selection are uninspected; Read-back signal cannot be completely classified.)	partial coverage (Known automatic trace-fed routes support yes without observed benefit. No local no finding is asserted without a bounded absence record. Model proposal provenance, some precompaction-path durability, initial content and host checkpoint production remain unresolved.); Automatic preference/learning fragment writes with later prompt/recall consumer: known — yes [wired] (Input activity becomes durable short guidance/context and reaches later turns; success and improved capacity remain unobserved.); Tool-result path capture with later Active Files consumer: known — yes [wired] (Captured attempted tool paths are retained and later supplied as context; source tool success is not checked.); Precompaction task extraction and selected path retention: partial — yes [wired] (Task text has JSON retention plus requested later consumption. Locally copied paths reach later injection, but independent durability depends on another result-hook write or host-retained details.); Automatic session-marked recall usage counters feeding promotion selection: known — yes [wired] (Saved access metadata changes later promotion eligibility. Counters are not truth evidence; the resulting external Markdown consumer remains only afforded.); Explicit model remember and schedule proposal provenance: not-determinable (Automatic admission is known but the proposals original content may be authored, imported or trace-derived inside the uninspected model. No complete trace-fed derivation is recorded.); Initially loaded facts/jobs, generic callers and existing Markdown: not-determinable (Missing original admission and derivation prevents qualifying or excluding automatic trace-fed writes.); External Pi-retained session payloads, compaction summaries and alternative checkpoints: uninspected (Gateway continuation and extension branch-detail access are recorded, but external host serialization, checkpoint production and later selection are uninspected; Trace learning cannot be completely classified.)	partial coverage (Authority follows each actual consumer: knowledge, instruction, learning, ranking, routing, validation and enforcement have local warrants. Exported context remains afforded and opaque host payload effects uninspected; static instructions and unrelated ingress controls are not memory findings.); Retained factual/path/task text at later Pi/model prompt or requested recall: known — knowledge [wired] (Delivery supplies reference/context without factual acceptance or obedience guarantees; _task has requested rather than ordinary push delivery.); Retained preferences as repeated host guidance: known — instruction [wired] (Preferences are supplied as standing behavioral directions, independently of extraction faithfulness.); Retained traces at automatic extraction and compaction update consumers: known — learning [wired] (Recorded input/tool/session activity feeds durable context artifacts with later consumers; this does not demonstrate improved capacity.); Retained keys/values and HRR state at recall matching/decoding: known — ranking [wired] (Fixed selection code consumes retained content and numerical features as ranking inputs; transient confidence itself is not accumulated memory.); Retained fact keys and named-store organization at dispatch/formatting: known — routing [wired] (Keys and store names select scopes and display sections, not factual admissibility.); Retained hit metadata at promotion writer: known — validation [wired], enforcement [wired] (The retained usage threshold governs candidate admissibility, not truth validation; missing extension session IDs limit normal activation.); Retained cron controls at scheduler and router: known — routing [wired], enforcement [wired] (These fields select later operation and recipient and gate whether a job fires.); Retained scheduled prompt at later Pi/model: known — instruction [wired] (The stored request supplies behavioral directions at its actual model consumer; no compliance guarantee follows.); Promoted entries at intended external Claude consumer: known — knowledge [afforded] (Actual loading and activation are outside the inspected implementation; only intended context supply is afforded.); External Pi-retained session payloads, compaction summaries and alternative checkpoints: uninspected (Gateway continuation and extension branch-detail access are recorded, but external host serialization, checkpoint production and later selection are uninspected; Behavioral authority cannot be completely classified.)	AAS-2026-10-14-nuggets-c2e87e6916cc-01
o-o	2	Retained local article prose/structure, embedded image payloads and captions, operative manifest, source/fact cache, changelog, operator-edited instantiated contract and shared-code replacements, rebuilt library projection, and browser oo-theme preference, with their admission and later-consumer routes. Excludes unchanged shipped instructions/machinery as accumulated memory, transient image downloads, external research pages, provider weights and external Claude session stores. Host handling of requested reads and possible compaction/checkpoints remains an explicit unresolved delivery/trace-route limitation, not a claim about all runtime storage.	known coverage (The inventory covers retained document parts, operator-replaced machinery, the library projection and UI preference. External Claude sessions, provider weights and host checkpoints are outside this shipped-memory boundary; they are not inferred from file reuse.); Article, manifest, instantiated contract, source/fact cache, changelog, embedded images, rebuilt index, and operator-replaced shell/JS/CSS: known — files [wired] (All included document parts persist in local polyglot files; encoded image data and JSON are not separate substrates.); Browser oo-theme preference: known — kv [wired] (The key/value mechanism is inspected; browser-origin deployment and actual persistence are not observed.)	known coverage (Admission mechanisms, rather than physical writes or authorship, are classified. Actual sample production remains unknown under lineage, but shipped creation, edits, projection, research admission and preference writes have named control owners.); Human creation parameters, explicit configuration supply/removal, shared-code propagation and theme choice: known — manual [wired] (A person explicitly supplies/selects the retained content or replacement; software applying it does not make admission automatic.); Template defaults/empty-state emission, index rebuild and agent-selected research edits: known — automatic [wired] (Starting these workflows is not per-content approval; agent-selected content is admitted directly without a separate human veto.)	known coverage (Inventory covers document/cache requests, library/browser reads, automatic projection selection and UI preference replay. The external host’s treatment after a requested read remains uninspected, not a new established local read-back route.); Requested document/cache read by later research agent: known — pull [claimed] (The path and contract request retained material; neither the filename nor returned read is another unsolicited push.); Operator-requested library rebuild/browser article or index viewing/configuration display: known — pull [wired] (Requested delivery is pull even though code renders it.); Automatic selection within index rebuilding, shell freshness extraction and browser theme replay: known — push [wired] (Separate selector operations choose retained input within the requested workflow, without a request for those particular retained parts.)	known coverage (Signals classify the automatic selectors only. Requested filenames and catalog identifiers alone do not establish targeted push; index card selection is coarse and theme replay performs an actual key lookup.); Index rebuild trigger: discovered files and dates select retained entries/excerpts; later dispatch trigger: first freshness metadata supplies gate input: known — coarse [wired] (Availability, date ordering, first-match extraction and eight-card limit are broad structural selection, not inferred semantic retrieval.); Browser initialization trigger: localStorage lookup selects saved oo-theme preference: known — identifier [wired] (The key is the actual selector of the delivered retained setting, not merely a label on a requested document.)	not-determinable coverage (No accepted bounded-absence record establishes a local no, and no accepted trace-fed write establishes yes. Research persistence and attributed change summaries are insufficient; external compaction remains uninspected rather than negative.); Research article/cache and product-change summaries with later whole-file consumption: not-determinable (Research outputs have unverified execution provenance. Change summaries do not supply a recorded trace-to-derived-artifact chain or evidence of their use as update experience; neither yes nor bounded no is warranted.); Index, operator-maintained guidance/machinery and UI preference routes: not-determinable (The records identify deterministic projection or explicit operator supply, not qualifying automatic trace-fed writes. They do not declare a searched bounded absence for all trace-learning alternatives.); Host compaction/checkpoint alternatives affecting later file-read context: uninspected (External host session retention, summaries and checkpoints are uninspected. No local profile value can be inferred from their opacity; this names a delivery limitation, not inclusion of all provider storage.)	partial coverage (Authority is assigned at consumers, not from storage labels. Wired shell/browser effects and requested model effects retain separate bases. No inspected consumer establishes image truth, trace-fed learning authority or external budget enforcement.); Prior article, sources/facts and change summaries at the later research agent: known — knowledge [claimed] (The later whole-file/cache read is requested, not observed delivery or dependence.); Operator-edited instantiated contract at the research agent: known — instruction [claimed] (Scope includes retained operator changes, not static shipped instructions standing alone.); Manifest dates and interval at direct/batch freshness gates: known — routing [wired], enforcement [wired] (The same metadata both selects dispatch and supports the local skip decision; force is an explicit alternate path.); Manifest as_of at the later research agent: known — routing [claimed] (Date-focus instruction is weaker than the executable freshness gate.); Article/reference text, manifest display and library excerpts at browser/human readers: known — knowledge [wired] (Rendering delivers reference content; no factual acceptance is inferred.); Retained document dates and displayed text at index/browser ordering and filtering consumers: known — ranking [wired] (Ranking concerns visibility, not research quality or a learned policy.); Theme value at later browser initialization: known — routing [wired] (The saved string selects presentation state only.); Operator-replaced shell at future Bash invocation: known — enforcement [wired] (Replacement code is an executable control, not advisory prose; individual new payload behavior is not inspected.); Changed contract budget at external CLI enforcement, replacement JS/CSS payload effects, and decoded image semantics: not-determinable (Budget is passed but external enforcement is uninspected. Arbitrary replacement presentation payloads and decoded images lack enough recorded consumer-effect detail for further controlled authority findings.)	AAS-2026-10-15-o-o-b6fd0055e5c1-01
oh-my-pi	2	Included: accumulated session journals and textual/opaque/image checkpoints, branch notebooks and persisted recall; local extraction/intermediate/publication/lesson parts; Mnemopi working/episodic/fact/numerical parts and alternative CLI/MCP/module routes; Hindsight target-side retention/recall and stable mental-model cache with opaque backing limits; Sharpshooter deltas/decision files; managed skill bodies/descriptions; optional child session/artifact later-consumer routes. Untraced producer, admission, advisor, import and host alternatives remain included unresolved units. Excluded: static shipped prompts/instructions, ordinary current-run return/control state, generic project work products not established as memory, external service/provider implementation and model weights. Their opaque operative backing is retained as uncertainty, not silently classified from readable displays.	partial coverage (File, process and SQLite witnesses cover named inspected parts, not opaque backing or alternative shipped storage paths.); JSONL journal/checkpoint/notebook/recall parts and persisted child sessions: known — files [wired], in-memory [wired] (JSONL persistence is the inspected journal branch; provider replay is stored metadata, not inspected provider weights.); Local stage-one outputs and maintenance access state: known — sqlite [wired] (The shared job database is separate from published guidance files.); Local published memory, lessons, procedures, scripts, templates and examples: known — files [wired], in-memory [wired] (All named publication buckets are files; the cached instruction snapshot is another operative part.); Mnemopi working/episodic/fact/JSON-embedding/binary-vector retained parts: known — sqlite [wired] (Numerical content in a SQLite column does not establish a separate vector storage service.); Sharpshooter queued deltas and project decision replacements: known — files [wired] (Queue and published decisions have separate admissions but the same file substrate.); Managed skill body and sanitized description: known — files [wired] (Canonical managed object after reconciliation; a directory name does not establish a prompt-registry service.); Hindsight stable mental-model rendered cache: known — in-memory [wired] (Only the local cache substrate is resolved; external backing is a separate unresolved part.); Hindsight upstream retained bank/mental-model state and provider-native checkpoint state: uninspected (Service requests and local replay metadata do not reveal the internal operative storage of excluded services/providers.); Snapcompact resolved frames/archive backing and alternative Redis/SQL journals, imports and other shipped host loops: uninspected (Archive references are retained, but full frame backing and alternative storage implementations are untraced; no storage-parity or exhaustive union.)	partial coverage (Automatic writes and a distinct conditional human-approval route are supported; initialization, provider/core and alternative API controls prevent complete coverage.); Local extraction/consolidation, intermediate synchronization, Sharpshooter extraction/replacement and private capture admission: known — automatic [wired] (Starting a session or supplying a triggering user sentence is not approving the transformed retained content.); Adapter transcript submission and first-turn recall persistence: known — automatic [wired] (The adapter controls automatic submission/commit; excluded server admission is assessed separately.); Foreground model writes when generic policy auto-allows: notebook, lesson, memory edit, managed mutation: known — automatic [wired] (Conditional software-admitted route; generic caller authorship remains unresolved on other invocations.); Explicit retained-write invocation when wrapper requires and obtains human approval of effective parameters: known — manual [wired] (A wired conditional approval mechanism, not evidence that any particular deployment enabled or exercised it; no semantic-correctness oracle.); Checkpoint producer/extension replacement admission, initial notebook/lessons/managed content and generic explicit save callers: uninspected (Snapshot commit checks are known, but exact admission control of every producer/initial content and generic caller identity is not.); Mnemopi core numerical/fact/episodic admission and alternate CLI/MCP/module mechanisms; Hindsight server/seed admission: uninspected (Invoked core maintenance and opaque server/seed routes do not resolve all per-content control or alternative caller authorization.)	partial coverage (Both automatic supply and explicit demand are wired, but untraced alternatives prevent a complete read-back inventory.); Later context reconstruction, local snapshot, decision injection and stable mental-model cache delivery: known — push [wired] (Runtime selectors supply retained parts without a foreground request for those parts.); First-turn automatic query recall, persisted reuse and Mnemopi change notices: known — push [wired] (An ordinary coding prompt is not an explicit recall request; the runtime performs automatic selection.); Explicit recall/reflect, local memory URL, Mnemopi ID, raw-history and notebook read requests: known — pull [wired] (Requested delivery is one pull, not an additional push merely because code returns text.); Selected managed skill invocation body read: known — pull [wired] (Invocation caller chooses the target; builder reads its body. Complete downstream dispatch remains unobserved.); Managed metadata full prompt supply, child artifact subsequent retrieval, compaction-time recall consumers and other host/advisor alternatives: uninspected (Discovery/preparation and pointer rendering do not establish every later delivery chain; API availability does not identify an unspecified later consumer.)	partial coverage (Signals characterize actual push selection only. Coarse filtering and identity matching remain separate; known pull routes do not make unresolved push alternatives inapplicable.); Automatic branch/checkpoint/notebook reconstruction: known — identifier [wired], coarse [wired] (Trigger: rebuild/resume/commit; selector: active leaf path, checkpoint IDs/reset and byte budgets; selected parts: checkpoint, kept tail and latest valid notes.); Automatic local summary/lesson and Sharpshooter decision instruction supply: known — coarse [wired] (Trigger: instruction/snapshot assembly; selector: project memory root, populated fixed files and token budget; selected parts: summary then lessons, or project decision files.); Hindsight stable mental-model cache publication/delivery: known — coarse [wired] (Trigger: startup/new transcript/explicit reload; selector: bank/tags, configured rendering bound and generation/publication status; selected part: rendered stable block. Upstream selection is separate.); Persisted first-turn recall reuse and Mnemopi tracked change notice: known — identifier [wired], coarse [wired] (Trigger: resumed/later prompt; selector: scope and tracked memory IDs/preview comparison; selected parts: persisted recall or changed-memory notice.); First-turn Mnemopi query relevance and Hindsight query/mental-model upstream selection: uninspected (Trigger/query composition/bank filters/budget are recorded; actual lexical/embedding/judgment ranking selector and opaque upstream selected payload are unresolved. FTS schema alone cannot classify push selection.); Managed metadata automatic supply and other included host/advisor push alternatives: uninspected (Enable/name filters establish discovery eligibility, not a complete unsolicited metadata delivery path. An ID on requested file/skill body is not identifier-based push.)	partial coverage (Three positive wired chains coexist with unresolved compaction/core/service/alternative paths. This axis means trace-fed durable update, not improved future capacity or exercised self-improvement.); Local rollout extraction to summary/procedures and later injection/read: known — yes [wired] (Automatic traces become durable behavior-shaping prose with inspected later supply routes; no observed improvement is required.); Sharpshooter user-turn extraction to durable project instructions: known — yes [wired] (Occurrence gates do not establish truth, but the update-to-later-guidance chain is wired.); Private autolearn capture to durable lessons/managed skill invocation: known — yes [wired] (Qualifies when capture elects an admitted write; the conditional branch is wired, not observed beneficial learning.); Text/provider/image compaction alternatives and notebook-directed context replacement: uninspected (Retained continuation context qualifies if automatically trace-produced, but producer algorithms/admission alternatives are not fully recorded. Notes durability alone does not establish an automatic trace-fed route.); Mnemopi core fact/episode/embedding transformations and Hindsight upstream updates/mental-model seeding: uninspected (Raw transcript remember/service submission alone does not prove a durable behavior-shaping transformation; core/upstream chain and initial seed controls remain unresolved.); Other foreground/CLI/MCP/module writes, imports, child revival and untraced host learning alternatives: uninspected (Generic caller identity/initial provenance and untraced derivation prevent an exhaustive qualifying-route inventory. No bounded absence record supports a local or system no.)	partial coverage (Authorities attach to consumed retained parts. Known update learning and downstream knowledge/instruction coexist with unresolved numerical/opaque/execution consumers.); Local memory summary and lessons at foreground/child/advisor guidance consumers: known — knowledge [wired] (Heuristics/process guidance is explicitly subordinate to current evidence; not a hard veto.); Sharpshooter decision files and Hindsight stable mental-model cache at inference consumers: known — instruction [wired] (These retained parts supply standing guidance, not executable enforcement.); Reconstructed journal/summary/notebook, recall blocks and explicitly read memory artifacts at later models: known — knowledge [wired] (This unit covers readable supplied context; it does not claim execution of code or semantic force of opaque replay.); Managed skill body at invocation-message preparation: known — instruction [wired] (The builder loads body guidance; complete dispatch and compliance are not observed.); Managed applicability description at skill choice: known — routing [afforded] (Description text can guide model choice; actual complete prompt dispatch and selection are unobserved.); Retained traces/intermediates at local consolidation, Sharpshooter consolidation and capture update consumers: known — learning [wired] (Learning here is authority at an update consumer, not demonstrated improved capacity.); Retained branch/checkpoint/reset/recall scope and ID metadata at reconstruction/tracking consumers: known — routing [wired], validation [wired], enforcement [wired] (Only consumed controls are classified; idle descriptive metadata has no authority.); Retained Mnemopi scope/retirement/store metadata at live lookup and edit eligibility: known — validation [wired], enforcement [wired], routing [wired] (Store-specific eligibility and retirement are separate from factual warrant; propagation to copies remains unresolved.); Retained local job tokens/watermarks/cwd and queued decision evidence at maintenance consumers: known — routing [wired], validation [wired], enforcement [wired], knowledge [wired] (Metadata shapes update ownership/input selection; evidence/rationale supplies interpreted context.); Mnemopi text ranking integration, numerical representations and alternative validation/shared-memory consumers: uninspected (FTS index inputs are recorded, but actual ranking integration, numerical consumer and alternate admission consumers are not traced; no inferred ranking authority.); Opaque provider replay, image recovery, external Hindsight updates and generated script execution/template rendering: uninspected (Readable context witnesses cannot upgrade opaque/unknown consumers to instruction, enforcement or learning; runtime permissions/static templates are excluded memory controls.)	AAS-2026-10-14-oh-my-pi-c0ffc54ff768-01
okf-harness	2	Retained registered file bytes and URL pointers; source manifest; wiki prose and retrieval/citation fields; reconciliation edge IDs and notes; persisted graph/backlinks reports; Git checkpoint trees and judgments; their admission, maintenance and later-consumer routes. Includes unresolved initial content, opaque captured files and generic caller control. Excludes unchanged shipped guidance provisioning, runtime-version configuration, transient evidence/check/doctor returns, reserved legacy logs as answer memory, external host session/compaction/provider state and model weights. Restored instructions/configuration remain included as checkpoint payload, not newly inferred online instruction learning.	known coverage (The recorded retained-object inventory uses local files and Git objects. Source encoding opacity does not obscure its disk substrate; external provider/session state and static provisioning are excluded.); Registered raw bytes, URL pointers and manifest: known — files [wired] (Local raw files and JSONL remain available to later source reading and checks.); Wiki prose and parsed retrieval fields: known — files [wired] (Markdown files retain both semantic content and access structures.); Reconciliation edge IDs and notes: known — files [wired] (Local JSONL retains the two distinct parts.); Persisted backlinks JSON and graph HTML: known — files [wired] (Generated reports are local files; an edge encoding does not make their substrate a graph database.); Checkpoint trees and judgment bodies: known — repo [wired] (Git retains versioned trees and commit bodies; repo classifies this recovery store, separately from working files.)	partial coverage (Automatic software/model admission and explicit human write-back permission are supported locally. Generic callers, historical content and restoration control remain unresolved.); Registration and manifest admission: known — automatic [wired] (Code applies registration checks and writes/reuses entries without per-content human approval; starting source ingestion is not approval of each retained part.); Ingest/reconcile wiki edit admission: known — automatic [claimed] (Guidance assigns proposal and adoption to the agent. The start request and occasional stop contract are not mandatory per-claim human decisions.); Conversation answer write-back admission: known — manual [claimed] (Explicit user yes supplies the decision to retain that answer; no/silence persists nothing under the protocol.); Derived graph report admission: known — automatic [wired] (Fixed code generates reports without a human approving individual nodes or edges.); Reconciliation acknowledgment and note admission control: not-determinable (Core rejects invalid IDs/empty notes, but generic caller identity does not establish human approval versus model admission.); Baseline/checkpoint/restore admission control: not-determinable (Caller selects judgment/completion; human-versus-agent content decision is not captured, including replacement by restoration.); Existing wiki and initial source-content admission: not-determinable (Neither scans nor external input bytes establish the original admission mechanism.)	partial coverage (Every established delivery is requested pull. Ledger-note semantic read-back remains unresolved; static host skill discovery and external host memory are excluded rather than recast as push.); Requested ingest metadata planning: known — pull [wired] (The wired finding covers only requested metadata planning and its returned recommendations.); Agent source/wiki reading under ingest and reconciliation guidance: known — pull [claimed] (Direction is resolved for the delivered agent protocol, with a claimed basis distinct from executable metadata planning; actual agent compliance is unobserved.); Evidence briefs, direct search/read and continuation: known — pull [wired] (User question leads to a host-issued retrieval request. Lexical selection and delivery within it are not another unsolicited push.); History and selected-tree recovery: known — pull [wired] (Caller requests history or restoration, and later requested workspace operations consume restored files.); Graph report browser inspection: known — pull [afforded] (Caller requests generation/opening; a persisted report affords later requested inspection.); Requested snapshot/check consumption of manifest/edge state: known — pull [wired] (Checks and evidence invocations reload access/bookkeeping state as part of the requested operation.); Semantic consumption of reconciliation note reasons: not-determinable (Retention and parsing establish no later semantic delivery route; no direction is invented.)	not-determinable coverage (No established accumulated-memory route is a push, but unresolved note consumption prevents a complete pull-only assessment for the entire inventory.); All established source/wiki/check/history/graph pull chains: inapplicable (Within these recorded requested routes there is no push selector to classify. Query words, read IDs and budgets are not push signals.); Unestablished reconciliation-note semantic consumer: not-determinable (Without a later delivery/selection route its direction and any push-selection input cannot be classified.)	not-determinable coverage (No qualifying automatic trace-fed write is established, and the accepted set declares no bounded absence record licensing a local no. Arbitrary input content, initial lineage and generic caller control prevent a stronger assessment.); Deterministic graph generation and Git tree capture: not-determinable (The recorded mechanisms compile links and snapshot current files rather than describing trace distillation, but no accepted bounded absence record supports a controlled no finding.); Explicitly accepted conversation write-back: inapplicable (This specific delivered protocol requires human approval of retained content, so it is outside automatic trace-fed admission. Its later knowledge route remains supported.); Agent ingestion of arbitrary source content and wiki revision: not-determinable (Automatic agent adoption is requested, but original source content is arbitrary and actual semantic transformation is unresolved. Records do not establish whether any included input is a qualifying activity trace.); Generic caller notes, judgments and initial retained content: not-determinable (Caller identity and derivation inputs are unresolved; retention of rationale alone establishes neither an automatic trace-fed write nor its qualifying later consumer.)	partial coverage (Authority is split by actual retained part and consumer. Reconciliation note semantics have no established later consumer, preventing complete classification.); Registered bytes to integrity checks: known — validation [wired] (Stored hash/bytes supply byte-identity criteria; this validates integrity, not source truth.); Registered inputs to ingest/reconcile agent: known — knowledge [claimed] (Requested source reading supplies evidence for semantic edits, not privileged directions.); Wiki prose to answering agent: known — knowledge [wired] (Selected prose supplies context for later answers. Delivery is wired; dependence and benefit are unobserved.); Wiki prose/title/description/tags/index at lexical scorer: known — ranking [wired] (Retained strings influence relative retrieval priority under fixed lexical code.); Manifest IDs/paths and wiki parsed fields at lookup and planning: known — routing [wired] (Retained identities select references, targets and planning candidates.); Manifest uniqueness and stored provenance at admission/integrity consumers: known — validation [wired], enforcement [wired] (Recorded IDs/paths/hash constrain admissibility and integrity checks; duplicate identities can reject ordinary registration.); Wiki source/citation identities at evidence seal consumer: known — enforcement [wired] (Retained links identify concepts excluded from planner delivery when source risk is detected; this is planner-local enforcement, not universal read denial.); Reconciliation exact pair at snapshot currency consumer: known — routing [wired] (Acknowledged pairs select which suspected edges remain pending; they do not validate changed claims or gate evidence merely because currency is false.); Retained reconciliation note semantics: not-determinable (Only a nonempty-string criterion and ID-driven filtering are established; no later semantic consumer licenses knowledge or learning authority.); Checkpoint identities and tree at restore: known — routing [wired] (Selected completion chooses the restored working state used by later commands. Individual restored parts retain their own downstream forces.); Checkpoint judgment at history caller: known — knowledge [wired] (History exposes rationale as information for caller interpretation and restore choice, not as a scored learning signal.); Persisted graph display to browser reader: known — knowledge [afforded] (Generated navigation and labels inform the requested browser reader; actual opening is unobserved.)	AAS-2026-10-15-okf-harness-6bf5507fbdbf-01
openclerk	2	Retained and revised vault Markdown, source assets/notes, SQLite mirrors/chunks/typed and graph projections, freshness/provenance, semantic vector caches, retrieval regression cases, changed installed module configuration at later dispatch, and shipped evaluation host-session continuation. Includes unresolved original-asset consumers, initial admission, untraced maintenance/report alternatives and opaque host checkpoints. Excludes static shipped skills/contracts, transient unadopted planning/report output, ordinary current-call state, external provider model weights/training and private deployment contents.	partial coverage (The recorded local objects have file and SQLite witnesses. The resumed host payload is included but its storage is opaque; no complete substrate union is established.); Canonical vault notes, raw source assets, semantic cache and regression cases: known — files [wired] (These retained objects are files; JSON and numerical encoding do not add a storage substrate.); Document mirror, chunks, typed projections, graph relationships and freshness/provenance state: known — sqlite [wired] (The graph is represented in SQLite; graph structure is not a separate demonstrated graph storage engine.); Installed module configuration used by subsequent semantic dispatch: not-determinable (Configuration persistence is established, but its physical substrate is not specified in this accepted record.); Resumed evaluation host session, including possible checkpoint/summary parts: uninspected (A session ID selects external retained state. Its physical storage is not inspected. Provider weights themselves are outside this accumulated-memory inventory.)	partial coverage (Software admission is supported on particular transformations; generic caller requests and an approval-mode label do not establish explicit human per-content decisions.); Vault parse/import and derived chunk/projection admission: known — automatic [wired] (Initial authorship is separate from software admission into the mirror and derived stores.); Source-note rendering, compiler assembly, admitted audit repair, cache rebuild and retrieval-row construction: known — automatic [wired] (An upstream request does not make the downstream software-selected content manual.); Caller-supplied original note content, replacement, move and configuration admission: not-determinable (Callers may be human or agent. Doctrine asks for approval but the mode check does not authenticate a human content decision; manual admission cannot be asserted.); Separate adoption of session candidate into a durable document: not-determinable (The planner writes no document and leaves adoption to a subsequent caller; no automatic end-to-end adoption or authenticated human choice is established.); Resumed host-session persistence and alternative checkpoint replacements: uninspected (The driver resumes an ID but does not expose internal content-write decisions.)	partial coverage (All inspected external retrieval deliveries fulfill explicit requests. Internal selection is distinguished from a second unsolicited delivery. Host continuation has opaque selection and delivery internals.); Context packs, primitive projection/document reads, policy recall and explicit semantic retrieval: known — pull [wired] (A task/query or explicit action requests retained material; fixed-path or embedding selection inside that chain is not by itself push.); Explicit regression replay and audit/compiler evidence reads: known — pull [wired] (Requested replay/maintenance calls consume retained cases and source state.); Original PDF/video metadata assets outside extracted-note reads: not-determinable (Retention is established but no later consumer route for these original payloads is established by the record.); External resumed host payload and untraced compound-report/host retention routes: uninspected (ID-based resume is wired; which retained parts are delivered by request versus independent host supply is not exposed.)	not-determinable coverage (No positive push selector is established in the inspected request routes. The included opaque host alternative prevents a completely assessed pull-only boundary and therefore prevents axis-level inapplicability.); Inspected explicit context, primitive, policy, semantic and replay request chains: inapplicable (These named operations fulfill retained-material requests rather than push. Requested IDs and fixed policy paths are not evidence of identifier-based push.); Original source assets with no established later delivery: not-determinable (No push trigger, selector input and selected original payload are recorded.); Resumed host-session selection and opaque checkpoint/compaction delivery: uninspected (The session ID alone does not establish an identity match selecting delivered retained parts. No controlled push signal can be assigned.)	not-determinable coverage (No accepted record establishes the complete qualifying automatic trace-to-durable-behavior-shaping-artifact chain. This is not bounded system-wide absence: host checkpoints and separate session adoption remain unresolved.); Explicit session artifact planning followed by separate possible lifecycle adoption: not-determinable (Planning has no durable candidate write. Neither automatic durable adoption nor bounded absence of such adoption is established across subsequent callers.); Retrieval execution capture and later regression replay: not-determinable (Recorded rows are reduced traces used as selectors/comparison baselines. They do not establish a new durable behavior-shaping update. No absence record warrants a controlled no finding.); Corpus-derived caches and synthesis/source maintenance: not-determinable (Automatic derivation exists, but the records do not establish qualifying trace-fed inputs on these paths. Imported transcripts and original notes have unresolved activity/content provenance.); Resumed host-session checkpoint, summary or compaction writes: uninspected (Later continuation is wired, but automatic summary production, retained payload and later consumption chain are opaque. This included alternative cannot be classified no.)	partial coverage (Classifications name actual retained-part consumers. Shipped static skill instructions are excluded; code-authored policy summaries do not establish consumption of all retained policy prose as standing instructions.); Retrieved note/chunk prose, typed fact values and synthesis summaries delivered to the external agent: known — knowledge [wired] (Knowledge identifies evidence/context supply, not observed compliance or answer improvement.); Retained chunk text and cached chunk vectors at lexical/semantic ranking: known — ranking [wired], ranking [wired] (Fixed scoring code can confer ranking force on retained content.); Stored document fields, projection source classes and module configuration at route/target selection: known — routing [wired], routing [wired] (Source roles select copied input; installed configuration selects later module execution, not model learning.); Module enabled state and verified binding at subsequent configured execution: known — enforcement [wired], validation [wired] (These controls act on the runner module path, not every possible direct executable call.); Cache identity metadata at vector reuse admission: known — validation [wired], enforcement [wired] (Stored identity fields are compared against the current request/corpus; rejected entries cannot supply reused vectors.); Captured retrieval query/action/filter and baseline result identities at replay: known — routing [wired], validation [wired] (The trace has operative replay/comparison consumers, but no recorded automatic behavior-shaping update.); Freshness reasons and policy evidence at audit/report consumers: partial — knowledge [wired] (Recorded metadata informs warnings and source-role selection. Full policy interpretation and every composite report conversion remain unresolved.); Host-retained session parts at resumed Codex turn: uninspected (Continuation is wired, but consumed payload parts and their distinct effects are not inspected.)	AAS-2026-10-16-openclerk-ec0e00c2bb0a-01
OpenSage	2	Use-accumulated sandbox wiki/TODO/pins/findings and tool-output files; session trajectories, compaction/tool summaries, retained hook/image parts; dynamic definitions/instructions and checkpoint selectors; shipped exploration/vulnerability handoffs, evaluation traces and retained scores/grades. Includes unresolved ADK service/provider retained state, live retention, trainer checkpoint/update and custom-extension/remaining benchmark alternatives. Excludes static shipped root/maintenance instructions, ordinary inference weights without a retained update, task-source code graphs and repository data not shown to accumulate through use, and external implementation bodies; their included interfaces remain uncertain.	partial coverage (The inventoried file/process stores are supported; opaque alternatives and incomplete extension/benchmark inspection prevent complete substrate enumeration.); Sandbox wiki, task notes, pins, saved outputs/findings; host trajectories, selectors, definitions and vulnerability checkpoints: known — files [wired] (Records identify sandbox Markdown/output files and host JSON/TOML checkpoints; serialization format is not a substrate.); Live event/compaction state, pending hook/image parts and restored objects: known — in-memory [wired] (These retained parts also reside in process session events and pending lists; persistence is conditional.); Evaluation traces and retained DevOps grades: known — files [wired] (Exported traces, score.json and grade files have later evaluation/report consumers.); Constructed ADK memory service handle: known — in-memory [wired] (The record establishes dependency-owned process memory, not its payload encoding or a consumer route.); Included ADK service/provider state, live retention, external trainer and extension-specific alternatives: uninspected (Provider caches, learned checkpoints and extension-specific retained storage are not classified by a service name or integration handle.)	partial coverage (Automatic routes are positive; no recorded explicit per-content human decision supports manual. Unresolved writers prevent an automatic-only conclusion.); Software seeds, checkpoints, dynamic-definition admission, compaction and tool reduction: known — automatic [wired] (Software/model admission has no per-content human decision; workflow startup is not manual admission.); Finding append/suppression, hook/image retention and benchmark handoffs/grades: known — automatic [wired] (Code/model accepts content under local predicates, not explicit operator approval of each retained item.); Foreground-agent wiki/TODO/pin writes: known — automatic [afforded] (The named model-directed maintenance branch admits its own text through granted file tools.); Other filesystem writers, initial non-seed content and externally replaced checkpoints: not-determinable (Generic writer identity and the possibility of file replacement do not prove a manual decision or automatic replacement mechanism.); Included ADK service/provider state, live retention, external trainer and extension-specific alternatives: uninspected (External trainer acceptance and custom extension write controls are uninspected.)	partial coverage (Both requested retrieval and automatic supply occur in distinct operations. Opaque alternatives prevent complete direction coverage.); Wiki/TODO/full tool-output and shared finding file requests: known — pull [afforded] (A named later agent requests file contents; a delivered path alone is not content retrieval.); Requested resume, child continuation/inheritance and explicit history consultation: known — pull [wired] (Requests select retained sessions/history. Their requested returns are not separate push operations.); History folding/pin embedding, reduced-response future history, retained hook/image delivery: known — push [wired] (Automatic selectors supply retained context without a content-retrieval request, conditional on plugins and later invocation.); Finding broadcast and benchmark explore/findings next-agent handoffs: known — push [wired] (Code supplies selected prior results to peer/main/PoC context; a configured stage request does not make automatic prompt inclusion a file pull.); Included ADK service/provider state, live retention, external trainer and extension-specific alternatives: uninspected (Constructing an ADK memory service does not establish retrieval; live/provider/trainer and custom-extension later routes remain uninspected.)	partial coverage (Push signals have local triggers and retained targets. No inferred lexical/embedding signal is supported, but missing inspection is not their absence.); Later history fold selects compaction markers and original events: known — coarse [wired], identifier [wired] (Later context construction uses timestamp-range/ordering and branch/invocation matching to select delivered retained parts; range selection is separate from identity matching.); Successful compaction appends current pinned file and output/history plugins supply retained context: known — coarse [wired] (Triggers are successful compaction, eligible tool result/later history, hook matcher/queued content and image request construction. Availability, nonempty text, length and image-size budgets select retained parts; file IDs alone are not targeted push.); Tool-summary model selects retained context according to inferred task intent: known — inferred-judgment [wired] (Threshold triggers summarization; recent retained events and tool arguments supply inferred intent for selected summary content; deterministic prefix branch remains coarse.); Finding peer broadcast and benchmark handoff selection: known — coarse [wired] (Nonduplicate branch broadcasts to all peers; enabled explore selects final/fallback text; resume/nonempty findings supplies structured content. No semantic ranking or matched content identity is established.); Included ADK service/provider state, live retention, external trainer and extension-specific alternatives: uninspected (Selectors for opaque service/live/provider/trainer/extension pushes are uninspected. Pull file names or checkpoint IDs do not establish identifier-based push.)	partial coverage (Qualifying summary routes establish yes locally, not improved capacity or a complete learning inventory. Unresolved routes are not converted into no.); Automatic history summary retained in compaction events and delivered on subsequent folded-history invocation: known — yes [wired] (User/model/tool traces produce durable continuation context; later-consumer wiring establishes the axis without observed improvement.); Automatic tool-response summary/prefix retained in events and consumed by later model/history consultation: known — yes [wired] (Tool execution feedback plus recent session history produces retained context, including deterministic preview alternatives.); Explore quota-exception trace-to-summary-to-main-agent chain: known — yes [wired] (Recorded exploration turns/tool responses feed automatic summary construction, prompt admission and later main-agent consumption.); Wiki/pin authored entries, finding comparison, hooks/images, raw checkpoints/exports and ordinary final-text handoffs: not-determinable (Use-derived content or raw retention alone does not establish a trace-fed transformation into behavior-shaping material for every branch; actual content origins and downstream updates are unresolved. No bounded absence record licenses no.); Included ADK service/provider state, live retention, external trainer and extension-specific alternatives: uninspected (External parameter updates, live retention and extension-specific trace-fed routes remain uninspected.)	partial coverage (Distinct consumers retain distinct forces. Static root/memory policy is excluded; knowledge, instruction and learning are not interchangeable.); Retained summary and vulnerability descriptions at later model consumers: known — knowledge [wired] (Context/advice supplies information; delivery is not compliance or truth acceptance.); Wiki, task notes and findings at explicit future file readers: known — knowledge [afforded] (Policy documents future agent readers; exposed finding-file reads afford reference use without proving deployed activation.); Finding and contradiction text at internal peer-delivery consumer: known — knowledge [wired] (The inspected ordinary broadcast branch supplies advisory text; its internal wiring does not upgrade unresolved shipped invocation.); Retained dynamic instruction at spawned/rebuilt child model: known — instruction [wired] (Saved full_instruction supplies standing directions, unlike the static maintenance policy excluded here.); Pinned operational directions and retained hook guidance at later model history: known — instruction [wired] (These recorded guidance parts carry interpreted directions; arbitrary command output is not upgraded to instruction.); Recorded older user/model/tool history at summary-update consumers: known — learning [wired] (Recorded experience feeds automatic production of retained behavior-shaping context; downstream summary knowledge is a different effect.); Previous findings at duplicate-check admission consumer: known — validation [wired], enforcement [wired] (Prior text supplies duplicate criteria; the interpreted DUPLICATE_OF result vetoes addition. This does not validate truth.); Retained checkpoint/definition/range selectors and vulnerability stage fields: known — routing [wired] (IDs, metadata and finding presence select restored state, capabilities, history ranges or next task stage.); Retained benchmark decision fields at report/scoring consumers: known — ranking [wired] (Scores and parsed grade/match fields determine reported relative task performance; no learned ranking policy is inferred.); Included ADK service/provider state, live retention, external trainer and extension-specific alternatives: uninspected (Authority of opaque service state, trained parameters and arbitrary retained extension output cannot be resolved from transport alone.)	AAS-2026-10-14-opensage-291d1ed79d37-01
OpenViking	2	Retained raw session archives and externalized tool results; archive overviews and cumulative/legacy checkpoints; ordinary memory, case, trajectory, experience and skill files; their vectors, abstracts and symbolic metadata; recall ledger and OpenCode capture/retry state; write, maintenance, extraction and selected OpenCode/VikingBot consumption routes. Resource processing, alternate admission, snapshot/batch variants, bot local files and other shipped consumers remain included with unresolved coverage. Excludes transient query plans and recall digests as new memory objects, static shipped prompts/schemas, provider weights and third-party host scheduling. Runtime settings are not substituted for the retained-memory boundary.	partial coverage (Files and vector storage are supported witnesses, not an exhaustive substrate union for all included paths.); Session raw archives, overviews and checkpoint metadata: known — files [wired] (Covers the recorded archive parts, including cumulative and legacy checkpoint metadata, not native backend internals.); Externalized tool payload and synopsis/reference fields: known — files [wired] (Opaque payload semantics do not obscure the inspected storage route.); Ordinary memory, case, trajectory and experience files: known — files [wired] (Policy snapshots used only in current computation are not a separate accumulated in-memory store. Optional snapshot retention remains unresolved below.); Processed skill package content and auxiliary files: known — files [wired] (File storage is established even though auxiliary content and later invocation are not completely classified.); Memory index vectors, URI fields and abstracts: known — vector [wired] (Backend engine identity and convergence are unresolved, but the recorded access substrate is vector storage.); RecallLedger and OpenCode session mapping/pending queue: known — files [wired] (Recorded file persistence is distinct from deployment durability and full queue housekeeping.); VikingBot local MEMORY.md and HISTORY.md: known — files [wired] (Substrate is known; admission and subsequent use are not.); Remaining resource products, optional policy snapshots, native storage and alternate shipped memory paths: uninspected (The records retain partial resource/processor, snapshot, native and bot-loop coverage. They do not enumerate every retained object or concrete storage alternative.)	partial coverage (Automatic admission is supported on inspected routes; unresolved explicit and initial paths neither establish manual control nor its absence.); Archive/capture/offload/checkpoint/overview admission: known — automatic [wired] (A caller starting commit or a host scheduling compaction does not make subsequent admissions manual.); Ordinary memory application, compile proposals, experience and conditional skill updates: known — automatic [wired] (Covers inspected persistent branches; dry-run and generic batch updater alternatives are retained below.); Index update, recall-ledger record and automatic linked-memory cleanup: known — automatic [wired] (The caller decision to remove the primary resource remains separate from computed downstream cleanup.); Initial/plain memory content, training-case fast path, skill/resource installs, primary explicit forgetting and bot local file writers: not-determinable (Caller identity does not establish an explicit human content decision. A description requesting confirmation is not an enforced approval path; no manual finding is warranted.); Alternate resource/watch admission, policy snapshot/restore and batch updater variants: uninspected (Incomplete producer/admission coverage prevents extending the known automatic mechanisms to every included write.)	partial coverage (Both pull and push exist in different operations of the same chains; unresolved alternative consumers prevent complete coverage.); Requested search/find/context returns, session reconstruction and explicit tool-result reload: known — pull [wired], pull [afforded] (Wired requested context and afforded optional reload retain their distinct bases.); OpenCode chat-hook recall and VikingBot query-time memory/reminder augmentation: known — push [wired] (Conditional on gates/host hooks; not a claim of deployment activation or universal automatic recall.); Extraction-only tool-payload hydration: known — push [wired] (Hydration failure leaves the stub; public explicit reload is separately pull.); Untraced bot local-file, full skill invocation, alternative host and generic training/executor delivery paths: uninspected (Storage and generic successor flow do not enumerate every requested versus automatic consumer operation.)	partial coverage (Coarse, embedding, judgment and actual identifier selection have route-local warrants. Lexical retrieval on a requested route is not promoted into a push-signal finding.); OpenCode automatic recall relevance chain: known — inferred-embedding [wired], inferred-judgment [wired], coarse [wired] (Covers the inspected automatic recall branch, not parallel session injection or every deployed callback.); VikingBot automatic case-to-experience and user-memory selection: partial — inferred-embedding [wired], identifier [wired], coarse [wired] (The explicit exact-case lookup and complete profile/skill loading are not fully audited; these alternatives remain unresolved within the unit.); Automatic extraction hydration of retained tool output: known — identifier [wired] (Reference/truncation flags gate hydration; the payload identity and offset/limit actually select supplied material.); Requested search/find/session returns and explicit tool-result reload: inapplicable (These bounded operations are pull; their URI/anchor, lexical or relevance selectors do not establish a push signal merely by selecting a requested return.); Untraced parallel session injection, bot local-file, skill and other host/training consumers: uninspected (The reports leave these delivery/payload paths incomplete; one known pull route cannot make unresolved alternative push selection inapplicable.)	partial coverage (Qualifying wired trace-fed writes exist; no bounded absence records support no, and alternative writes prevent a complete assessment.); Automatic archive overview/checkpoint production followed by session/search consumption: known — yes [wired] (Includes the recorded cumulative and legacy summary consumer branches, not a guarantee of faithful continuity.); Session-derived ordinary memory/case files followed by retrieval or later extraction: known — yes [wired] (Covers ordinary session producers; initial content and training-case fast path remain unresolved below.); Trace-to-trajectory/experience and conditional session-skill maintenance: known — yes [wired] (This write-route classification is independent of the reports' uninspected improved-capacity, reflection or theory-builder properties.); Raw trace/offload/index/bookkeeping routes as independent learning mechanisms: not-determinable (These parts participate in qualifying chains, but the records do not completely establish distinct behavior-shaping learning writes or a bounded absence for every offload/selector alternative. No system-wide or local no is inferred from transport terminology.); Resource-derived producers, initial/manual memory, training-case fast path, bot local writers and alternate batch/snapshot variants: uninspected (Missing producer, persistent-updater or later-consumer facts prevent complete classification; dry-run policy snapshots are not silently treated as persisted learning.)	partial coverage (All seven authority values have local warrants; unresolved consumers prevent complete authority inventory and no finding proves compliance or benefit.); Memory and derived session text at later retrieval/planning and host-context consumers: known — knowledge [wired] (Wired delivery is the authority path; actual useful interpretation remains unobserved.); Experience bodies at VikingBot reminder and schema-declared future instruction consumer: known — instruction [wired], instruction [claimed] (Known channel-local classifications preserve the claimed/wired distinction, not a claim of universal host adoption.); Retained prior memory, trajectory diagnostics and experience/skill policy at update consumers: known — learning [wired] (Update consumption is separate from output knowledge/instruction delivery and from the unresolved general learning property.); Index vectors, stored abstracts and case relevance at scored retrieval: known — ranking [wired] (Filter-only find is a separate unranked branch, not evidence against the scored branch.); Terminal/archive/checkpoint controls, case links, tool refs and OpenCode association/retry state: known — routing [wired] (Names actual control consumers; unconsumed provenance labels are not granted force.); Recall ledger URI/turn/detail state at server gather: known — enforcement [wired] (Bare-URI, expired and malformed records have explicit exceptions; persistence failures weaken coverage rather than changing this local force.); Retained existing-file/link state at patch and replacement-delete checks: known — validation [wired], enforcement [wired] (Mechanical checks and vetoes are distinct from epistemic acceptance.); Full skill invocation, resource derivatives, bot local files and alternate host/batch consumers: uninspected (Readable skill text or a file method alone does not establish every actual downstream force; opaque payload effects and alternate consumers remain unresolved.)	AAS-2026-10-14-openviking-7b5b57541af5-01
OpenWiki	2	Use-derived repository/personal wiki pages; Claim statements, resource/version/envelope and projected metadata; resumable PageJob guidance/queue and coverage manifest; connector raw evidence/state and LangSmith-derived records/documents; persistent chat checkpoints and home-filesystem history offload; their maintenance, requested retrieval, visualization, Git/cache transfer and benchmark verdict/carry consumers. Includes unresolved dependency payload/selection and source-specific connector/evaluator retention alternatives. Excludes static shipped/installed prompts, skills and setup guidance as non-use-derived context; current-only nonchat state, transient search index/snippets, provider weights and external host-owned memory are not retained OpenWiki memory classifications.	partial coverage (Files and chat SQLite are established independently of opaque payload semantics. Optional evaluator output/cache persistence and source-specific connector alternatives lack complete substrate coverage.); Wiki Markdown, Claim parts/envelopes, page manifest, resumable queue and assignment guidance: known — files [wired] (Local Markdown and JSON files persist across later invocations; ClaimSession also retains process-local working copies. Filesystem is the operative durable substrate.); Process-local ClaimSession: known — in-memory [wired] (Session working copies of retained sidecar Claims participate in reconciliation before persistence.); Repository publication of accumulated wiki and manifest: known — repo [wired] (Scheduled CI proposes retained wiki/manifest changes in Git; actual merging is unobserved.); Raw connector files/state and compact LangSmith records: known — files [wired] (Private local JSON retains fetched payloads, recent-run metadata and compact traces.); Persistent chat checkpoint database: known — sqlite [wired] (Chat selects disk-backed openwiki.sqlite; nonchat :memory: state alone is not durable read-back memory.); Home-directory conversation-history offload: known — files [wired] (The wired filesystem mount is separate from SQLite; payload opacity does not obscure its substrate.); LEDGER verdict/carry working state: known — in-memory [wired] (Result and carry objects feed later benchmark evaluation in the continuing runner.); Optional LEDGER output persistence and DeepSWE cache backing: not-determinable (Records establish captures, archives and transfers but not a complete classification of their optional backing stores.); Connector-specific retained alternatives: uninspected (Individual cursor, expiry and replacement implementations were not inspected; shared file I/O does not enumerate every operative source-specific retained part.)	partial coverage (Native/model/code admission is automatic even when an operator starts a workflow. Host control, Git merge approval and independent initial entries remain unresolved; no generic caller identity establishes manual admission.); Native pages, Claims and persisted planner jobs: known — automatic [wired] (Models propose and core software admits without a per-content human decision.); Evidence/envelope/manifest and projected provenance/events: known — automatic [wired] (Deterministic code resolves, validates and persists retained metadata.); Personal model writes and deterministic connector acquisition: known — automatic [wired] (The model/filesystem and connector software admit content; initiating ingestion is not per-item approval.); Trace-fed documentation and compact trace persistence: known — automatic [wired] (Code selects/imports traces and model/core admission writes local documents without per-finding human approval.); Translation replacement and retry metadata: known — automatic [wired] (Code admits checked model output or retry markers, with no per-page human approval.); Chat state persistence: known — automatic [wired] (Graph/dependency state persistence is software admission without a human decision for each retained item. Exact restored payload is independently uninspected.); Intended middleware history offload admission: partial — automatic [afforded] (The intended middleware producer has direct backend write access without a per-offload human decision. The contract affords automatic admission; dependency production wiring and content policy are not inspected.); Benchmark capture, verdict/cache/carry and DeepSWE artifact replacement: known — automatic [wired] (Shipped software selects captures, judgment reuse and workspace/cache replacement.); Host-written content and external edits: not-determinable (Host identity and editable files do not establish whether a human supplies, edits or approves each admitted item.); PR merge into future Git baseline: not-determinable (CI creates a proposal automatically, but deployed merge approver/policy is not inspected; manual merge cannot be inferred.); Initial independent sources/foreign verification and source-specific state: not-determinable (Preserved entries and shared I/O do not establish their original content-control mechanisms.)	partial coverage (Requested retrieval and tool reads coexist with automatic assignment/context and evaluation carry. Dependency payload selection and actual history reload are unresolved rather than converted into pull-only inapplicability.); Requested host wiki search/section reads and personal wiki/raw tool reads: known — pull [wired] (Foreground consumers request retained material; ranking and returning the request do not add an unsolicited push.); Requested full page Claim inspection: known — pull [wired] (Worker/host explicitly requests complete compact current Claims.); Automatic persisted PageJob guidance and issue Claims in assignments: known — push [wired] (At worker construction/resume, core selects retained assignment and attention Claims without a content read request by that worker.); Automatic lifecycle reload of queue, manifest and sidecars: known — push [wired] (Begin/resume loads retained control state to drive the later lifecycle invocation.); Chat continuation thread request: known — pull [wired] (Caller supplies a thread for graph continuation; satisfying that keyed continuation request is pull at the request boundary.); Checkpoint model-context delivery operations: uninspected (The later graph restore is wired, but dependency selection of opaque retained turns/summaries into model context is uninspected; additional push operations cannot be excluded.); Visualizer requested graph projection: known — pull [wired] (A later visualization request consumes the retained wiki; browser details do not change the inspected requested build.); LEDGER automatic carry and artifact supply to next evaluation: known — push [wired] (Runner supplies accumulated previous artifact/watch state to a subsequent evaluator call automatically.); DeepSWE artifact transfer into downstream task workspace: known — push [wired] (Preparation selects generated/cached retained wiki and copies it before a separate task host run; later host section requests remain pull.); Native history offload reload and cross-worker selection: not-determinable (Backend read affordance is known, but no actual automatic delivery chain or exercised requesting consumer is established. It cannot resolve a direction unit by itself.); Source-specific connector and remaining evaluator consumer routes: uninspected (Uninspected alternatives prevent complete read-back direction coverage.)	partial coverage (Signals classify automatic push operations, not lexical ranking of requested search results. Assignment identity and issue filtering are distinct from coarse availability and opaque continuation/history selection.); Persisted guidance selected for assigned job: known — identifier [wired] (Trigger is worker/host assignment or resume; pending job/page identity selects the retained guidance actually delivered.); Issue Claims selected into assignment: known — identifier [wired], coarse [wired] (Trigger is next page assignment; exact owning page/job selects retained Claims, then stale/unresolved issue status selects the attention subset.); Lifecycle control-state reload: known — identifier [wired] (Trigger is begin/resume; wiki/page paths select retained run, manifest and sidecar state for code consumers.); LEDGER previous artifact and obsolete watch-state carry: known — coarse [wired] (Trigger is advancing benchmark checkpoint; availability of continuing carry supplies prior artifact/watch state. No semantic similarity selector is established by the runner.); DeepSWE generated/cached artifact supply: partial — identifier [wired] (Trigger is task preparation; task base/package digest/model cache identity selects a retained wiki for copying. Compatible older-cache selection is recorded but complete key trust remains unassessed.); Opaque checkpoint delivery selector: uninspected (Thread identifies graph namespace, but that does not identify which opaque retained parts are automatically delivered to the model.); History automatic replay selector: not-determinable (Mount prefix is an access mapping, not an established selector for delivered history parts; production trigger/input/budget are unresolved.); Requested search, explicit section/Claim/raw reads and visualization builds: inapplicable (These local units fulfill requests rather than push; lexical ranking of a requested search does not establish inferred-lexical push.); Residual connector and evaluator push alternatives: uninspected (Source-specific routes and full evaluator selection were not inventoried in accepted records.)	partial coverage (The external operational-trace to retained documentation to later context chain supports a wired positive without improved-capacity evidence. Opaque state/offload and personal-source alternatives prevent a complete yes/no inventory; no system-wide negative is asserted.); LangSmith operational traces to durable runtime documentation and later agents: known — yes [wired] (Automatic trace-fed authoring produces retained behavior-shaping documents available to later requested wiki context. Architectural qualification does not require observed improvement.); Checkpoint and native history raw-to-derived-to-later-model alternatives: uninspected (Recorded graph persistence and intended summarization do not resolve an automatic durable summary production/consumption chain; raw storage alone is insufficient.); Personal connector trace-like source updates: uninspected (Individual source contents/transforms were not inspected, preventing classification of session/event-fed derived documentation routes.); LEDGER automatic assessment retention and later evaluator reuse: not-determinable (Verdicts/cache/carry are consequential evaluation state, but accepted records do not establish a qualifying trace-fed behavior-shaping update; no score-to-product revision gate is wired in the inspected runner.); Nontrace repository authoring, metadata, queue, translation and transfer operations: inapplicable (These bounded ordinary operations are not a qualifying trace-fed update merely because code executes them. The separately scoped trace-documentation branch remains included.)	partial coverage (Forces are attached to actual consumers, not JSON/file labels. Supported knowledge, instructions, routing, validation, enforcement, ranking and trace-update learning coexist with unresolved dependency payload consumers.); Wiki prose at later authors, retrieval hosts, personal chat and human visualization: known — knowledge [wired] (Retained prose supplies evidence/context/reference, not standing host instructions under shipped guidance.); Retained wiki text/frontmatter/source hints at retrieval ranker: known — ranking [wired] (Text is scored lexically and source/path metadata affects relative priority; no learned policy is required.); Deprecated status at search eligibility: known — enforcement [wired] (Search status excludes deprecated pages from delivery; explicit section reads do not repeat the gate.); Claim statements and associated resources at maintenance model/host: known — knowledge [wired] (Statements and resource identities supply maintenance evidence; freshness issues do not prove truth.); Evidence versions/Claim membership/envelope at reconciliation and durability checks: known — validation [wired], enforcement [wired] (Retained IDs, versions and state provide admissibility criteria and veto invalid transitions.); Queue, manifest and freshness issue state at lifecycle selection: known — routing [wired] (Retained progress/checkpoints and debt select phase, next work, source windows and required page work.); Persisted PageJob guidance at native worker or resumed host: known — instruction [wired] (Purpose/instructions supply directions for assigned work across stages/resume.); Connector run/cursor access metadata at raw selection: partial — routing [wired] (Known run/access fields select retained raw files for reads; detailed source-specific cursor authority is unresolved.); Raw source text at personal curation and compact traces at documentation update: known — knowledge [wired] (The authoring model receives source content/execution evidence as context.); Retained operational traces at automatic documentation update: known — learning [wired] (Retained experience feeds durable behavior-shaping documents; this is learning force at the update consumer, not measured improved capacity.); Checkpoint thread/namespace keys at continuation selector: known — routing [wired] (Retained keyed state selects the graph continuation namespace. Opaque content force at the next model is separate.); LEDGER verdict/carry at benchmark reporting and subsequent evaluation: known — knowledge [wired], routing [wired] (Verdicts supply assessment context; obsolete targets and previous artifacts select later evaluation scope. No product-edit gate is established.); Opaque checkpoint content and actual history payload at later model: uninspected (The exact delivered part/effect and automatic replay are unresolved; a known key or backend read capability does not resolve knowledge versus instruction force.); Source-specific connectors and optional evaluator consumers: uninspected (Uninspected source-specific state and output consumers prevent a complete authority inventory.)	AAS-2026-10-14-openwiki-8f502f680f75-01
origin	2	Retained captures and imported source content; episodes, fact-child/vector and summary access structures; page prose/projections, claims, support/review state and proposals; typed Space Briefs, handoff artifacts, queued writes, usage/maintenance state, cached diagnostics and repair artifacts; mutable prompt overrides and their later consumers. Includes unresolved graph/community/vocabulary, full import/distillation, repair/refinement, client projection and evaluation-persistence alternatives recorded in the inventory. Excludes shipped static prompts/skills as accumulated memory, current-call evaluator compression without later reload, host-private transcripts/compaction, provider-internal state and inference weights not shown to be accumulated through Origin use. Provider mutability remains unresolved externally, not classified as absent learning.	partial coverage (The recorded inventory covers database objects, retrieval indexes, file projections/configuration, repair artifacts and cached diagnostics. The lint-cache carrier, graph/community/vocabulary objects, queue encodings and complete client/evaluator persistence remain unresolved; these prevent a complete substrate inventory.); Memory, episode, fact-child, summary-node and Brief relational carriers: known — sqlite [wired], rdbms [wired] (These named objects persist in local libSQL; classification concerns their recorded carriers, not every table in the product.); Dense retrieval indexes for memory, episode, fact-child and summary text: known — vector [wired] (Dense vectors are retained retrieval structures in the same database, not a claim of an independent vector service.); Claim, support and page-review state: known — rdbms [wired] (The records establish persisted claim/revision and SQL truth-state carriers, separately from readable page prose.); Markdown page and Brief projections, session-log artifacts and prompt override files: known — files [wired], files [afforded] (Files are operative but not interchangeable authorities: page edits reload, Brief receipts do not. Session-log creation is instructed, not observed.); Cached lint reports: not-determinable (Exact-scope MCP caching and later repair planning are recorded, but the accepted records do not identify the cache carrier. Caching alone cannot distinguish volatile memory from persisted storage, so no controlled substrate value is established for this unit.); Repair store, offline outbox, graph/community/vocabulary state and remaining page/client/evaluation artifacts: not-determinable (A private repair store and queued writes are established, but their complete carriers are not specified in the accepted records. The memory inventory leaves graph payloads and other persistence paths uninspected. No graph, files or service-object value is inferred from a feature or store name alone.)	partial coverage (Known software admissions and instructed operator choices coexist. Generic caller identity, external editing and uninspected producers prevent a universal manual/automatic classification.); Capture gate, inspected import branches, index co-writes, enrichment, recaps, rollup and page synthesis admission: known — automatic [wired] (The scoped admission steps decide by software/model rules; starting a workflow or authoring its input is not item-level approval.); Host-authored handoff delta admission: known — automatic [afforded], automatic [wired] (The user starts handoff, but instructions explicitly omit approval for ordinary Brief changes/captures. Daemon typed commit is wired separately from instructed host authorship.); Support publication/withdrawal, access logging and automatic withdrawal maintenance: known — automatic [wired] (This includes the automatic maintenance branch, not explicit forget whose caller identity is unresolved.); Human page edits and instructed revision/refinement decisions: partial — manual [afforded], manual [claimed] (Explicit user-provided edits/decisions establish the manual alternative, but handlers and filesystem scans do not authenticate a human; other callers remain unresolved. Forced rebuild confirmation does not make its generated candidate manually approved.); Version-bound page-review mark and operator truth cutover: partial — manual [wired], manual [afforded] (Cutover explicitly requires operator confirmation of the plan. Page-review human provenance is not audited end-to-end despite its bound presence protocol.); Prompt-file replacement, direct forget/review/repair callers and remaining phase/client/import producers: not-determinable (A file writer, local caller or approval string does not establish an explicit human content decision. Full repair approval provenance, source preparation and graph/refinement admissions remain unresolved; software physical I/O alone does not resolve control.)	partial coverage (Requested foreground retrieval and automatic background selection are different operations. The SessionStart hook is not evidence of knowledge injection. Remaining graph, import, projection and repair consumers prevent complete direction coverage.); Requested recall, page/source/revision access and Space Brief reload: known — pull [wired] (Query ranking, supplemental pages, prelude and delivery fulfill a consumer request; delivery alone is not an extra push.); Automatic enrichment, burst/rollup model inputs and source-change page refresh: known — push [wired] (Daemon selectors supply retained material to a later inference/task invocation without a query from that consumer. This is background push, not unsolicited foreground chat injection.); Claim/support and cached diagnostic/repair-artifact reuse: partial — push [wired], pull [wired] (The truth worker supplies retained claim/evidence pairs automatically. Explicit planning/apply/verification calls consume requested retained artifacts; every internal selection/delivery leg is not recorded.); Prompt override reload and remaining graph/import/client/evaluation routes: not-determinable (Override loader consumer and task guidance are known, but complete reload cadence and request-versus-independent-selection chains are not. Other included routes lack full later-consumer tracing; known pull cannot eliminate these alternatives.)	partial coverage (Only automatic push selectors are classified. Query embedding/reranking on requested recall is not a push signal. Coarse workload selection and identity matching remain separate findings; unspecified ranking algorithms are not inferred.); Automatic durable-candidate enrichment and recent-burst summary selection: known — coarse [wired] (Trigger is an automatic stage; selector inputs are stage eligibility/attempt generation or lookback, temporal grouping, minimum size and coverage. Selected parts are durable candidate content or burst contents.); Automatic source-change page refresh source matching: known — identifier [wired] (The automatic refresh trigger selects a page; current page-source links or its stored source list match retained source identities and determine delivered snippets. The foreground rebuild record corroborates the shared source lookup, not an additional push receipt.); Automatic rollup bucket/root input and truth-worker candidate selection: partial — coarse [wired] (Rollup eligibility/minimum membership/provider availability and truth-worker leases/budgets are recorded coarse filters. Complete graph-to-bucket derivation and the ranked linked-evidence candidate selector are not audited; no inferred-embedding, inferred-lexical or inferred-judgment value is guessed.); Remaining automatic graph/import/repair selections and override delivery: not-determinable (The records do not establish every trigger, selector input and delivered retained part for these alternatives. Prompt filenames and manifest IDs alone are not identifier-based push; an actual selection/delivery chain is required.)	partial coverage (Qualifying automatic trace-fed retained-artifact routes exist, without establishing improved capacity or parameter training. Mixed input origins and uninspected update alternatives prevent a complete classification; no system-wide no is asserted.); Session handoff to durable Brief/captures to later resume/recall: known — yes [afforded] (The shipped host procedure interprets session evidence and writes durable summary/items without per-item approval; later Brief reads reload them. Host execution is instructed, not observed.); Captured conversational input to durable enrichment and later retrieval: known — yes [wired] (This scope is the explicitly recorded conversational-input branch, not every candidate or imported file. Persisted derived fields/content supply later retrieval/context even without weight updates.); Recap and summary-rollup inputs across the mixed retained corpus: partial — yes [wired] (The corpus can contain captured conversational experience, so that branch continues a qualifying automatic chain; other inputs may be authored/imported/static or previously derived. Exact experiential ancestry of each selected item is unresolved.); Imports through full downstream distillation, graph/refinement/repair updates, access-feedback and prompt-file authors: not-determinable (Retaining conversation exports or access logs alone is not trace learning. Full downstream selectors, raw-to-derived links and later consumers remain incomplete; production count flushing is unresolved. Loader persistence does not establish trace-fed override authorship.); Read-time evaluator compression output: inapplicable (This inspected branch retains no durable checkpoint for a later invocation; its substituted current-case answer context is outside accumulated memory. It neither establishes trace-learning yes nor a system-wide no.)	partial coverage (Authority is assigned at actual consumers, with learning at update consumers separate from downstream knowledge. Missing graph, refinement-writer, repair-payload and projection detail prevents a complete authority inventory.); Memory/page/summary/episode prose delivered to later recall and synthesis consumers: known — knowledge [wired] (Knowledge means supplied context, not warranted truth or observed answer change. Standard MCP omits source_text and separate summary.); Memory, fact-child, episode and summary retrieval features: partial — ranking [wired] (Retained text/vectors affect relative selection in inspected retrieval paths; full ranking and every field contribution remain uninspected.); Brief summary/items at the instructed resume consumer: known — knowledge [afforded], instruction [afforded] (The external agent is instructed to treat the typed Brief as current-work authority; compliance is not observed and the Markdown receipt is not authority.); Retained prompt overrides at inference consumers: known — instruction [wired] (Loader and task request wiring support model guidance, not semantic enforcement or feedback-based improvement.); Retained version, pending/supersession/eviction and ownership fields at admission and retrieval: known — validation [wired], enforcement [wired] (These are controls at the named gates, not a universal permissions envelope.); Scheduler cursors, enrichment receipts and source links at subsequent automatic work: known — routing [wired] (State selects subsequent work or source input windows; it is not automatically knowledge.); Page truth/support and human-review state at exposure and reconciliation: known — enforcement [wired], validation [wired] (Enforcement is conditional on generation/grant and has documented exceptions; absence of activation evidence is not absence of wiring.); Repair manifests, receipts and cached lint reports at apply/planning/verification consumers: known — knowledge [wired], validation [wired], enforcement [wired] (The recorded controls authorize/check bounded repairs, not all content as true.); Session evidence at the automatic durable-artifact update consumer: partial — learning [afforded], learning [wired] (This learning authority is the trace-fed artifact-update force, not improved capacity. Original provenance is unresolved for mixed recap/rollup inputs.); Remaining graph/community/vocabulary consumers, refinement writers and complete projections: uninspected (The recorded coverage limits leave these consumer/effect branches unresolved, so the positive authority set is not complete.)	AAS-2026-10-15-origin-805d03bc599f-01
OS-Copilot	2	Retained tool code, descriptions, vector/name access state, live reloaded manager copies, course lessons and their later curriculum/task routes, and GAIA result rows consumed for accounting/resume. Includes unresolved initial/direct-caller provenance and possible external-provider/page-service retained alternatives. Excludes shipped static prompts and inference weights as established accumulated memory, ordinary graph/repair state and example histories, raw display-only UI logs/screenshots, and arbitrary host effects without a recorded later consumer; no exclusion resolves opaque provider or page-cache alternatives.	partial coverage (Known payload, index, curriculum and benchmark stores coexist with unresolved service/provider retention. No repository or model-weight update substrate is inferred from source hosting or inference.); Retained tool code and description payloads, including live manager copies: known — files [wired], in-memory [wired] (JSON and mirrored source/text files persist content; getters consume the reloaded dictionary.); Tool description access index: known — vector [wired] (The persistent Chroma collection is a vector substrate; dependency-internal SQL or file layout is not separately inferred.); Course lessons and their live curriculum copy: known — files [wired], in-memory [wired] (Course JSON reload and in-process dictionary updates are both explicit.); GAIA retained comparison rows used for accounting/resume: known — files [wired] (The script appends JSONL and reloads rows on later launches.); Possible external-provider retained state and page-service cross-request state: uninspected (Provider retention/update internals and Chroma page collection lifecycle were not inspected. Inference alone does not classify accumulated provider memory; request-local page processing does not establish a later-memory route.)	partial coverage (Known software admission and an afforded human content-supply mechanism coexist with unresolved generic and initialization control.); Automatic candidate tool admission, extracted descriptions and index writes: known — automatic [wired] (Starting the task is not approval of each retained candidate.); Course design, dictionary update and JSON merge: known — automatic [wired] (Parsed model output is software-admitted, with no lesson-success approval gate.); Documented human tool-maintenance use: known — manual [afforded] (This finding covers only a human following the tutorial and supplying/editing the selected file, not every API caller.); Generic manager writes and initial tool/course files: not-determinable (Caller identity and loader behavior do not determine per-content human control. Do not label every maintenance API write manual.); GAIA result persistence: known — automatic [wired] (The driver appends result rows without per-row operator approval.); Possible external-provider retained state and page-service cross-request state: uninspected (Provider retention/update internals and Chroma page collection lifecycle were not inspected. Inference alone does not classify accumulated provider memory; request-local page processing does not establish a later-memory route.)	partial coverage (Both pull and push are supported, at distinct operations. Page HTTP response delivery alone does not establish a later-memory route; provider alternatives remain uninspected.); Tool retrieval request, similarity selection and payload return: known — pull [wired] (Task/replan/subtask transitions issue retrieval requests. Similarity fulfills those requests; returned names are internal joins, not a second unsolicited push.); Course filename initialization load: known — pull [wired] (Software/package selects the requested local course file.); Prior-course insertion into subsequent course-design prompts: known — push [wired] (At curriculum design, the driver automatically supplies up to the last fifty dictionary entries without a model request for those entries.); GAIA result-file reload for counters and resume: known — pull [wired] (The benchmark caller requests previous result rows through get_numbers; subsequent skip logic is a control effect, not additional content delivery.); Possible external-provider retained state and page-service cross-request state: uninspected (Provider retention/update internals and Chroma page collection lifecycle were not inspected. Inference alone does not classify accumulated provider memory; request-local page processing does not establish a later-memory route.)	partial coverage (The known push has a coarse selector. The unresolved service/provider alternatives prevent treating all other included memory as assessed pull-only.); Course prior-context push selector: known — coarse [wired] (Trigger: next course-design call. Input: insertion-ordered course dictionary and length. Selected retained part: the last fifty lesson entries, or all entries if shorter. There is no semantic match or token cap.); Requested tool retrieval, course initialization and GAIA reload: inapplicable (These operations are pull. Embedding matching and exact filenames/IDs on requested returns do not establish push signals.); Possible external-provider retained state and page-service cross-request state: uninspected (Provider retention/update internals and Chroma page collection lifecycle were not inspected. Inference alone does not classify accumulated provider memory; request-local page processing does not establish a later-memory route.)	partial coverage (A wired yes is not observed learning, improved capacity or causal self-improvement. GAIA accounting does not establish a qualifying positive, but lacks a canonical bounded absence warrant for no; unresolved alternatives remain explicit.); Feedback-gated reusable tool artifacts and later generation context: known — yes [wired] (Automatic execution-feedback-responsive admission yields durable code/descriptions and a wired later consumer. Original-code/repaired-score divergence limits quality and candidate identity, not existence of the route.); Course generation from prior lesson text, configuration and optional demo-derived content: not-determinable (The saved material is proposed curriculum, not a compacted outcome trace. The optional agent-derived demo payload and initial/prior content provenance are not characterized enough to establish every trace-fed alternative or a complete local negative.); Generic maintenance and initial tool content: not-determinable (Generic callers leave both admission control and original trace-fed derivation unresolved; the afforded human path does not settle other callers.); GAIA result rows used only for accounting/resume: not-determinable (The accepted record establishes raw-result accounting and resume, not a qualifying learned artifact. It declares no bounded absence record, so a controlled local no cannot be established.); Possible external-provider retained state and page-service cross-request state: uninspected (Provider retention/update internals and Chroma page collection lifecycle were not inspected. Inference alone does not classify accumulated provider memory; request-local page processing does not establish a later-memory route.)	partial coverage (Authority is attached to actual retained parts and consumers. Transient judge scores enforce automatic admission but are not retained controls. No learning authority is inferred for unretained feedback or from trace-learning membership alone.); Tool code and readable descriptions at later planner/generator consumers: known — knowledge [wired] (Delivery is contextual, not an enforced instruction or automatic stored-code invocation.); Descriptions and dense vectors at tool similarity selector: known — ranking [wired] (Text and its embedding affect relative retrieval priority under fixed similarity code.); Tool name metadata at retained-payload dereference: known — routing [wired] (Selected identity metadata chooses which JSON payload enters the context channel.); Prior lesson values at the next course-design call: known — knowledge [wired] (The prior-course text is advisory design context, not verified completion evidence.); Retained/generated lesson values at agent task dispatch: known — instruction [wired] (Lessons become operative requests to the task agent rather than background reference alone.); GAIA saved answer/reference pairs at equality accounting: known — validation [wired] (The reference is a bounded validation criterion for exact string agreement, not tool admission or semantic truth.); GAIA saved final index at resume selection: known — routing [wired], enforcement [wired] (The checkpoint selects the subsequent input window; the code comparison enforces the skip.); Possible external-provider retained state and page-service cross-request state: uninspected (Provider retention/update internals and Chroma page collection lifecycle were not inspected. Inference alone does not classify accumulated provider memory; request-local page processing does not establish a later-memory route.)	AAS-2026-10-16-os-copilot-7cef03148c53-01
PageIndex	2	Retained local pages, tree parts, metadata and manifest; alternative builder JSON; external heading cache; child summaries consumed by later parent calls; caller-held appendable transcripts; notebook outline and page-image reuse; and remote document memory exposed by cloud clients. Includes admission, maintenance and later-consumer paths. Excludes static shipped instructions, fixed optimizer formulae, inference weights without a recorded memory update, ordinary transient runner state and diagnostic logs without a later consumer. Remote internals, cache initialization and external history control remain included uncertainties, not inferred local equivalents.	partial coverage (Local file/value inventory is supported; remote and external persistence prevent a complete system substrate classification.); Local pages, tree, metadata/manifest, heading-cache file and alternate builder JSON: known — files [wired] (These records identify each included file surface and its persistence boundary.); Summary values between child and parent invocations; caller-held protocol transcript values: known — in-memory [wired] (Known for the supplied interfaces and within-run summary path, not for a host's additional persistence.); Vision notebook retained/fetched tree and PDF page images: known — files [wired], in-memory [wired] (Source wiring covers notebook files/variables, not cross-session restoration.); Remote document objects and any caller-added transcript persistence: not-determinable (Remote IDs do not identify a physical substrate; hosts own any persistence beyond returned transcript values.)	partial coverage (Automatic downstream admission is positive evidence; unknown initialization and host/server control prevent a complete agency union.); Managed local derived-document admission, summary/refinement admission and alternate builder JSON production: known — automatic [wired] (Caller-selected options do not make subsequent generated-content admission manual.); Local deletion and manifest repair inside the exposed management/store mechanisms: known — automatic [wired] (Classifies internal admission; the external actor initiating a deletion remains caller/model dependent.); PDF and metadata initial supply, heading-cache creation, external transcript retention and cloud admission: not-determinable (Generic caller identity does not establish an explicit human content decision; cache and remote admission controls are not recorded.)	partial coverage (Both pull and push exist locally at distinct operations. Remote and caller-owned routes prevent complete direction coverage.); Local document tools/getters connected to later QA or framework consumers: known — pull [wired] (Known request-led delivery covers the shipped local callbacks; actual framework use is not observed.); Settled child summaries automatically supplied to later parent-summary calls: known — push [wired] (Within-run later invocation qualifies; ordinary unretained current-call state does not.); Vision notebook request for tree followed by automatic outline/image input assembly: known — pull [wired], push [wired] (Distinct request, selection and delivery operations are not conflated with managed tool chat.); Cloud client getters and tool returns: partial — pull [wired] (Client request wiring does not enumerate server-side automatic selection or managed-chat delivery.); Caller-selected protocol-history reuse and unspecified alternate JSON consumers: not-determinable (Appendable values afford reuse, but external selection/request ownership is unresolved; file persistence alone establishes no later consumer.)	partial coverage (Inspected automatic frame construction supports coarse and targeted signals; opaque alternatives remain unresolved.); Child-summary supply to parent and full-outline supply to notebook selector: known — coarse [wired] (Names the automatic stage trigger, structural input and retained material entering the next frame.); Vision notebook selection and delivery of retained page images: known — inferred-judgment [wired], identifier [wired] (Judgment and subsequent ID/range mapping are separate selector mechanisms in one automatic delivery chain.); Local request-led discovery/structure/page responses: inapplicable (This completely described local route fulfills requests; its page IDs, ordering and response budgets are not push signals.); Remote automatic delivery and caller history selection: not-determinable (Remote policy and external history selection are unknown; known local pull cannot make these alternatives inapplicable.)	not-determinable coverage (No qualifying automatic trace-fed update is established by the accepted records. No absence record closes every included write alternative, so neither a system-wide no nor yes is asserted.); Local document-derived trees, summaries and metadata: inapplicable (The recorded transformation inputs are document content, not activity/experience traces; these particular writes are not trace-fed learning routes.); Raw protocol transcripts, cache provenance, remote updates and external host reuse: not-determinable (Raw trace exposure does not establish an automatic durable behavior-shaping write. Unknown cache origin, external retention and server updates prevent either a positive or a complete bounded negative.)	partial coverage (Known evidence, priority, addressing and availability paths do not settle opaque alternatives. Static managed instructions and cost criteria are outside this memory scope.); Page text and outline/document descriptions supplied to answering or selection models: known — knowledge [wired], ranking [wired] (Knowledge and ranking name separate consumer effects of retained descriptive content, not policy authority.); Tree spans/node IDs and heading-cache page/kind data at deterministic selection consumers: known — routing [wired] (Addressing effects are separate from the semantic relevance of the chosen material.); Document marker and valid metadata at local store/getter consumers: known — enforcement [wired] (The stored marker gates availability; fixed schema code is excluded static machinery rather than retained validation memory.); Child summaries at parent-summary consumer and raw histories at later chat consumer: known — knowledge [wired] (No learning authority follows from content reaching an inference call.); Cloud retained parts, arbitrary host histories and caller metadata beyond documented selectors: not-determinable (Remote internals and additional host effects are unknown. Remote guidance is not shown to be an artifact accumulated through use.)	AAS-2026-10-15-pageindex-50056ef72767-01
pal	2	Retained raw/wiki/user files and access metadata, optional context Git replication, PostgreSQL user facts, routing and learning text/vector stores, personal memory and session history; includes model, direct API, loader, synchronization, evaluator-cleanup and opaque registered-service alternatives. Excludes unchanged shipped instructions/templates, current replies and transient evaluation maps, external source services unless saved, and provider model weights not shown to accumulate Pal experience.	known coverage (The accepted inventory locates all included retained objects, including opaque framework state and the two registered service handles. Known is configuration-level substrate coverage, not observed persistence or complete payload knowledge.); Raw sources, wiki articles/index/state/lint, manifest and edited or generated user files: known — files [wired] (Covers substantive file content and its separate navigation/status objects, not merely the article body.); Optional Git copies and history of retained context files: known — repo [wired] (Git supplements filesystem storage; it does not replicate database memories.); Routing and operational-learning contents and dense retrieval indexes: known — rdbms [wired], vector [wired] (Opaque entry serialization does not obscure these configured storage substrates. Registered services expose these same handles, not an evidenced additional store.); Structured user facts, framework personal memory and persisted sessions: known — rdbms [wired] (Payload schemas and exact table contents remain unknown; substrate classification does not require guessing them.)	partial coverage (Automatic mechanisms and an afforded explicit human-edit path coexist with unresolved generic caller, initialization, remote and service admission.); Researcher ingest, Compiler/Linter files and metadata, Navigator files/SQL and specialist routing writes: known — automatic [wired] (Classifies these model-proposed admission mechanisms, not every direct caller of the same underlying API.); Loader metadata admission and evaluator raw/wiki cleanup: known — automatic [wired] (A caller's recreate or suite-launch decision is not approval of each retained item.); Configured learning writes, personal-memory updates and session persistence: partial — automatic [wired] (Exact dependency rejection/replacement control and registered-service alternatives remain unknown.); Explicit human edits to retained user guidance/content: known — manual [afforded] (Static shipped guidance is excluded; no actual human editing episode is observed.); Generic direct-ingest submissions, initial retained contents, remote Git authorship and service API admission: not-determinable (Caller/remote identity and dependency operation contracts do not establish who decided the exact content. Fixed physical writes or a Git pull cannot settle its upstream admission agency.)	partial coverage (Both foreground pull and configured automatic push exist. Dependency-owned memory and service alternatives prevent complete route coverage, not the positive direction findings.); Requested wiki/raw/user-file/SQL retrieval and routing or learning search: known — pull [wired] (Includes maintenance article/raw reads; file identifiers implement requested fetches rather than autonomous delivery.); Configured learning addition, recent history supply and explicit history/session search: partial — push [wired], pull [wired] (Personal-memory delivery, exact automatic selection and possible checkpoint alternatives remain dependency-internal.); Lint report later repair use and registered knowledge-service consumers: partial — pull [afforded] (Registered service handles alone do not establish another read-back route; their concrete consumers and operations remain unresolved.)	partial coverage (The history window supports a coarse component. Opaque automatic learning/personal-memory selection and service alternatives prevent complete signal classification; requested identifier reads do not add identifier push.); Recent-history push to leader and Navigator: partial — coarse [wired] (Exact turn/tool filtering and scope keys are external. Run/session IDs do not establish a separately inspected identifier-matched push.); Automatic learning-context supply and personal-memory inclusion: not-determinable (Trigger is a later configured invocation; selected parts are intended learnings/preferences. Selector inputs, matching rule and budget are unknown. Hybrid store configuration alone cannot establish the signal of automatic delivery.); Requested file/SQL/knowledge/learning/history reads: inapplicable (These named operations fulfill consumer requests; push selection signal does not apply to their requested returns. This local inapplicability does not cover automatic alternatives.); Unresolved registered-service read-back alternatives: uninspected (No concrete service consumer or push selector is established by registration; cannot classify its signal or declare it inapplicable.)	partial coverage (Qualifying configured guidance-update chains support yes. No bounded-absence record supports no for other routes; unknown products, checkpoints and service writes remain included. Trace learning is not measured learning capacity or exercised self-improvement.); Interaction-fed operational learning writes and later guidance consumers: partial — yes [wired] (Exact extraction, stored entries and all alternative input/write paths are opaque; the positive is architectural, not observed uptake or benefit.); Agentic personal-memory updates and later preference guidance: partial — yes [wired] (Precise payload, extraction and delivery remain external. This does not classify raw history persistence itself as learning.); Generated/saved user products, ordinary source compilation and opaque session checkpoints: not-determinable (Saved summaries/reviews can survive, but records do not resolve every product's original input and later behavior-shaping route. Document compilation need not be trace-fed; raw history alone is insufficient. External continuation-summary/checkpoint production is uninspected, not absent.); Additional admissions and consumers through registered knowledge services: uninspected (Service operations and callers are dependency-owned; registration does not prove or exclude a qualifying trace-fed write.)	partial coverage (Consumer-specific knowledge, instruction, ranking, routing and learning effects are supported at distinct layers. Opaque payload/service effects remain included; no system-wide enforcement/validation negative is asserted.); Selected raw/wiki/SQL/saved content at answering or maintenance models: known — knowledge [wired] (Actual content supplies knowledge rather than a hard control; compliance and answer benefit are unobserved.); Wiki index, manifest eligibility, routing descriptions and dense features at retrieval/compile selectors: partial — routing [claimed], ranking [wired] (Exact dependency ranking, service consumers and automatic selectors are uninspected. Bookkeeping timestamps alone are not classified as enforcement or validation.); Retained operational guidance, personal preferences and edited voice/template files at later generation: partial — instruction [wired], instruction [claimed], ranking [claimed] (Payload contents, precedence enforcement and all dependency-owned effects remain unresolved.); Recorded interactions/corrections at learning and personal-memory update consumers: partial — learning [wired], knowledge [wired] (Exact trace contents, extraction and delivered payloads are not inspected; no improvement or additional veto authority follows.); Retained lint criticism at intended later repair readers: partial — knowledge [afforded] (No enforced article withdrawal or validation criterion is established by storing a diagnosis. Actual repair consumption remains unresolved.)	AAS-2026-10-17-pal-e313a0902289-01
Phantom	2	Retained material accumulated or changed through use: episode/fact/procedure payloads and dense/sparse access indexes; evolved and reflection-only files, working memory, append notes and dashboard user Markdown; evolution queue/poison/version/reason state; committed chat text, attachment payloads, translated events and durable timelines; host SDK-session selectors and opaque dependency-owned histories/checkpoints; changed dynamic capabilities and operator extensions/settings; durable scheduled task handoffs. Includes write/admission, maintenance and later model/software/client consumption routes. Excludes unmodified shipped roles/security/constitution/teaching prompts and baseline skills as accumulated memory, transient query coordination/staging/snapshots as standalone durable stores, arbitrary work products without a recorded later consumer, and provider model weights without an established in-scope update. External SDK retained state is included as opaque, not silently excluded.	partial coverage (The inventory covers vector memory and indexes, file memories and extensions, chat stores, learning controls, session selectors and scheduled handoffs. SDK-owned retained conversation/checkpoint storage is unresolved; external inference weights are not assumed to be accumulated memory.); Qdrant episode, semantic-fact and procedure payloads and their retrieval indexes: known — vector [wired] (Payloads and access vectors live in the vector store, not in model weights.); Evolved files, reflection-only files, working memory, agent notes and user-scope Markdown: known — files [wired] (The live EvolvedConfig copy is backed by retained filesystem text.); Evolution queue and poison state: known — sqlite [wired] (Staging JSONL is a transient delivery copy, not an additional durable memory substrate.); Evolution versions, gate logs, change summaries and metrics: known — files [wired] (Transient snapshots are not a retained Git repository of prior versions.); Committed web transcript, translated events and derived timelines: known — sqlite [wired] (Attachment payloads are split from descriptors and message rows.); Chat binary attachments: known — files [wired] (The record establishes attachment files, not a text rendering of their contents.); SDK session access rows: known — sqlite [wired] (These are selectors, not the selected SDK conversation.); Dynamic capability definitions and executable script files: known — sqlite [wired], in-memory [wired], files [wired] (Definitions are durable in SQLite and reloaded into a Map; script bodies can mutate separately.); Changed skills, subagent instructions, hooks, plugin settings and runtime YAML: known — files [wired] (Only changed or installed content is included, not shipped static defaults.); Scheduled job instruction and dispatch metadata: known — sqlite [wired] (A retained future-work handoff, distinct from learned vector procedures.); Dependency-owned persisted SDK conversations and compaction payloads: uninspected (The records distinguish SDK persistence from host rows but do not inspect its storage implementation. No host SQLite or file classification is transferred to the opaque payload.)	partial coverage (Software/model admission is automatic even when a user starts work or authors source messages. Authenticated caller identity does not establish a per-content human decision. Initial procedure/vector admission and SDK writes remain unresolved; no manual value is inferred from dashboard labels.); Channel episode/fact extraction and subsequent index/store updates: known — automatic [wired] (No per-fact approval is required.); Reflection queue admission, model edits and branch-limited software acceptance: known — automatic [wired] (User corrections are evidence rather than per-edit approval.); Foreground note admission and ordinary model file/capability edits: known — automatic [afforded], automatic [wired] (Foreground persistence is distinct from reflection admission.); Startup note seeding, host chat persistence/timeline construction, logs and session bookkeeping: known — automatic [wired] (Physical I/O is not the criterion: these are software-selected admissions.); Scheduled job creation by foreground model: known — automatic [wired] (An initiating request is not approval of the resulting task body.); Dashboard memory Markdown edits/vector deletion and external extension/dynamic-tool admission: not-determinable (The authenticated caller could be human or software. No record establishes the human decision required for manual classification on every such request.); Initial vector/procedure and evolved-file admission; optional historical session log: not-determinable (Subsequent known transformations do not resolve initial control or an unestablished writer.); Dependency-owned SDK conversation persistence and compaction admission: uninspected (The wrapper requests persistence, but does not expose internal content selection/admission or checkpoint replacement.)	partial coverage (Both automatic supply and requested retrieval are established. Operations within each chain are distinguished; returning a requested tool/resource result is not a second push. SDK history selection is wired but its internal summary delivery and arbitrary settings consumption remain unresolved.); Automatic vector selection during channel/chat prompt building: known — push [wired] (Foreground query triggers code selection without a memory-tool request.); Fixed evolved snapshot loading and working-memory projection: known — push [wired] (Only fixed evolved files passing formatting conditions are automatically delivered; notes/corrections are not silently included.); Host chat artifact/checkpoint supplement: known — push [wired] (Fresh host supplement replaces prior tagged context under Murph; it is not full transcript compaction.); Requested vector/config/changelog resources and transcript search: known — pull [wired] (External resource exposure does not prove a deployed client.); Requested direct note/current-file reads: known — pull [afforded] (Agent notes remain outside automatic injection.); Dashboard memory inspection: known — pull [wired] (This read route does not establish arbitrary SDK injection.); Reflection batch delivery of queued evidence: known — push [wired] (Cadence selects and stages retained summaries for the update invocation.); Retained SDK-ID lookup for resumed foreground execution: known — push [wired] (This describes the host selection and requested resume, not complete SDK payload supply.); Dynamic-definition reload and scheduled job dispatch: known — push [wired], push [wired] (These are later consumption of retained capabilities and handoffs, not vector-memory retrieval.); Arbitrary user Markdown, installed settings/skills/hooks and dependency conversation/summary delivery: uninspected (The records do not resolve every actual dependency selection and delivery route. The known host resume-ID push does not establish complete SDK read-back coverage.)	partial coverage (Push signals are classified at actual selectors. Coarse readiness/budget/availability and targeted matches are separate findings. Requested lexical transcript and vector searches are pull operations and add no push-signal values. SDK/settings selectors remain unresolved.); Automatic vector query relevance selection: known — inferred-embedding [wired], inferred-lexical [wired] (Trigger: foreground text. Inputs: text embeddings and sparse token features. Selected parts: episode/fact/procedure candidates.); Vector availability, validity, recency and prompt-budget filtering: known — coarse [wired] (Trigger: prompt construction. Inputs: readiness, score/age/importance/access and estimated remaining budget. Selected parts: whole fact/procedure sections and bounded episodes.); Fixed evolved-file and working-memory projection: known — coarse [wired] (Trigger: assembly/reload. Inputs: fixed file slots, nonempty/content-line conditions and line count. Selected parts: evolved sections and working-memory head/tail.); Chat artifact/checkpoint continuity selection: known — identifier [wired], coarse [wired] (Trigger: later chat query. Inputs: current session ID and supported page-tool identities; event/timeline tails, parseability, filters and caps. Selected parts: matching artifacts and recent checkpoint metadata.); Evolution batch selection: known — coarse [wired], inferred-judgment [wired] (Trigger: cadence drain. Inputs: queue age, eligibility/retries and batching. Selected parts: staged session summaries and gate reasons.); Runtime SDK session-ID selection: known — identifier [wired], coarse [wired] (Trigger: foreground query. Input: channel/conversation key and expiry. Selected retained part: matching active SDK-history handle; payload is opaque.); Dynamic catalog enumeration and scheduled dispatch: known — coarse [wired] (Trigger: server creation or scheduler tick. Inputs: catalog availability or due/enabled/status. Selected parts: retained definitions or due tasks.); SDK internal history/checkpoint selection and user/project extension selection: uninspected (Known host IDs and display summaries do not classify the hidden selector or selected payload.)	partial coverage (The profile uses the automatic trace-fed durable-artifact definition, not demonstrated improved capacity. Wired vector extraction, reflection and durable chat summaries qualify; raw retention alone does not. Local procedure absence does not become a system-wide no. Opaque SDK summaries, mixed foreground writes and initial state prevent complete coverage.); Automatic channel trace-to-episode/fact-to-later-model chain: known — yes [wired] (Durable contextual artifacts qualify without novel claims or measured benefit.); Queued conversational evidence to evolved/reflection files to later consumers: known — yes [wired] (Fixed-loader files reach foreground prompts; other files can shape later reflection and require promotion for automatic foreground delivery.); Foreground experience-fed note writing and later use: known — yes [afforded] (This remains an affordance, not an observed or deterministically scheduled note-writing episode.); Translated chat events to durable timelines to later artifact context: known — yes [wired] (The generated summary qualifies through retention and later contextual supply, not merely browser display.); Shipped automatic trace-to-procedure producer/outcome loop: known — no [wired] (The bounded non-test source caller search establishes no shipped automatic procedure-learning chain; external/preexisting procedure admission is assessed separately.); Initial/external procedure and vector writes: not-determinable (Unresolved initial producers cannot inherit the local shipped-procedure no; their trace inputs and later behavior-shaping updates are not established.); SDK compaction/checkpoint production and later consumption: uninspected (Host compaction counters do not reveal whether or how an automatic retained summary reaches a later invocation.); External memory/extension/handler edits, initial files and optional session-log population: not-determinable (Generic caller edits or a named log path do not establish the trace-fed durable-update chain.); Scheduled handoff generation and direct executable capability creation: not-determinable (Automatic admission and later use are known, but records do not establish that the task/handler was extracted from recorded activity rather than an immediate open request. Neither yes nor no follows for all these writes.)	partial coverage (Authority follows each actual consumer, not the prompt container or write agency. Retained evidence can shape reflection while its output gives later instruction/knowledge. Arbitrary Markdown, dependency settings effects and SDK payload authority remain unresolved; static check code/constitution are excluded from accumulated memory.); Vector prose and suggested steps at the foreground model/tool-result consumer: known — knowledge [wired] (Advice/evidence delivery is wired; compliance and improved action are unobserved.); Retained vector content, embeddings and episode salience metadata at retrieval/reranking: known — ranking [wired] (Fixed scoring code does not remove the ranking force of retained inputs.); Fact validity state at default recall and conflict selection: known — enforcement [wired], routing [wired] (Temporal recall can bypass the default expiry exclusion.); Evolved persona/profile/strategy prose at later foreground prompt consumer: known — instruction [wired], knowledge [wired] (The fixed loader and more-than-one-content-line formatter condition bound delivery.); Queued messages, current reflection files and notes at reflection-update consumer: known — learning [wired], learning [afforded], knowledge [afforded] (Learning authority denotes an update input, not demonstrated improved capacity.); Working-memory projection at foreground model: known — knowledge [wired], instruction [wired] (The host explicitly asks the model to trust the selected notes; this is not enforcement.); Append notes read through foreground tools: known — knowledge [afforded] (These notes are deliberately excluded from automatic injection.); Queue key/age/retry/poison state at batching and rejection consumers: known — routing [wired], enforcement [wired] (Control fields differ from the natural-language evidence they gate.); SDK-ID and timestamp rows at session-resume consumer: known — routing [wired] (This is authority of the host selector, not of opaque transcript contents.); Chat transcript and host event/timeline artifacts at recovery and continuity consumers: known — knowledge [wired] (Checkpoint counters do not become hidden standing instructions.); Chat session/sequence/status metadata at bound transcript access and continuity selector: known — enforcement [wired], routing [wired] (Access and selection consequences are distinct from prose advice.); Evolution reason/change metadata at requesting MCP/dashboard reader: known — knowledge [wired] (Automatic evolved prompts omit these reasons.); Dynamic handler identity/type/path at dispatch: known — routing [wired] (Dispatch fields select executable operations; executable force alone is not the controlled enforcement value.); Arbitrary retained handler code at OS execution: not-determinable (The records establish executable force under process rights, but not whether each body has a controlled instruction, knowledge, enforcement, validation or other listed effect. Execution alone is not a permit/refuse mechanism.); Retained dynamic descriptions at model tool selection: known — routing [afforded] (This is selection guidance, not observed choice or an enforced route.); Changed skill/subagent prose at requested SDK settings consumption: known — instruction [afforded] (Host settings exposure is wired; exact dependency loading/activation remains uninspected.); Scheduled task and dispatch policy at timer and later runtime: known — instruction [wired], routing [wired], enforcement [wired] (Run classification/delivery are not task correctness acceptance.); User-scope Markdown at dashboard reader and optional foreground file reader: known — knowledge [wired], knowledge [afforded] (These reader paths do not establish arbitrary SDK automatic injection.); Arbitrary user-scope Markdown at automatic SDK consumers: uninspected (SDK project/user settings exposure does not establish automatic consumption of every file or its downstream force.); Changed plugin/hook/YAML controls and SDK retained payload effects: uninspected (Settings exposure does not establish all dependency-specific consumers or effects; opaque history may include instructions, tools and summaries.)	AAS-2026-10-15-phantom-cc566f03ac40-01
pi-posthorse	2	Retained note content and revisions; host session-file content, images, IDs, projection/edit relationships and explicit/custom/native/legacy checkpoints consumed by Posthorse; retained reminder prose and eligibility fingerprints; their admission, maintenance and later-consumer routes. Includes unresolved initial notes, external sectional editors and native checkpoint production as named alternatives. Excludes external model weights/provider state, transient reset maps, page reservations and caches, regenerated static guidance, display-only cards/metadata as an independent memory payload, and development evaluator/checker outputs not consumed by production updates.	known coverage (The accepted inventory identifies notes and host session files as the retained substrates. External model weights and transient access state are not this memory boundary.); Durable notes, including pre-existing content and migrated locations: known — files [wired] (Initial authorship uncertainty does not obscure the file substrate.); Host journal: raw entries, edited versions, images, reminders, explicit/custom/native and legacy checkpoints: known — files [wired] (The recorded host-file contract covers these retained parts. In-process projections are access paths, not additional durable stores; host persistence correctness is unverified.)	partial coverage (Model/software admission is positive on inspected routes. No manual finding is established merely by optional human authorship or workflow initiation; external and initial controls remain unresolved.); Foreground model note-content admission through notes write/append: known — automatic [wired] (This classifies the evidenced model path, not generic caller identity or physical filesystem I/O.); Other caller admission and pre-existing note initialization: not-determinable (Generic caller identity and unknown initialization do not establish either explicit human content decision or autonomous admission.); External sectional note editing: uninspected (The external writer is afforded, but its implementation and per-content approval policy are uninspected.); Foreground model explicit-handoff admission: known — automatic [wired] (A user initiating the task is not a human approval gate for this write.); Automatic custom recovery admission and note migration: known — automatic [wired], automatic [wired] (Construction/admission and relocation are separate mechanisms, both software-controlled.); Reminder record admission: known — automatic [wired] (No per-reminder human content approval is required by the recorded route.); Host raw-journal persistence, context-edit replacement and native checkpoint admission: uninspected (External host generation, capture and replacement controls are not inspected. Reading/selecting these checkpoints does not establish their write agency.)	known coverage (The memory report inventory distinguishes requested recall from checkpoint and reminder-metadata automatic supply. Host deployment/operation uncertainty does not make the recorded directions indeterminate.); Notes and detailed history recall, including stored images and old checkpoints: known — pull [wired] (Automatic paging/delivery fulfills the request; it is not an additional unsolicited push.); Explicit, custom automatic and existing nonautomatic checkpoint delivery: known — push [wired] (The known consumption path includes native/legacy checkpoint material reused by custom recovery; its external generation is a separate uncertainty.); Retained reminder metadata lookup and model-visible filtering: known — push [wired] (Reminder prose is also available through the separately classified history pull route.)	known coverage (Known push units have coarse or identity-matching selectors. Pull units are locally inapplicable, not a system-wide pull-only conclusion; no task-relevance inference is attributed to these fixed selectors.); Automatic recovery source selection on compaction hook or threshold fallback: known — coarse [wired], identifier [wired] (Budget/recency selection and actual provenance identity matching are distinct warrants; no semantic embedding or lexical push is recorded.); Explicit handoff selected for committed next-window projection: known — coarse [wired] (Callback ID matches authenticate the newly proposed write; they are not independent identity-targeted retrieval of an older retained memory part.); Nonautomatic checkpoint carried by automatic ancestry selection: known — coarse [wired] (Unknown external checkpoint generation does not obscure the inspected coarse reuse selector.); Retained reminder metadata selected by automatic hook lookup: known — identifier [wired] (An actual fingerprint match affects selection, not merely a catalog ID.); Caller-requested note/history pages: inapplicable (These units are pull; their literal/path/ID targeting is not a push-selection signal. All other direction units have been classified separately.)	partial coverage (The custom continuation update supports yes at wired strength. Other included writes remain unresolved; no local or system-wide no is asserted without a bounded absence record.); Automatic custom recovery: retained session/tool records to durable next-window continuation: known — yes [wired] (The wired route meets this axis even though it reshapes existing claims and no improved capacity or exercised benefit is observed.); Model note maintenance and explicit handoff formulation: not-determinable (The records establish content writes and later consumption but do not establish which traces actually feed model formulation. Requested task-state maintenance alone cannot establish a qualifying route or bounded no.); Native checkpoint and external note-editor updates: uninspected (Their implementations and input-to-update chains are uninspected; custom recovery's positive witness cannot resolve them.); Raw journal and retained reminder production: not-determinable (Raw capture alone is not trace learning. Fixed reminder generation and retained metadata effects do not establish a qualifying trace-fed behavior-shaping update, nor do these records warrant a universal bounded negative.)	partial coverage (Actual recorded consumers support multiple forces. Unknown arbitrary-content effects prevent a complete authority inventory; transient page/callback controls and shipped static guidance are excluded.); Retrieved note prose at the foreground model: known — knowledge [wired] (No standing-preference consumer or executable truth gate is recorded for arbitrary note instructions.); Journal prose and image payloads at the foreground model: known — knowledge [wired] (Knowledge force does not imply truth, fresh authorization or actual reliance.); Explicit and nonautomatic checkpoint prose at the next model: known — knowledge [wired], instruction [wired] (Instruction is supported for the fixed explicit fallback, not inferred for every native or caller-authored claim.); Retained traces and checkpoints at the automatic recovery builder: known — learning [wired], routing [wired], validation [wired], enforcement [wired] (Update-consumer force is separate from the output's later knowledge and instruction force; learning here does not assert improved capacity.); Automatic recovery output at the later model: known — knowledge [wired], instruction [wired] (The model interprets these warnings; there is no enforced truth license.); Retained content and metadata at recall selectors: known — ranking [wired], routing [wired] (Fixed ranking code still consumes retained features; it need not be learned ranking policy.); Retained reminder prose and fingerprints at later model/eligibility code: known — instruction [wired], routing [wired], enforcement [wired] (The metadata effect is executable while prose compliance remains unobserved.); Other arbitrary note content and unrecorded native-checkpoint consumer effects: not-determinable (The records do not individuate all possible downstream instructions or other forces of arbitrary retained content. No additional force is inferred from possibility alone.)	AAS-2026-10-14-pi-posthorse-58b9647d2f29-01
pi-self-learning	2	Accumulated daily reflections, canonical index text and ranking metadata, core and long-term renders, monthly summaries and process-local reflection notes; their initialization/migration, automatic updates, maintenance and later model/code consumers. Existing file content is included with unresolved original provenance and admission control. Branch conversations and tool results are original learning inputs, not extension-owned durable stores. Optional Git history is included as conditional retention, not a shipped read-back route. Excludes static boilerplates and consultation policy, ordinary settings and branch controls, enclosing-host transcript persistence and provider parameters: these are dependencies or configuration, not accumulated memory maintained by this extension.	known coverage (Inventory covers every extension-maintained accumulated object in scope. External model state and host transcript storage are not asserted as extension memory substrates.); Daily journal, canonical index text and ranking metadata, core, long-term and monthly current files: known — files [wired] (Records enumerate both canonical parts and all rendered/summary file alternatives; unavailable contents do not obscure their implemented substrate.); Optional version history of accumulated memory files: known — repo [wired] (Covers commit paths for reflection and both maintenance producers; no shipped history consumer is implied.); Process-local recent reflection notes: known — in-memory [wired] (This alternative is volatile and has no durable reload or branch/root key.)	partial coverage (All shipped write mechanisms are classifiable; included initial/external content control remains unresolved, so the axis is not wholly known.); Reflection/repair admission, daily append, index/ranking updates, rendering and process-note admission: known — automatic [wired] (Starting or enabling reflection does not make its generated content manually admitted.); Monthly summary admission, redistill replacement and migration/default admission: known — automatic [wired] (Known automatic import/admission remains distinct from unknown original creation of its inputs.); Original admission and replacement of pre-existing or externally edited memory files: not-determinable (Files can be externally edited, but no actual operator content decision or initial control is recorded. Physical editability alone does not establish manual admission.)	known coverage (The inventory separates request, automatic selection and delivery within each chain. Git history has no shipped later consumer; current-file and note paths cover the recorded alternatives.); Unrequested future-turn notes/file supply and selected retained content supplied to maintenance completions: known — push [wired] (Operator-started maintenance still includes a distinct automatic selection and delivery operation to its model.); Update-code requests for existing canonical state and explicit month/redistill requests for retained input: known — pull [wired] (Files are actual later update inputs, not merely storage capabilities.); Foreground on-demand consultation of core, daily, monthly and long-term paths: known — pull [afforded] (Affordance remains weaker than wired delivery and is not upgraded by the automatic bundle witness.)	known coverage (All recorded push operations have named triggers, selectors and delivered parts. Pull path identifiers alone add no signal finding.); Before-agent-start note and file push: known — coarse [wired] (Selected parts are note fragments and configured core/daily/monthly prefixes; a fixed path or chronological filename ordering is not inferred lexical relevance.); Monthly command's selected daily-file push to summary model: known — identifier [wired], coarse [wired] (The upstream month request supplies the selector input; the downstream model receives selected retained prose automatically.); Redistill command's selected canonical-rule push to rewrite model: known — coarse [wired] (Selected retained part is rule text/kind; ranking metadata chooses targets but is not supplied as model evidence.)	partial coverage (Positive trace-fed routes coexist with unresolved legacy inputs. No system-wide no or observed-learning claim follows from missing execution evidence.); Automatic or manually initiated reflection over branch activity, including repair and durable daily/index/render descendants: known — yes [wired] (A wired trace-learning route does not establish observed use, improved performance or parameter training.); Monthly summary and redistill over reflection-derived retained material: partial — yes [wired] (Inputs may instead include legacy/externally edited prose of unknown origin; the records cannot classify every maintenance input as trace-derived.); Existing-content initialization and CORE fallback with unresolved original source: not-determinable (Automatic import is known, but original activity provenance is not. Neither a qualifying trace-fed write nor a local no can be established for all imported content.); Process-local note as a standalone output: inapplicable (The note has later consumers but no durable persistence/reload. The encompassing reflection route qualifies through its durable siblings, not through this volatile fragment alone.)	known coverage (Coverage is of implemented or afforded consumers/effects, not actual use. Structural parsers and static policy are not accumulated validation criteria; no retained-memory enforcement authority is inferred from strict-mode wording.); Retained prose and notes supplied to later foreground agents as historical evidence: known — knowledge [wired] (Covers automatic delivery; host application and changed answers remain unobserved.); Retained core, daily, monthly and long-term prose on requested external-agent consultation paths: known — knowledge [afforded] (Long-term content belongs to this channel rather than the automatic bundle.); Retained prevention rules and watch-outs at later foreground interpretation: known — instruction [afforded] (This is the retained advice's potential standing force, separate from the excluded static system-prompt consultation instructions.); Retained experience at reflection, monthly-summary and canonical-rule rewrite consumers: known — learning [wired] (Names the actual update consumers rather than inferring authority from storage alone.); Index content identity/kind and ranking metadata at rendering and redistill targeting: known — ranking [wired], routing [wired] (Symbolic consumer effects are classified separately from prose interpretation; recurrence metadata is not a correctness criterion.)	AAS-2026-10-15-pi-self-learning-1a318610d4b1-01
playground	2	Includes use-accumulated workspace files and subsequently modified profiles; retained job output/access state; self/shared pile content and opaque native-worker alternatives; standalone timestamp, wiki and telemetry retained material; and separately labelled summary/cover doctrine. Excludes initial static profiles and shipped prompts as retained memory, signer keys, credentials, worker mappings, bundle caches and transport-session machinery, as well as external model weights and excluded archive-import execution. Declared context mechanisms are compared only at their local claimed/afforded basis, never as current provider wiring.	partial coverage (Files and RAM are established for the inspected retained inventory. Additional native-worker retained state is opaque; file custody does not enumerate that state.); Accumulated workspace files, subsequently modified profiles, self/shared piles and standalone script input/output piles: known — files [wired] (The records identify filesystem storage, including opaque pile bytes. Native encoding is not another established substrate.); Retained job output and accumulated polling/terminal-observation metadata: known — in-memory [wired] (Retention is bounded to the same daemon; delivered JSON is not a separate durable substrate.); Documented summary/cognition/archive branch storage: not-determinable (Branch and blob descriptions do not establish a controlled storage substrate for the unwired context system.); Additional retained state inside the native worker: not-determinable (The fixed mounted pile is covered separately. The external implementation prevents enumeration of other operative storage.)	partial coverage (Software admission is established for output capture, onboarding wrappers and maintenance. Caller-driven content decisions and external native admission remain unresolved.); MCP file replacement and arbitrary shell/profile mutation: not-determinable (No per-content human gate is imposed, but generic callers can be humans or programs; absence of a gate does not identify their admission decisions.); Job output capture and retention-state updates: known — automatic [wired] (Capture/state updates are software-driven once work is launched; workflow launch is not content approval.); Fixed jail onboarding import/tenant registration wrapper: partial — automatic [wired] (Software initiates fixed additions; per-fact native acceptance and initial corpus/shared seed control remain external.); Timestamp migration and changed-text wiki edit wrappers: known — automatic [wired] (Operator invocation/dry-run choice does not manually approve every retained change; external semantic acceptance is a separate unresolved mechanism.); Requested model episode consolidation: known — automatic [claimed] (This unit covers the model-directed request only, not every explicit memory-create caller.); Initial piles, explicit memory-create callers, telemetry producers and external native admission: not-determinable (Manual-curation terminology and an operator-supplied file do not establish an explicit human decision for every retained item. Native and trace-producer admission are opaque.)	partial coverage (Requested files, shell access, replay and diagnostic reads are pull; documented cover and modified-profile application are push. Native semantic alternatives remain unresolved.); Requested workspace read/exec and retained job polling: known — pull [wired] (Delivery fulfills a consumer request; literal paths/job IDs do not turn it into independent push.); Use-modified session profile on fresh login-shell invocation: known — push [wired] (Only the subsequently changed profile is use-accumulated memory; clean file tools bypass it.); Standalone native maintenance/telemetry input reads: known — pull [wired] (Operator script invocation requests retained material; internal filtering within that request is not a second push.); Documented automatic cover and explicit temporal recall/provenance: known — push [claimed], pull [afforded] (Directions are classified for the described mechanisms only, not attributed as implemented provider behavior.); Native faculty/worker semantic read-back: not-determinable (Mounted state and forwarded calls do not reveal whether requested or unsolicited native content reaches later semantic consumers.)	partial coverage (Coarse profile availability and declared temporal/budget cover selection are positive push signals. Pull selectors are inapplicable locally, not evidence that opaque native alternatives are pull-only.); Automatic application of a use-modified profile: known — coarse [wired] (Trigger is a login-shell start; selector is login-mode/profile availability; selected part is the entire persisted profile. No content identity matching is recorded.); Documented cover push: known — coarse [claimed] (Trigger is model invocation; inputs are temporal root/child structure and remaining context budget; selected parts are retained summary chunks.); Explicit file/job/script/temporal recall requests: inapplicable (These units fulfill explicit requests. Path, job ID, sequence cursor and branch/tag IDs are pull selectors, not identifier-based push.); Native worker/faculty unsolicited delivery alternatives: not-determinable (No native selector input or delivered part is inspectable; known provider pull cannot resolve this included alternative.)	partial coverage (Requested model consolidation supplies a claimed positive. Raw replay, diagnostics and maintenance have no recorded qualifying update, but lack canonical bounded-absence records for local no values. External automatic transforms and actual summary admission remain unresolved; no observed benefit is asserted.); Requested episode-to-memory-to-later-cover route: partial — yes [claimed] (Prompts request a model update and documents retain/deliver summaries, but actual producer/admission/wiring and legacy alternatives remain unresolved.); Provider job-output capture/replay: not-determinable (Records describe raw retention/rendering, not a durable derived summary or parameter update. They supply no canonical bounded-absence record for a controlled no; later caller updates are external.); Telemetry diagnostic aggregate production: not-determinable (The inspected report ends at stdout without a retained adaptation. No qualifying yes or canonical bounded-absence record for no is supplied; downstream operator updates are external.); Timestamp/link maintenance wrappers: not-determinable (Encoding/reference transformations do not establish a qualifying trace-fed update. A canonical bounded-absence record for no and external update-consumer coverage are not supplied.); Arbitrary workspace programs, initial native corpus, telemetry producer and native-worker transformations: not-determinable (Uninspected producers/admission and later consumers prevent a complete positive or negative trace-learning classification.)	partial coverage (Authority is attached to particular consumers; native semantic effects and arbitrary payload interpretation remain unresolved. No learning benefit or compliance is inferred.); Accumulated workspace prose at MCP clients: known — knowledge [wired] (This classifies the delivered reference/context channel, not truth or actual client reliance.); Stored scripts and use-modified profile instructions at fresh shells: known — instruction [wired] (Only executable persisted content and defaults are included; initial shipped/provisioned directions are excluded.); Retained job output at the client and sequence metadata at polling code: known — knowledge [wired], routing [wired] (Client delivery is advisory; fixed selectors act on metadata rather than semantic truth.); Retained handle terminal/final-page-observation state at subsequent job admission: known — validation [wired], enforcement [wired] (These criteria govern capacity recycling and refusal, not acceptance of an answer.); Retained timestamp facts at migration and wiki fragments at rewrite: partial — routing [wired] (Wrapper selection is known; actual later semantic readers of changed native content are external.); Retained telemetry measurements at diagnostic arithmetic and operator display: known — ranking [wired], knowledge [wired] (Fixed arithmetic and ordering are inspected; causal measurement truth and adaptation are not.); Documented summaries at external context assembly/model and episode inputs at requested consolidation: partial — knowledge [claimed], learning [claimed] (Declared context supply is supported, while actual update control/consumer wiring and alternative legacy mechanisms remain unresolved.); Self/shared pile content at external faculty/native-worker consumers: not-determinable (Custody, signer identity and policy labels do not establish downstream knowledge, enforcement or other semantic force.)	AAS-2026-10-14-playground-88ee8a292f9b-01
Pond	2	Included are accumulated canonical sessions/messages/parts and raw provider rows, upstream summary/checkpoint carriers, durable search text and message vectors, row-meta caches and sync progress, changed capture configuration, their admission, maintenance, transfer, recall and restored-session routes. Other codecs' opaque payload/checkpoint alternatives, unspecified library admissions, export consumers and untraced interactive delivery remain included uncertainties. Excluded are fixed embedding weights, static shipped skills/tool guidelines, transient check results and benchmark outputs without an established later behavioral consumer, and external hosts' internal context management. Canonical storage is not assumed to resolve the meaning or generation of its opaque contents.	partial coverage (Positive local and advertised remote branches remain distinct; included codec/export alternatives prevent complete inventory coverage.); Local canonical Lance datasets, full parts/raw rows/carriers, and archive/restored files: known — files [wired] (Covers the local canonical/file branch, including opaque bytes as stored objects, without claiming codec completeness or host loading.); Advertised remote canonical storage in an operator-owned S3 bucket: partial — service-object [claimed] (Remote object-store implementation internals and deployed storage are not inspected; exact backing behavior remains unresolved.); Version-keyed row-meta cache, resident snapshots, sync cursor/status and changed adapter configuration: known — files [wired], in-memory [wired] (Both persistent files and operative resident cache parts are classified; ordinary current-batch buffers are excluded.); Additional codec-specific retained state and standalone export alternatives: uninspected (Shared canonical storage is known, but uninspected codecs and subsequent export consumers cannot establish an exhaustive substrate inventory for their operative alternatives.)	partial coverage (Automatic admissions and explicit manual content/configuration choices coexist with unresolved initial and generic-caller control.); Adapter/HTTP canonical admission, searchable projection, vector updates and cursor/cache production: known — automatic [wired] (Known subsequent admission is independent of upstream authorship.); Explicit capture configuration and first-run bootstrap admission: known — manual [wired], automatic [wired] (The two evidenced admissions are retained independently; a recorded decline can prevent bootstrap.); Operator-selected session/target replacement through CLI/Pi restoration: known — manual [wired] (This is content-selected restoration, not the read that chooses a checkpoint and not evidence of automatic summary generation.); Bulk copy/import and generic library write admission: partial — automatic [wired] (Generic caller identity does not resolve human control over initial supplied/replacement content; wire admission rules do not cover every library write.); Original upstream checkpoint generation and other codecs' admission alternatives: not-determinable (Original generation/control and uninspected alternatives are unresolved; known Pond capture does not establish upstream approval or automatic generation.)	partial coverage (Requested human/model reads coexist with automatic internal metadata/update supply; untraced alternate consumers prevent a complete direction inventory.); Requested search/get/SQL, transfer and operator-selected restoration: known — pull [wired] (Tool/model answering and restored host continuation are unobserved but the requested delivery paths are wired.); Restarted sync's automatic row-map/cursor fallback into freshness planning: known — push [wired] (This is a software metadata consumer, not unsolicited host-model context insertion.); Inline/backlog embedding row selection into numerical update: known — push [wired] (Automatic update-input delivery is separate from the later requested semantic-search return.); Interactive editor-reference insertion, export-file consumers and other codec checkpoint consumption alternatives: uninspected (The supplied requested-tool account is selective; no accepted record traces editor-reference insertion or later export reading, and remaining codec consumers are uninspected. Their directions cannot be inferred from known pull routes.)	partial coverage (Known internal push selectors establish positives; pull-only inapplicability is local and does not erase untraced alternate delivery.); Known requested search/get/SQL/restore/transfer operations: inapplicable (These are pull operations. Query lexical/embedding scores and requested IDs do not classify a push signal.); Automatic restarted-sync supply of row-map or validated cursor progress: known — coarse [wired], identifier [wired] (Availability and actual lineage identity comparison are distinct selectors; subsequent freshness decisions are effects of the selected progress.); Automatic eligible search-text supply to inline/backlog/force embedding: known — coarse [wired] (Eligibility filtering is not embedding-inferred selection; embeddings are produced after selection.); Untraced editor insertion, export consumers and opaque checkpoint alternatives: uninspected (Unknown direction/selector facts prevent treating all remaining units as pull-only or assigning a push signal.)	partial coverage (Two wired artifact-update witnesses establish yes locally; unresolved original summary and alternate write/consumer routes prevent complete coverage.); Automatic recorded conversational parts to durable search text to later recall/ranking: known — yes [wired] (The derived route qualifies; retaining untouched logs alone is not the warrant.); Automatic session projection to retained message vectors to later semantic ranking: known — yes [wired] (Includes initial and maintenance production; no observed benefit, revised method or weight training is asserted.); Original upstream checkpoint/branch-summary generation and other codec trace-fed alternatives: not-determinable (Summary labels and native replay establish retention/possible delivery, not the automatic original write or complete later-consumer chain; other codecs remain uninspected.); Capture configuration, progress/cache state, bulk/library admissions and opaque restored payloads: not-determinable (Operational persistence alone does not establish a qualifying experience-to-behavior-shaping write; generic content control and opaque host consumption remain unresolved. No system-wide no follows.)	partial coverage (Actual advisory, ranking, update, routing and retained-key veto consumers are classified separately; opaque restoration prevents complete authority coverage.); Historical readable material delivered as requested search/get/SQL results to a host model or human: known — knowledge [wired] (Known delivery force is advisory historical evidence; interpretation and changed answers remain unobserved.); Search text and message vectors consumed by later lexical/semantic ranking: known — ranking [wired] (Ranking influence does not imply factual endorsement or retrieval benefit.); Retained conversational traces/search text at automatic projection and embedding update consumers: known — learning [wired] (This names the update consumer's force, not improved capacity or training of embedder parameters.); Canonical identity/lineage, row-meta, watermarks and persisted capture configuration at fixed selectors: known — routing [wired] (Access metadata routes operations rather than supplying model instructions.); Stored canonical keys and labels consumed by additive admission: known — validation [wired], enforcement [wired] (These are retained-state effects, distinct from static schema checks, HTTP gates or a nonexistent purge denylist.); Native checkpoint/raw payload at restored host loader; foreign reconstruction and other codec alternatives: partial — routing [wired] (File handoff is wired, but opaque payload force at external host interpretation and other codecs remains unresolved; recorded System labels cannot establish instruction authority.)	AAS-2026-10-14-pond-7a589a889ca1-01
prime-agent	2	Included: all five use-authored or use-refined local/global harness kinds and their access metadata; accumulated session messages, compaction/branch summaries and refinement audit/event projections; serialized/revived namespace and distinct manifest/notice; use-authored or changed skill/resource files; retained child results; incompletely traced goals/quota/scheduling, fork and observe/message/recovery alternatives; unresolved provider-side accumulated/adapted state, if any. Excluded: unchanged shipped instructions and library machine templates, ordinary live factory run registry/ports as such, fixed inference weights as such, release deployment payloads, and export/telemetry/workspace artifacts without an established later memory consumer. External provider implementation is not frozen; its unresolved alternative is retained as uncertainty, not a positive storage or learning claim.	partial coverage (File and live-memory substrates are supported for inspected parts. Unresolved retained controls, fork/recovery alternatives and opaque provider state prevent complete storage coverage; JSON and dill are encodings, not additional substrates.); Local/global harness text, skill contracts, subagent templates and stored factory specifications: known — files [wired] (All five harness kinds share the retained JSON file store; the live factory registry is a different object.); Session messages, summaries, branch-summary entries, refinement audits/history and delivered custom rows: known — files [wired], in-memory [wired] (Persistent sessions have file-backed histories and live reconstructed context; no-session managers retain only live memory.); Namespace payload and manifest, with later revived namespace: known — files [wired], in-memory [wired] (Payload files are distinct from reporting metadata and their restored process values.); Changed skill bodies and executable packages: known — files [afforded] (The filesystem admission route is afforded; whole-route implementation characterization remains disputed, so it is not upgraded.); Included persisted goals, quota parks, autonomous/scheduled wake-up controls, fork transfer and child observe/message/recovery alternatives: uninspected (The memory inventory identifies these included alternatives but does not fully trace their retained parts, admission or replay. Their storage classification remains uninspected; the known routes do not enumerate the whole boundary.); Opaque provider-side accumulated state, if any; fixed inference weights are not use-derived memory: uninspected (Provider caching, adaptation and parameter provenance are uninspected. No storage value is inferred from ordinary remote inference or a mutable endpoint. External implementation is outside frozen evidence, preventing resolution of this included alternative.)	partial coverage (Automatic admission is wired for refinement, model cells, summaries and checkpoints; manual resource authoring is afforded. Generic caller identity, raw store replacements, initialization and untraced controls retain uncertainty.); Refinement proposal admission, including operator-triggered and automatic review routes: known — automatic [wired] (Operator scope/focus and factory opt-in do not supply per-edit content approval.); Foreground-model direct harness edits through Python: known — automatic [wired] (The explicitly model-generated path is known; generic helper callers are separated below.); Other direct helper callers, initial populated-store admission and unrestricted raw file replacement: not-determinable (Caller-authored content is accepted, but generic caller identity does not establish whether an explicit human content decision occurred. Initial store control and raw replacement approval are also unresolved.); Session accumulation, summary commitment and namespace capture: known — automatic [wired] (Calling compaction/navigation does not make the resulting model summary manually admitted; checkpoint serialization likewise uses software checks.); Human-authored and model-authored changed skill/resource files: known — manual [afforded], automatic [afforded] (The supplied route explicitly permits humans to choose/edit content and models to author it; discovery is not a mandatory human approval stage.); Included persisted goals, quota parks, autonomous/scheduled wake-up controls, fork transfer and child observe/message/recovery alternatives: uninspected (The memory inventory identifies these included alternatives but does not fully trace their retained parts, admission or replay. Their admission control classification remains uninspected; the known routes do not enumerate the whole boundary.); Opaque provider-side accumulated state, if any; fixed inference weights are not use-derived memory: uninspected (Provider caching, adaptation and parameter provenance are uninspected. No admission control value is inferred from ordinary remote inference or a mutable endpoint. External implementation is outside frozen evidence, preventing resolution of this included alternative.)	partial coverage (Both pull and push occur at distinct operations, sometimes within one request chain. Live factory port bindings alone are not accumulated read-back; retained child result collection is. Untraced alternatives prevent a complete direction assessment.); Harness digest to foreground context: known — push [wired] (Code selects retained hints without a foreground request for those entries.); Foreground requested harness get/list/search/overview/snapshot: known — pull [afforded] (Requested returns are pull, not another push because helper code chooses matches.); Factory stored specification request and automatic subagent-template resolution: known — pull [wired], push [wired] (The requested machine return/launch is distinct from code selecting full referenced retained child templates.); Refiner automatic state/history inputs: known — push [wired] (Starting refine requests a transformation, not each retained history row. Host supplies bounded state/history to planner.); Session restore/navigation request and automatic reconstructed summary/tail context: known — pull [wired], push [wired] (Caller session/branch selection is a request; host context reconstruction selects and supplies retained content within that chain.); Kernel revival and independent namespace notice: known — push [wired] (Kernel boot restoration and notice delivery do not require a later cell to request each retained value/name.); Changed skill inventory and requested full instruction body: known — push [afforded], pull [afforded] (Inventory is coarse automatic supply; full body invocation/inspection is requested. Whole-route classification stays afforded.); Retained completed child results: known — pull [wired] (Explicit collect is distinct from live factory bindings and status previews.); Included persisted goals, quota parks, autonomous/scheduled wake-up controls, fork transfer and child observe/message/recovery alternatives: uninspected (The memory inventory identifies these included alternatives but does not fully trace their retained parts, admission or replay. Their read-back direction classification remains uninspected; the known routes do not enumerate the whole boundary.); Opaque provider-side accumulated state, if any; fixed inference weights are not use-derived memory: uninspected (Provider caching, adaptation and parameter provenance are uninspected. No read-back direction value is inferred from ordinary remote inference or a mutable endpoint. External implementation is outside frozen evidence, preventing resolution of this included alternative.)	partial coverage (Lexical, coarse and identifier-based pushes are supported on separate selectors. Requested search is pull despite lexical ranking. Unresolved push alternatives prevent axis inapplicability or complete signal coverage.); Boundary-triggered harness digest entry selection: known — inferred-lexical [wired], coarse [wired] (Trigger is cold start/resume/compaction or stale/missing digest; goal and recent text terms rank retained entries, while availability and clipping bound delivery.); Factory automatic stored subagent-template push: known — identifier [wired] (Triggered while resolving the requested machine; its retained reference names select delivered full prompt templates.); Refiner automatic history/state projection: known — coarse [wired] (Trigger is planner request; latest history and bounded overview select retained result summaries and current entries.); Automatic compaction/branch context reconstruction: known — coarse [wired] (Trigger is commit/rebuild for the selected path; cut, tail and abandoned-path budget select retained summary/context. A stored boundary ID alone does not prove identifier-based push.); Checkpoint restore and namespace notice: known — coarse [wired] (Trigger is artifact-backed boot or post-compaction probe; snapshot target, eligible names and size/serialization limits bound retained payload/notice.); Changed skill inventory push: known — coarse [afforded] (Trigger is startup/reload; discovered availability provides names/descriptions/imports. A catalog name alone is not identity-matched body delivery.); Requested harness detail/search, stored factory lookup, session selection and child collection operations: inapplicable (Only the requested-return operations are covered here; independent pushes in those chains are separate units.); Included persisted goals, quota parks, autonomous/scheduled wake-up controls, fork transfer and child observe/message/recovery alternatives: uninspected (The memory inventory identifies these included alternatives but does not fully trace their retained parts, admission or replay. Their push selection classification remains uninspected; the known routes do not enumerate the whole boundary.); Opaque provider-side accumulated state, if any; fixed inference weights are not use-derived memory: uninspected (Provider caching, adaptation and parameter provenance are uninspected. No push selection value is inferred from ordinary remote inference or a mutable endpoint. External implementation is outside frozen evidence, preventing resolution of this included alternative.)	partial coverage (Three wired automatic trace-fed durable summary/edit routes support local yes findings. No local no is invented without bounded absence records. Other write inputs and later consumers remain unresolved; this is not demonstrated improvement.); Trajectory-fed refinement to durable harness entry to later foreground/child consumer: known — yes [wired] (Automatic software admission produces retained behavior-shaping content; no novelty, truth acceptance or measured capacity gain is required for this route classification.); Compaction summary to later model invocation: known — yes [wired] (Compression qualifies because it automatically transforms traces into retained continuation guidance; old raw history can also remain.); Optional abandoned-branch summary to later selected context: known — yes [wired] (Caller choosing summarize is not per-summary content approval; host persists the generated summary for later model context.); Direct caller edits, changed resources and initialization: not-determinable (These routes can write retained behavior-shaping content, but their original trace inputs are not established for every caller. Model authorship alone does not establish trace learning.); Namespace capture and arbitrary revived values: not-determinable (Known capture preserves current values rather than deriving lessons from recorded activity. Enclosed values and their producers are uninspected, so no system-wide or payload-wide negative is asserted.); Included persisted goals, quota parks, autonomous/scheduled wake-up controls, fork transfer and child observe/message/recovery alternatives: uninspected (The memory inventory identifies these included alternatives but does not fully trace their retained parts, admission or replay. Their qualifying trace-fed writes classification remains uninspected; the known routes do not enumerate the whole boundary.); Opaque provider-side accumulated state, if any; fixed inference weights are not use-derived memory: uninspected (Provider caching, adaptation and parameter provenance are uninspected. No qualifying trace-fed writes value is inferred from ordinary remote inference or a mutable endpoint. External implementation is outside frozen evidence, preventing resolution of this included alternative.)	partial coverage (Knowledge, instruction, learning, ranking, routing, validation and enforcement are attached to actual consumers. They do not establish compliance, factual acceptance or benefit. Unresolved payload and control consumers prevent complete authority coverage.); Memory-kind facts/preferences/decisions at foreground model: known — knowledge [wired], knowledge [afforded] (Digest gives advisory context; requested full detail is a separately afforded consumer path.); Prompt policies and skill explanations at foreground model: known — instruction [wired], instruction [afforded] (Standing supplemental directions are model-interpreted, not immutable base-prompt enforcement.); Stored subagent task content and metadata at factory/child consumers: known — instruction [wired], routing [wired] (Full prompt reaches the child; model/thinking selectors are consumed separately by host code.); Stored machine task templates and symbolic control at factory consumers: known — instruction [wired], routing [wired], enforcement [wired], validation [wired] (Task directions and executable transition constraints are distinct effects of retained specification parts.); Entry text/access fields at digest and requested-search rankers: known — ranking [wired], ranking [afforded] (Fixed lexical scoring still gives retained content ranking force.); Session messages and retained refinement history at update consumers: known — learning [wired], learning [wired] (Trace-fed guidance updates and continuation summaries consume retained experience. Expected outcome is not a measured benefit.); Conversation, summary, audit projection and namespace notice at later foreground model: known — knowledge [wired] (These supply context/reports, not authenticated truth; bounded projections omit much retained detail.); Compaction boundary, digest fingerprint and retained audit snapshots at host consumers: known — routing [wired] (Access/control metadata is separated from advisory content.); Changed skill body/program at model/kernel admission: known — instruction [afforded] (Whole installation route retains afforded evidence in light of the cross-report disagreement; arbitrary program effects are not blanket enforcement authority.); Opaque checkpoint values and executable payload effects: not-determinable (Restored values can alter later Python execution, but actual content and consumers are absent. Generic execution dependence does not determine a controlled authority for every enclosed value.); Included persisted goals, quota parks, autonomous/scheduled wake-up controls, fork transfer and child observe/message/recovery alternatives: uninspected (The memory inventory identifies these included alternatives but does not fully trace their retained parts, admission or replay. Their authority classification remains uninspected; the known routes do not enumerate the whole boundary.); Opaque provider-side accumulated state, if any; fixed inference weights are not use-derived memory: uninspected (Provider caching, adaptation and parameter provenance are uninspected. No authority value is inferred from ordinary remote inference or a mutable endpoint. External implementation is outside frozen evidence, preventing resolution of this included alternative.)	AAS-2026-10-14-prime-agent-8ec59ae12476-01
PrimeScientist	2	Accumulated experiment-tree plans, priors/reasons, packets/statistics, winning logs and diagnostic reductions, failure digests, baseline history and reusable workspaces; cross-task insight files and copies; reflector session access metadata and opaque resumed payload. Includes their automatic admission, transformation, withdrawal/reset and later driver/reflector/worker consumer routes, including native-skill loading uncertainty and arbitrary workspace data/model parts. Excludes shipped static task instructions/references, fixed reflector/search policy, transient scoring intermediates without a later consumer, the disabled temporary error-analysis helper, and external model weights/runtime internals except the included session/native-skill memory interfaces. No deployment, benefit or model-weight-learning assertion.	partial coverage (The operative file inventory is recorded; external session storage remains opaque. Workspace file storage does not resolve the encodings inside it.); Tree plans, statistics, proposal artifacts, outcome packets, raw logs and reductions, baseline history, retained workspaces, insight copies and session access metadata: known — files [wired] (These operative retained objects and staging copies use the filesystem; JSON and Markdown are encodings, not different substrates.); External reflector session payload selected through saved session ID: uninspected (The JSON access file is inspected, but neither client/provider payload storage nor retention internals are inspected. No substrate value is assigned to the payload.)	partial coverage (Software/model admission is automatic despite an operator launching the workflow. Initial raw proposals/insights and host payload writes retain unknown control; mere filesystem editability establishes no manual route.); Reflector plan admission, packet/statistics/artifact retention, brief/digest/history production and session-metadata updates: known — automatic [wired] (Inspected routes admit without per-content human approval, including score-gated workspace replacement.); Insight copying/loading after files exist: known — automatic [wired] (Known automatic copy is independent of unresolved initial content admission.); Initial cross-task insights and pre-existing raw proposal content admission: not-determinable (Original controller/approval is unresolved. Unknown caller identity cannot establish manual or automatic initial admission.); External session content writes and compaction; arbitrary workspace data/model admission: uninspected (Internal host and task-specific write mechanisms were not inspected; inspected metadata/copy operations do not classify them.)	partial coverage (Both requested filesystem reads and automatic subsequent-invocation supply are supported. External restoration and native-skill loading add unresolved operations; neither known direction covers them.); Reflector/worker explicitly requested artifact reads and driver state reloads: known — pull [wired], pull [afforded] (A consumer reading a named file fulfills its request; file availability alone is not another automatic delivery.); Automatic plan/history/insight and inherited-workspace supply to a new invocation: known — push [wired] (Code prepares selected retained inputs without a material request from the new worker.); External reflector session restoration and native-skill body insertion: partial — pull [wired] (Wrapper requests the selected session, but host window selection/restoration and implicit skill insertion are uninspected.)	partial coverage (Signals are local to automatic input preparation. Coarse filters and actual node/task identity matching are separate from unknown native-skill and host session selectors; no semantic insight retrieval is inferred.); FIRE-Bench insight supply at worker invocation: known — coarse [wired], identifier [wired] (Trigger is worker invocation; selector uses Markdown availability, nonempty size, sorting and current-task identity; selected parts are other-task insight text.); Baseline history supply at next iteration: known — coarse [wired] (Trigger is the next attempt; selector takes the last ten records and truncates summaries to 400 characters, then stages the selected history.); Selected node plan and parent/best workspace staging: known — coarse [wired], identifier [wired] (Trigger is driver-selected execution; ancestry/visibility and recorded score eligibility precede resolving the selected node/parent or best-workspace path into delivered retained parts. A catalog identifier alone would not suffice.); Native-skill selection/body insertion and external resumed-session window selection: uninspected (Implicit invocation configuration and saved session ID do not reveal host selectors, selected payload parts or restored budget. No signal is assigned to opaque content.)	partial coverage (Known automatic trace-fed reductions and control/guidance updates coexist with unresolved insight publication and opaque host/workspace updates. Yes means a durable later-consumed update path, not improved capacity or a weight change.); Log-brief, failure-digest and baseline-history writes for later model consumers: known — yes [wired] (Execution experience is automatically reduced into retained diagnostic/context guidance delivered to subsequent invocations.); Outcome-fed Q/visits updates and reflector successor-plan admission: known — yes [wired] (Evaluation feedback changes durable selection state and feeds the wired proposal-to-later-worker path; model evidence responsiveness is not observed.); Cross-task insight publication from completed attempts: not-determinable (Existing file delivery is wired but its original producer and admission are unresolved. Neither a closed learning loop nor a local no follows.); External session-derived host memory and task-specific workspace data/model updates: uninspected (Automatic trace-to-durable-artifact transformations inside those parts were not inspected; copied code or session access metadata alone cannot decide them.)	partial coverage (Authority follows each consumed part, not its filename. Model effects remain guidance/affordance rather than observed compliance; opaque payload and workspace consumers prevent an exhaustive authority inventory.); Selected node plans and change explanation at coding-worker guidance consumer: known — instruction [wired] (Standing invocation guidance is delivered through native skills or inline instructions; it is not procedural enforcement.); Diagnostic packets, brief, failure digest, prior reasons and insight text at reflector/worker reference consumers: known — knowledge [wired], knowledge [afforded] (These parts supply evidence or advice; score truth and model compliance are not guaranteed.); Baseline history at subsequent worker guidance consumer: known — knowledge [wired], instruction [wired] (Prior scores/actions and short summaries supply context, with a request to avoid failures and improve.); Retained outcome/diagnostic inputs at successor-plan and history update consumers: known — learning [wired] (Experience feeds durable behavior-shaping outputs; this authority does not assert improved capacity.); Q/P and admitted prior estimate at BAVT weighting consumer: known — ranking [wired] (Retained numbers influence relative sampling priority, not epistemic truth probability.); Packet presence/replay fields and session ID/counter at driver/continuation consumers: known — routing [wired] (Retained fields determine execute-versus-expand, replay, continuation identity and budget accounting.); Retained branch visibility and score state at selection/inheritance/keep-revert consumers: known — enforcement [wired], routing [wired] (The driver excludes hidden branches and conditions seed reuse/replacement on recorded scores. This veto is on later search use, not on earlier worker actions.); External session payload and task-specific workspace content at opaque downstream consumers: uninspected (The continuation/copy path is visible, but payload effects and arbitrary data/model consumers are not. Executable content alone does not justify a blanket instruction or learning authority for the entire workspace.)	AAS-2026-10-14-primescientist-f880686cf98a-01
quicky-wiki	2	Use-derived or edited retained source documents, SQLite claims/control/provenance/relations/events/FTS, page summaries, project configuration, optional saved answers and rendered/exported views; their admission, maintenance and later consumers, plus unresolved provider-retained use-derived state and embedding initialization. Excludes static shipped prompts/defaults, static model weights, transient extraction/diff/consensus/review arrays except as transformation inputs or outputs, package release artifacts and external host planning. Remaining export payload details and opaque alternatives stay included as unresolved.	partial coverage (The recorded inventory covers SQLite parts, retained files and configuration, but provider-retained alternatives remain opaque. Graph relations and FTS are SQLite structures, not additional physical substrates.); Claims, controls, summaries, source register, events, relations and FTS indexes: known — sqlite [wired] (All six retained graph parts and their access structures are accounted for.); Source files, saved answer sources, rendered views and edited project configuration: known — files [wired] (These operative retained objects are filesystem files; deployment contents are not supplied.); Opaque provider-retained state, if any, accumulated through use: uninspected (The substrate of any provider-retained use-derived state cannot be classified. Provider internals were not inspected; inference calls do not resolve provider retention or updates. Static model weights are not themselves classified as accumulated memory.)	partial coverage (Compiler admission is automatic despite user initiation; explicit operator choices are manual. Generic caller identities and provider updates remain unresolved.); Compiler claim/summary/control/page-link admission, event recording, decay and generated-view writes: known — automatic [wired] (No per-content human decision intervenes on these recorded transformations.); Explicit human setup overrides and opt-in answer-file save: known — manual [wired] (Save is a human decision to retain that answer, not per-claim factual approval or approval of subsequent extraction.); Dashboard/MCP/library content edits, external files and embedding initialization: not-determinable (Handlers accept caller content, but generic HTTP/API access does not identify an explicit human content decision or an autonomous producer; no blanket manual or automatic classification is warranted.); Opaque provider-retained state, if any, accumulated through use: uninspected (Provider admission/update control is uninspected. Provider internals were not inspected; inference calls do not resolve provider retention or updates. Static model weights are not themselves classified as accumulated memory.)	partial coverage (Requested query/client delivery is pull; automatic compilation-context selection and watched-file admission are push operations. These are distinct operations, not duplicate labels on one return.); Question/consensus, explicit page/search/entity/history/file-display requests and maintenance requests: known — pull [wired] (Requests for retained material lead to code selection and delivery; lexical selection within a requested answer does not create another push.); Automatic comparison, page assignment/summary and configuration guidance selection during ingestion: known — push [wired] (The source-ingest request is not a separate request by the receiving model for existing graph/configuration material.); Filesystem-event watcher supplying retained file content to extraction: known — push [wired] (This is push to the admission pipeline, not injection into an ongoing answering conversation.); Export files at external human/application readers: known — pull [afforded] (Affordance is not deployed wiring.); Opaque provider-retained state, if any, accumulated through use: uninspected (Provider-retained state delivery direction cannot be determined from inference calls. Provider internals were not inspected; inference calls do not resolve provider retention or updates. Static model weights are not themselves classified as accumulated memory.)	partial coverage (Signals classify push selectors only. Ordinary question lexical search is a pull selector, not inferred-lexical push. Opaque alternatives prevent declaring the entire signal inventory complete.); Automatic all-existing-statement and all-page-title supply to compiler models: known — coarse [wired] (Trigger is a new/changed source compilation; availability in the existing store is the selector input.); Automatic per-page summary context and entity-specific extraction guidance: known — identifier [wired] (Trigger is summary refresh or extraction. Page ID and resolved entity kind perform actual matches selecting delivered page claims or entityPrompts text.); Watched changed-file content delivered to compiler: known — identifier [wired] (Trigger is debounced filesystem change; event path identifies the actual file read. Startup unregistered-path and hash skip predicates additionally constrain processing.); Opaque provider-retained state, if any, accumulated through use: uninspected (Neither a push selector nor its input and selected opaque payload is recorded. Provider internals were not inspected; inference calls do not resolve provider retention or updates. Static model weights are not themselves classified as accumulated memory.)	partial coverage (A qualifying answer-trace route is wired after optional human retention, but opaque provider updates and unknown arbitrary source origins preclude a system-complete classification. Trace learning here means an automatic trace-fed durable update, not demonstrated improvement.); Saved question/answer document automatically re-extracted into graph for later questions: known — yes [wired] (Manual save does not make the later model/software claim-admission mechanism manual. Explicit ingest or active watch is required; activation is unobserved.); Reduced graph-mutation events and return-only critique/consensus outputs: not-determinable (Events alone are raw reduced traces, not an established learned derivative; review/consensus callers do not persist revisions. No qualifying automatic trace-fed transformation is established for these parts, but no bounded absence record supports a local no.); Arbitrary imported documents, initial embedding state and external caller updates: not-determinable (Unknown original content/control prevents enumerating further trace-fed learning routes; generic documents are not assumed to be session logs.); Opaque provider-retained state, if any, accumulated through use: uninspected (Provider-side trace-fed durable updates are not inspected. Provider internals were not inspected; inference calls do not resolve provider retention or updates. Static model weights are not themselves classified as accumulated memory.)	partial coverage (Authority follows actual retained parts and consumers, not the fact that host code enforces a write. Provider paths and external adoption remain unresolved.); Statements, scores, summaries and saved/raw documents at informational model/human consumers: known — knowledge [wired] (Delivered evidence and context are not verified truth or enforced obedience.); Statements and summaries in lexical retrieval; confidence/age in ordered review selection: known — ranking [wired] (Fixed scoring and ordering give retained content ranking force without learned ranking policy.); Source hashes, kind/metadata, IDs and graph edges at compiler/selector consumers: known — routing [wired] (Routing includes skip/extract, entity scope, page grouping and conditional cascade destinations.); Retained confidence at default slide/export eligibility checks: known — validation [wired], enforcement [wired] (Default slide eligibility is bounded: explicit topic selection bypasses the confidence floor.); Edited configuration entity prompts and symbolic runtime choices: known — instruction [wired], routing [wired] (Standing entity guidance is distinct from graph evidence.); Saved question/answer material at subsequent automatic extraction: known — learning [wired] (Update-consumer force is distinct from the resulting claims knowledge force.); Event history and rendered Markdown at shipped display consumers: known — knowledge [wired] (Stored reasons are inspectable, not ordinary answer inputs; dashboard reads files for display.); Exported content at external presentation/application readers: known — knowledge [afforded] (Only the documented/export-afforded informational consumer role is established.); Opaque provider-retained state, if any, accumulated through use: uninspected (No actual consumer/effect path for opaque provider-retained material can be classified. Provider internals were not inspected; inference calls do not resolve provider retention or updates. Static model weights are not themselves classified as accumulated memory.)	AAS-2026-10-15-quicky-wiki-628f7d7cc628-01
QwenPaw	2	Retained objects accumulated or changed through use: native session live/summary/control parts; Scroll SQLite history, continuation summary and eviction index; ReMe daily/session/digest notes, catalogs, cached embeddings and notifications; ADBPG retained service memory and local fallback; edited workspace prompt files; archive/replay artifacts; harness thread continuation; fork saved-context and retained project/registry paths. Opaque provider payloads, native dependency compaction, initial content and auxiliary governance/plugin/checkpoint/backup memory remain unresolved included alternatives. Auxiliary mail-thread/browser stores are included in the reports coverage limit, not inspected inventory witnesses. Excludes shipped unchanged instruction templates, inference/reranker model weights not shown to accumulate through use, ordinary current-run tool results without retention, and turn-local visual projection/recovery.	partial coverage (Local file, SQLite, vector-cache and external-service witnesses do not enumerate all included auxiliary stores or provider-owned payloads.); Daily/session notes and dream/paper digest files: known — files [wired] (Configured workspace Markdown/vault and local catalogs are file storage, not a graph substrate merely because they contain links.); Session JSON: live state, native summary, Scroll summary/index and pending extraction state: known — files [wired] (Session-directory files persist these separately produced parts.); Scroll history rows: known — sqlite [wired] (HistoryStore uses workspace history.db.); Native JSONL archives, debug replay files and saved large tool text: known — files [wired] (JSONL and UUID text artifacts are retained files.); Mutable workspace prompt bodies: known — files [wired] (Only edited/accumulated workspace bodies are included; packaged defaults alone are excluded.); Optional ReMe cached embeddings: partial — vector [wired] (Cached vectors are established, but complete cache backing and provider-side retained state are not exposed.); ADBPG retained memory service: partial — service-object [wired] (The external retained service object is known; database, embedding and other internal substrates cannot be inferred from its name.); ReMe inbox notifications: not-determinable (Application inbox retention is identified, but its storage implementation is uninspected.); Alternative harness retained threads and session bridge: partial — service-object [wired] (Thread resume and persisted mappings are wired; complete provider storage is opaque.); Fork project product retained for integration: known — repo [wired] (Git commits/branches and worktrees persist as project products; registry backing is a separate unresolved part.); Fork registry backing, governance reuse, plugin-owned memory and checkpoint/backup payload storage: uninspected (Records expose administrative retention or restoration but not a complete storage inventory.)	partial coverage (Automatic admission is supported on multiple routes. Generic API/command callers and initialization do not establish explicit human content decisions, so no blanket manual classification follows.); ReMe extraction, dream and Daily Paper computational admissions: known — automatic [wired] (Software/model selects generated content without per-content human veto; initiating a job or /memorize is not reviewing its output.); Scroll archival, index generation, summary replacement and quality gate: known — automatic [wired] (Code/model admission includes deterministic summary rejection; no per-summary human approval is shown.); ADBPG user-message forwarding: known — automatic [wired] (Software chooses unseen user messages and schedules add; human authorship is not review of remote derived facts.); Normal native session snapshot admission: known — automatic [wired] (Post-response hooks serialize agent/mode state without a per-snapshot content decision.); Tool-text archive and inbox projection admission: known — automatic [wired] (Threshold pruning and notification gates are computational admission mechanisms.); Fork dirty-worktree fallback commit and registry transition: known — automatic [wired] (Finalizer stages/commits worker changes and records administrative state without a human content gate.); Managed memory/workspace PUT, explicit debug replay and checkpoint replacement: not-determinable (Generic caller selection/replacement is inspected, but explicit human approval versus software admission is unresolved for these mechanisms.); Initial retained contents, provider-owned thread admission and auxiliary writes: uninspected (Known subsequent automatic writes do not resolve all initialization or provider/auxiliary admissions.)	partial coverage (Explicit requested return and unsolicited context selection coexist. Outer harness resume is known without complete provider-owned read-back coverage.); Foreground memory search and file expansion: known — pull [wired] (Foreground query requests retained snippets; guidance affords subsequent file expansion. The return itself is not an additional push.); Structured history recall and explicit debug replay: known — pull [wired] (Model/caller requests rows or a selected saved checkpoint for later context.); Automatic new-user-turn memory search: known — push [wired] (Middleware selects and injects retained snippets without a foreground tool request.); Native session restore and Scroll background state: known — push [wired] (Code restores selected retained state and renders summary/index into later calls.); Configured mutable workspace prompt contribution: known — push [wired] (Builder supplies selected edited bodies at future construction without the answering model requesting them.); Idle proactive saved-session context: known — push [wired] (Idle/cooldown/busy checks trigger automatic supply of cleaned retained session text to a separate agent.); Forked saved-parent context delivery: known — push [wired] (Code gives child a saved-session snapshot, not the latest evolving parent window.); Alternative harness resume: partial — push [wired] (QwenPaw selects saved thread identity for backend continuation; provider internal selectors/returns remain opaque.); Interests/catalog, inbox subsequent model adoption and auxiliary retained-store consumers: not-determinable (Display does not establish automatic model adoption; complete auxiliary consumer routes are not inspected.)	partial coverage (Signals apply only to automatic selection operations. Explicit search/recall returns stay pull; opaque ReMe/provider selectors and incomplete proactive predicates prevent complete signal coverage.); Native session restore on later request: known — identifier [wired] (Trigger: native build. Selector: session/user/channel-derived checkpoint identity. Selected part: matching saved state.); Scroll background summary/index/tail availability: known — coarse [wired] (Trigger: later model call. Selector: current retained checkpoint, validity and render budgets. Selected part: surviving summary/index/raw tail; no query-specific identity match is established for this rendering.); Configured mutable prompt-file contribution: known — identifier [wired] (Trigger: agent build. Configured filenames actually match the selected retained bodies; empty/missing filtering is a separate availability condition.); New-user-turn automatic ReMe retrieval: partial — inferred-lexical [wired], inferred-embedding [wired], inferred-judgment [wired] (Trigger: new external user turn. Latest user text queries daily/digest corpus; keyword and optional embedding retrieval are configured, while complete dependency selection is opaque.); Automatic ADBPG local Markdown fallback: known — inferred-lexical [wired] (Trigger: automatic memory search. Selector: user query keywords scored against MEMORY.md/direct memory paragraphs. Selected part: capped snippets.); ADBPG remote search selector: not-determinable (Query and score filtering are visible, but remote lexical/embedding/judgment candidate selection is not exposed.); Idle proactive context selection: partial — coarse [wired] (Trigger: idle monitor. Recent-session filtering and 100-message/50,000-character budgets supply cleaned text; exact recency predicate is uninspected.); Fork saved-parent snapshot selection: known — identifier [wired] (Trigger: fork dispatch. Parent session identity selects the saved session file copied to child; missing/unreadable file may yield empty state.); Harness saved-thread mapping: partial — identifier [wired] (Trigger: later backend turn. Outer mapping matches session to thread, but SDK-internal context selection is opaque.); Digest metadata and auxiliary push alternatives: not-determinable (Missing consumer/selector facts prevent inapplicability or a complete signal union.)	partial coverage (Automatic session-derived note and continuation-summary writes have later consumers. This is architectural trace learning, not demonstrated improvement; native/remote/auxiliary alternatives remain unresolved.); Conversation batches to ReMe notes to later search/model context: known — yes [wired] (Automatic trace-fed generated notes enter a wired durable behavior-shaping route; backend outputs and benefit are unobserved.); Archived turns/results to Scroll continuation summary to later context: known — yes [wired] (Automatic model summary is checkpointed and later rendered; reshaping qualifies without new claims or measured improvement.); ADBPG remote trace transformation: not-determinable (Automatic user-message forwarding and later text consumption are wired, but remote retained artifact/transformation semantics are opaque; no local negative follows.); Dream derived-note maintenance: not-determinable (Retained-note input and future search are declared, but exact trace lineage and backend update semantics remain unresolved.); Native compressed summary including debug replay alternative: uninspected (Summary persistence/replay is identified, but dependency-owned generation and full trace-fed admission chain are not inspected.); Daily Paper resource-fed route: inapplicable (Original inputs are external paper resources, not established activity traces from agent work; resource analysis alone is not trace learning.); Raw history capture and unchanged replay alone: inapplicable (These units record/recover raw evidence; their separate derived-summary route is classified above. Raw logging alone is not a qualifying derived update.); Harness-owned updates, fork products, initialization and auxiliary adaptation: uninspected (Records do not expose every automatic trace-fed artifact or later adaptation consumer.)	partial coverage (Authority follows retained part and actual consumer. Supported update, retrieval, guidance and administrative paths do not establish all interests, remote-payload or auxiliary consumers.); Retained notes and raw context at later answering/proactive models: known — knowledge [wired] (Retrieved text, restored context and cleaned saved-session text supply evidence/background to later inference.); Edited workspace prose at native system-prompt composition: known — instruction [wired] (Selected retained bodies supply standing directions; delivery does not prove compliance.); Session/tool traces at ReMe extraction and Scroll summary update consumers: known — learning [wired] (Retained experience feeds durable behavior-shaping notes or continuation summaries; this update force is distinct from downstream knowledge delivery.); Retained note snippets at lexical search and optional reranker: known — ranking [wired] (Note text influences relative retrieval priority even when scoring code is fixed.); Scroll source ranges and recovery IDs at reconciliation/recall: known — routing [wired], validation [wired], enforcement [wired] (Retained IDs/ranges select recovery scope and source-existence checks; missing provenance endpoints can withdraw summary reliance.); Pending-turn markers/snapshots at extraction batch selection: known — routing [wired] (Retained batching state selects the next extraction inputs, not their factual endorsement.); Fork registry status/head/scope at repeated finalization: known — routing [wired], validation [wired], enforcement [wired] (Registry values select reuse/transition and enforce scope/active-state eligibility.); Inbox notification body at operator display: known — knowledge [wired] (Display supplies information to the operator; no automatic model adoption is established.); Digest interests/catalog downstream effects, opaque ADBPG/harness payload consumers and auxiliary policy reuse: not-determinable (Known retrieval does not resolve every stored part or consumer; configured interests are not proven proactive instructions.)	AAS-2026-10-15-qwenpaw-e41495d47fc0-01
ReasoningBank	2	Retained WebArena and SWE-Bench advice banks and row metadata, query-vector caches, saved task trajectories and feedback used by later invocations, and web delivery text; ordinary admission/retrieval, successful AWM/Synapse alternatives, scaling and standalone induction. Initial nonempty bank/cache provenance and admission remain unresolved. Scaling output compatibility and standalone launch/later-consumer wiring remain included uncertainties. Ancillary retained-memory routes in other runners/adapters remain uninspected, not evidenced absent. Excludes static shipped prompts, provider/model weights not updated by these memory routes, ordinary episode-local accumulation and view truncation without a retained summary, and optional response memoization with no established shipped activation. No observed operation or benefit is asserted.	partial coverage (Files and RAM are supported locally; ancillary coverage prevents a system-complete inventory. Dense vectors stored in JSONL do not establish a vector database.); Ordinary bank content and ledger, query cache, delivery text, saved trajectories and web feedback: known — files [wired] (The records inventory the operative retained objects; initial content uncertainty does not obscure their inspected file-backed storage.); Loaded cache tensors and retained feedback's host-memory consumption view: known — in-memory [wired] (RAM is an access/consumption part of the file-backed memory chain.); Scaling and standalone intended output banks: known — files [afforded] (Storage destination is classified at affordance strength, independently of reachability and output-shape defects.); Ancillary memory paths in other runners/adapters: uninspected (The accepted runtime bounds inspection to traced drivers; alternative implementations prevent a complete substrate inventory.)	partial coverage (Automatic subsequent admissions are supported without resolving external initialization, web-host persistence internals or ancillary writes.); Ordinary advice admission, cache creation/append, bank metadata, feedback persistence and delivery-file regeneration: known — automatic [wired] (Generation authorship and physical filesystem I/O are kept separate from who admits content.); Coding trajectory saver and shipped consumption of saved web trajectory artifacts: partial — automatic [wired] (Installed BrowserGym's full persistence implementation is outside captured code; the complete web-artifact admission mechanism cannot be classified from its later reader alone.); Scaling and standalone advice admission bodies: known — automatic [afforded] (CLI path selection supplies configuration, not a decision approving the retained content.); Initial nonempty bank/cache contents and external edits: not-determinable (The records do not identify a content-supplying human decision or a software admission path for externally prepared or changed material. Manual is not inferred.); Ancillary retained-memory admissions: uninspected (Alternative entry points are not exhaustively inspected.)	partial coverage (Both request-side pull and automatic task-model supply push exist in the ordinary chain; unresolved alternatives prevent complete coverage.); Runner requests to bank/cache selectors and explicit readers of retained trajectories/feedback: known — pull [wired] (Pull describes requests within the chain, not whether the foreground task model asked to receive memory.); Automatic prior-advice selection and system-message supply during ordinary task setup: known — push [wired] (The user requested task work, not a memory lookup by the task model. This local supply operation coexists with the runner's selector pull.); Standalone output and scaling payload reaching a later consumer: not-determinable (Standalone destination has no established later consumer; scaling's invocation/return-shape defects prevent establishing delivery of compatible retained content.); Ancillary memory read-back: uninspected (Alternative runners/adapters prevent complete direction coverage.)	partial coverage (Embedding inference and an operative identifier match are wired witnesses; unresolved read-back branches prevent a complete signal inventory.); Automatic web/coding advice supply: task intent/problem to prior cached queries: known — inferred-embedding [wired], identifier [wired] (Embedding priority and identity join are distinct selector operations in the same push chain. Availability checks and n=1 do not add coarse targeting.); Standalone/scaling and ancillary push alternatives: not-determinable (Later delivery is unresolved for standalone/scaling and ancillary routes are uninspected. Their signals cannot be inferred from the ordinary selector or declared inapplicable.)	partial coverage (A positive ordinary-loop witness does not resolve other write-to-consumer chains; no unsupported local or system-wide negative is inferred.); Ordinary web and coding automatic trajectory-to-advice writes and later task-model consumption: known — yes [wired] (The consumer chain is wired; actual learning as improved capacity and downstream causal dependence remain unobserved.); Scaling/standalone writes, initial external content and ancillary automatic updates: not-determinable (Scaling lacks complete executable compatible production, standalone lacks a later consumer, and external/ancillary update chains remain unresolved. Raw persistence or query caching alone does not establish a qualifying trace-fed update.)	partial coverage (Separate downstream instruction/reference, update learning, numerical ranking and symbolic routing paths. No retained semantic acceptance criterion or advice execution enforcement is established; that is not a system-wide absence claim.); Selected ordinary advice consumed by web/coding task models through system messages: known — instruction [wired] (Channel and advisory force are established; compliance and benefit are not.); Synapse demonstrations at the task model and saved trajectory evidence at evaluators/inspectors: known — knowledge [wired] (This is the evidence/reference consumption path, not the downstream procedural direction path.); Retained trajectories and web explanation at ordinary induction consumers: known — learning [wired] (Coding strips system messages and does not forward judge rationale; the updater's actual view is narrower than the saved trace.); Prior cached vectors at similarity scoring: known — ranking [wired] (Stored query prose is not itself rescored after load; advice content and truth do not determine this ranking.); Selected cache/bank IDs at lookup and saved web outcome labels at induction prompt choice: known — routing [wired] (Bank status is not a retrieval filter; unconsumed template/status metadata is not assigned authority merely because it is stored.); Scaling and standalone experience at intended advice updaters: known — learning [afforded] (This local update-consumer authority does not establish downstream behavior dependence.); Scaling payload and standalone output at later task consumers: not-determinable (Scaling's return contract is incompatible/unresolved; standalone has no wired destination consumer. Intended advice cannot establish actual consumed force.); Ancillary memory consumers: uninspected (Other runners/adapters prevent a complete authority inventory.)	AAS-2026-10-14-reasoning-bank-f0bedcab086c-01
Reflexion	2	Included: ALFWorld/WebShop retained task plans, experience logs and configuration controls; HotPotQA diagnoses, prior-attempt excerpts, logs and whole-agent checkpoints; programming retained candidates, feedback, hints, UCS parent state and JSONL products, resume and offline execution consumers; unresolved search/grading internals and possible provider-side accumulated state. Excluded: shipped static prompts/examples, ordinary current-attempt windows and fixed pretrained weights as such; live external source contents and unrelated repository housekeeping.	partial coverage (Files and in-memory objects cover the inspected parent memory inventory. Missing UCS internals and opaque provider-side accumulated state prevent a complete inventory; repository packaging and pretrained weights are not counted as memory substrates.); ALFWorld and WebShop plan lists, configuration flags and trial/world logs: known — in-memory [wired], files [wired] (The cited records cover this named part and its consumer.); HotPotQA diagnosis and prior-attempt text in agent state; serialization and independently constructed logs: known — in-memory [wired], files [wired] (The cited records cover this named part and its consumer.); Programming candidates, hints, feedback, UCS nodes and result rows: known — in-memory [wired], files [wired] (The cited records cover this named part and its consumer.); Additional retention inside unavailable UCS search: not-determinable (Parent State objects are inspected; search-node retention, eviction and additional substrate are not. This prevents complete UCS storage classification.); Possible provider-side state accumulated through use, separate from shipped pretrained model weights: not-determinable (Generation adapters do not expose whether providers retain additional use-dependent state or its substrate. Pretrained weights alone do not establish accumulated memory.)	partial coverage (Subsequent host/model admission is automatic. Initial content, resumed-payload control, unavailable search admission and provider writes remain unresolved; human-edit renderer labels and workflow launch are not manual-content witnesses.); Failure-triggered plan/diagnosis admission and HotPotQA prior-trace replacement: known — automatic [wired] (The cited records cover this named part and its consumer.); Programming candidate replacement, feedback/hint append, UCS parent-node construction and result/checkpoint saving: known — automatic [wired] (The cited records cover this named part and its consumer.); Initial dataset and caller-selected resumed configuration content admission: not-determinable (Generic caller file selection does not establish explicit human approval/editing of retained content; initial authorship and admission control remain unresolved. No manual finding is licensed.); Additional UCS search retention/admission: not-determinable (The parent creates State values automatically, but missing search implementation prevents classifying additional node admission mechanisms.); Possible provider-side state accumulated through use, separate from shipped pretrained model weights: not-determinable (Provider-side use-dependent write and admission control are not visible in generation adapters.)	partial coverage (Requested configuration/file acquisition and automatic later selection are separate operations. Both pull and push are supported; checkpoint reload, external search, defective remote audit and provider routes retain local uncertainty.); Driver/agent retry prompt assembly and updater scenario delivery: known — push [wired] (The cited records cover this named part and its consumer.); ALFWorld/WebShop explicit resume and operator-selected standalone programming result read: known — pull [wired] (The cited records cover this named part and its consumer.); Standalone validator solved-row selection after requested file acquisition: known — push [wired] (The cited records cover this named part and its consumer.); UCS visible state expansion delivery: partial — push [wired] (Parent callbacks are wired; external scheduler invocation and further selection routes remain uninspected.); Intended LeetCode audit file acquisition and candidate delivery: partial — pull [afforded], push [afforded] (Entry-point defects block executable delivery; the scaffold does not establish wired remote read-back.); HotPotQA saved-agent reload: not-determinable (The writer was inspected but no saved-agent reload caller was established. Retention alone cannot establish either direction or a bounded absence.); Possible provider-side state accumulated through use, separate from shipped pretrained model weights: not-determinable (No later-consumer route for possible provider-accumulated memory is exposed.)	partial coverage (Identifier matching is supported for indexed environment/session configurations, while recency, strategy, outcome and budget rules are coarse. No embedding, lexical or model-judgment relevance selector is established; unresolved routes prevent complete coverage or blanket inapplicability.); ALFWorld/WebShop per-task plan and scenario push: known — identifier [wired], coarse [wired] (The cited records cover this named part and its consumer.); HotPotQA agent-owned retained strategy material: known — coarse [wired] (The cited records cover this named part and its consumer.); Ordinary/immediate programming retry and standalone solved-row execution: known — coarse [wired] (The cited records cover this named part and its consumer.); UCS state-specific expansion and unavailable search selection: partial — coarse [wired] (Missing search code prevents classifying its actual candidate selector. No identifier-based or inferred selector is inferred from the node ID API.); Intended LeetCode implementation iteration: partial — coarse [afforded] (Parent iteration is visible but executable delivery and remote internals remain unresolved.); HotPotQA checkpoint read-back selection: not-determinable (No reload route is established, so no push selector can be classified and pull-only inapplicability is not justified.); Possible provider-side state accumulated through use, separate from shipped pretrained model weights: not-determinable (No provider-side push trigger, selector input or selected retained part is inspectable.)	partial coverage (Qualifying automatic retained-guidance chains support yes without claiming improved capacity or exercised learning. WebShop has a later reflection consumer, not action-plan delivery. Candidate-only, checkpoint, search and provider alternatives remain unresolved, so no system-wide no is authored.); ALFWorld failure-log-to-plan-to-later-action/update chain: known — yes [wired] (The cited records cover this named part and its consumer.); WebShop failure-log-to-plan-to-later-reflection chain: known — yes [wired] (The cited records cover this named part and its consumer.); HotPotQA scratchpad-to-diagnosis or retained last-attempt excerpt-to-next-call chain: known — yes [wired] (The cited records cover this named part and its consumer.); Ordinary Python feedback-fed hints and immediate-refinement feedback-fed candidate replacement: known — yes [wired] (The cited records cover this named part and its consumer.); Immediate-reflexion with omitted execution feedback: not-determinable (The recorded derivation is candidate-fed, not executed-test-trace-fed. Records do not provide a bounded qualifying-trace absence warrant, so no local no or tool-trace positive is assigned.); UCS visible feedback-fed successor production: partial — yes [wired] (Visible Python callback chain is wired; missing search scheduling, horizon and additional write routes prevent complete assessment.); Whole-object checkpoint extras and standalone audit artifacts: not-determinable (Saving alone does not establish an additional behavior-shaping trace-fed write/later consumer. No audit result writeback is implemented in the named validator, but this does not warrant a whole-boundary no.); Possible provider-side state accumulated through use, separate from shipped pretrained model weights: not-determinable (Possible provider trace-fed adaptation or retained continuation state is inaccessible; generation-only client behavior is not a negative warrant.)	partial coverage (Authority is consumer-relative: advice, update input, fixed host controls, ranking inputs and executable products have distinct force. No semantic compliance or accepted explanation is inferred. Unavailable search/grading consumers and opaque payloads prevent complete coverage.); ALFWorld action/reflection models, WebShop reflection model and HotPotQA/programming successor generators consuming advice: known — knowledge [wired] (The cited records cover this named part and its consumer.); Failure logs, prior scratchpads, candidate feedback and past plans at automatic guidance producers: known — learning [wired] (The cited records cover this named part and its consumer.); Configuration success/skip flags and positional result-file completion accounting: known — routing [wired], enforcement [wired] (The cited records cover this named part and its consumer.); UCS retained outcome tuple/code at visible cost and successor callbacks: partial — ranking [wired], knowledge [wired] (Parent callback effects are wired in the Python-scoped account; missing search scheduling prevents complete consumer and horizon classification.); Persisted solution and tests at standalone Python/Rust validators: known — instruction [wired], validation [wired], routing [wired] (The cited records cover this named part and its consumer.); HotPotQA serialized library-object parts beyond identified text/state: not-determinable (No saved-agent reload consumer was inspected and full payload composition is unresolved; serialization alone supplies no additional downstream authority.); Intended offline LeetCode persisted-implementation consumer: partial — instruction [afforded] (The scaffold describes a consumer but does not wire an end-to-end submission; grader semantics remain unavailable.); Possible provider-side state accumulated through use, separate from shipped pretrained model weights: not-determinable (Possible provider memory consumers and effects are not exposed; known model generation cannot assign authority to an unknown retained part.)	AAS-2026-10-14-reflexion-19159b69b712-01
ReframeWeb	2	Included: retained graph guidance/preferences, sessions/messages/locators, instruction/task links and task/provider/startup configuration; tool histories and derived later-invocation summaries; directory/checkpoint locators, retained workspace/recovery bytes, image evidence and private startup-result evidence, including unresolved provenance, payload and later-consumer alternatives. Excluded: shipped static prompts/instructions before use-time adoption, transient orchestration/attempt controls as independent durable memory, provider model weights and hidden Codex/provider compaction, arbitrary native Store/UI internal state without demonstrated agent-memory delivery, and benchmark metrics as an automatic production updater. External exclusions do not resolve unknown included initialization or native maintenance.	partial coverage (Graph content, local/remote database adaptation, checkpoint/recovery storage and evidence files are classified. The remote database backing and mutable-directory contents prevent a complete physical-substrate inventory.); Guidance, preferences, session/conversation content, instruction bodies, filesystem locators, task receipts and saved Store/task configuration in the graph: known — graph [wired] (These are operative graph nodes and relations, not the checkpoint bytes they identify.); Configured local SurrealKV backing for graph context collections: known — kv [wired] (This is the local backing alternative; the graph organization remains a separate finding.); Configured remote graph database physical backing: not-determinable (The adapter establishes remote graph storage, but no accepted record establishes its underlying physical substrate. This does not add vector or service-object storage.); Checkpoint and recovery backing store: known — sqlite [wired], files [wired] (Native Store layout identifies both metadata database and payload/session directories.); Derived summary retained between producer and later completion/retry invocation: known — in-memory [wired] (The later-consumer summary occupies current BAML values, not a demonstrated cross-session summary index.); Screenshot originals/previews and private candidate result.json: known — files [wired] (Both records identify filesystem retention independently of graph locators.); Registered mutable directory payload backing: not-determinable (A directory path is wired as a materialization source, but the records do not classify every backing implementation or contained artifact.)	partial coverage (Known computational admissions and exact-content browser UI decisions coexist with unspecified generic API/script callers and unresolved initialization. Workflow initiation does not settle those mechanisms.); Reviewed candidate guidance admission: known — automatic [wired] (Models propose/keep; code intersects IDs and persists without a per-content human decision.); Human/validation reply recording, tool outcome recording, history summarization, image persistence and startup-result generation: known — automatic [wired] (Host/model production and admission is not replaced by human approval of each recorded object.); Human browser UI exact-content edits and confirmed deletion: known — manual [wired] (The record distinguishes this UI branch from arbitrary HTTP callers; no interaction was observed.); Schema-driven locator visibility update: known — automatic [wired] (Automatic migration is distinct from caller-controlled graph import.); Direct browser API edits, assigned Store instruction edits, project locator/save calls, graph import and Store/application publication: not-determinable (Generic caller identity and invocation alone do not establish explicit human content approval or exclude it. Their guards are known, but per-content admission control is unresolved.); Initial preferences, guidance, arbitrary sessions, instructions and directory/checkpoint source content: not-determinable (Known later writers do not settle initialization agency.)	partial coverage (Requested history/graph retrieval and automatic memory supply coexist within chains. Private result-file later use and arbitrary workspace/provider consumers remain unresolved, so known routes are not a complete read-back inventory.); Graph search and history rendering requests: known — pull [wired] (A consumer requests stored material by search plan or task history ID; fulfilling that request is pull.); Automatic cycle context reload: known — push [wired] (Cycle start supplies collections and acceptance-bound session/conversation context without an individual memory request.); Relevance selection within requested graph retrieval: known — push [wired] (The relevance model automatically selects retained candidates for composition; the upstream requested retrieval remains pull.); Task-linked instructions, selected project files, retry workspace and startup artifact supply: known — push [wired] (Task/startup triggers select associated retained material rather than a new consumer query for every delivered part.); Derived completion summary and newest-eight retained image supply: known — push [wired] (Later completion/visual model invocations receive selected retained representations automatically.); Private test result.json, arbitrary workspace payloads and benchmark-report later consumers: not-determinable (Return/storage capability alone does not establish all later invocation routes.)	partial coverage (Each supported signal names a push selector. Search requests are pull even when lexical hints determine the requested results; relevance-model selection supplies a separate inferred-judgment push. Other retained evidence/payload delivery remains unresolved.); Cycle-start guidance/preference collection supply: known — coarse [wired] (Trigger is a BAML cycle; selector input is collection membership; selected parts are all returned entries. Timestamp ordering is not a numerical budget.); Cycle-start session/conversation supply: known — identifier [wired], coarse [wired] (Trigger is context loading; acceptance-bound IDs select the current records, with relevance flags and role filters further bounding parts.); Relevance-model candidate selection for composer: known — inferred-judgment [wired] (Trigger is retrieval completion; model examines request plus candidate text/metadata and keeps candidate IDs. IDs encode its decision rather than replacing semantic judgment.); Task/selected-Store associations and next-launch startup selection: known — identifier [wired] (Task execution/preparation and Host launch trigger actual link/reference resolution. Selected identity matches determine delivered instructions, checkpoint sources and executable startup artifacts.); Completion summary delivery and image attachment window: known — coarse [wired] (Attempt completion supplies its derived summary; before next model turn completed observations/image MIME and newest-eight availability filter selected originals/previews.); Other workspace payload/evidence-report delivery selectors: not-determinable (The records do not establish a later selector for all included retained payloads; no pull-only inapplicability follows.)	partial coverage (Qualifying reviewed decision guidance and within-request retained summary supply are wired, not observed improvement. Image/log retention alone is insufficient. Generic locator control, initial content and arbitrary payload updates prevent a complete positive/negative inventory.); Decision-layer observation to reviewed durable guidance to later decision prompt: known — yes [wired] (The memory annotation explicitly establishes this qualifying automatic trace-fed guidance architecture. No accepted factual generalization or improved capacity is implied.); Recorded action history to retained completion/retry summary: known — yes [wired] (The summary replaces raw retained action history at a later consumer invocation within the request. Its attained retention horizon is that later invocation, not a demonstrated cross-session summary index.); Verified/saved locator admission and later relevance consumption: not-determinable (Activity/provenance is retained and later selected, but per-content human control and whether this is a qualifying automatic behavior-shaping update are unresolved.); Initial guidance/preferences/session content and arbitrary checkpoint/instruction update histories: not-determinable (No complete automatic trace-fed derivation is established for these alternatives. Known instruction delivery does not supply the missing update provenance.); Private startup test evidence and retained screenshots beyond evidence supply: not-determinable (These establish evidence retention and advisory context, not a recorded automatic trace-to-derived behavior-shaping update for every evidence part. No bounded absence record supports a local no.)	partial coverage (Authority follows the consumed retained part, not a uniform memory role. Retained tool history has learning force at summary production; transient guidance proposals do not establish a retained-observation learning consumer. Advice, standing directions, selector inputs and saved controls differ. Guidance-update input authority, opaque payloads and durable test-report consumers remain incompletely resolved.); Guidance text at later decision-layer model prompts: known — instruction [wired] (Reusable decision guidance provides standing directions, without guaranteed compliance.); Retained tool history at the action-summary update consumer: known — learning [wired] (This learning authority belongs to the recorded-history input at summary production, not transient guidance candidates. The summary's later knowledge authority is classified separately; cross-session retention and improved capacity are not established.); Decision observations as retained inputs at the guidance-update consumer: not-determinable (Candidates are transient staged groups; the records establish no retained decision-observation input delivered to review. Stored guidance is supplied, but no separately supported learning force for it is established. The trace-learning write positive does not establish this input authority.); Conversation/session/preference context and tool history at later relevance, composition and summary models: known — knowledge [wired], ranking [wired] (Context and evidence are distinct from enforced task permissions.); Linked instruction bodies at direct task execution and Codex preparation: known — instruction [wired] (Prose interpretation is not permission enforcement.); Graph identities, relevance flags and filesystem dependencies at host selection/materialization: known — routing [wired] (The consumed fields select memory input and preparation sources.); Saved task tool grants and Store activation/reference configuration at dispatch and guest checks: known — routing [wired], enforcement [wired], validation [wired] (Only retained grant/configuration parts are included; transient attempt state is not a new durable corpus.); Stored instruction body at guarded instruction replacement: known — validation [wired], enforcement [wired] (The expected-text comparison uses retained content as an admission criterion, independently of downstream instruction force.); Retained source files supplied as editable task workspace inputs: known — knowledge [wired] (This covers inspected materialization for later task work, not every opaque payload or startup executable effect.); Retained image evidence at subsequent model turns: known — knowledge [wired] (Rendering is advisory visual evidence, not a universal success oracle.); Derived action summary at completion/retry consumers: known — knowledge [wired] (The summary supplies outcome context; completion verdicts are separate consumers and not retained-summary enforcement.); Arbitrary checkpoint/directory payloads and private test/report evidence beyond their immediate caller: not-determinable (Executable startup and model-visible instructions are classified above. Other payload effects and mandatory later production report consumption are not established.)	AAS-2026-10-14-reframeweb-cff918d7eb1e-01
REM	2	Retained episode descriptions, parser metadata, envelope/access feedback, reusable embedding cache, episode vectors and temporal topology; semantic SQL facts/provenance/lifecycle fields, Qdrant vectors/payload and graph storage capabilities, their proposed automatic admission and known later read routes. Includes generic caller provenance and initial semantic population as unresolved alternatives. Excludes unchanged provider weights, static shipped prompts/code, auth/config/queue plumbing, current request-local candidate objects as retained memory, benchmark reporting files and the enclosing host context window. Candidate computations remain relevant only to proposed retained writes and bounded nonretention findings.	known coverage (The record inventory covers retained SQL, vector/cache and graph parts, including intended semantic representations. Provider weights, fixed source prompts, queues, request-local candidates and benchmark report files are outside the memory scope; unknown contents do not make the named substrates unknown.); Episode raw text, parsed interpretation, envelope and access feedback in PostgreSQL: known — rdbms [wired] (The named episode parts share one SQL substrate; store success remains conditional and unobserved.); Episode search vectors and cached episode/query embeddings: known — vector [wired], kv [wired] (Dense encoding is not the substrate. These two retained copies have distinct consumers.); Episode temporal identity/topology in Neo4j: known — graph [wired] (The retained part is topology, not SQL raw content.); Semantic SQL representation and Qdrant payload/vector representation: known — rdbms [afforded], vector [afforded] (Substrate coverage is resolved despite unknown initial population; this does not assert synchronized SQL/vector state.); Semantic compression nodes and edges: known — graph [afforded] (This is a library affordance rather than populated semantic provenance.)	partial coverage (Automatic mechanisms are supported without converting human authorship or workflow launch into manual admission. Generic supply and initial semantic population remain unresolved; no manual positive is established.); Structural episode admission, host exchange serialization and synthetic driver submission: known — automatic [wired] (This classifies admission, not the original author of the supplied text.); Generic caller supplying or replacing initial episode content: not-determinable (The caller may be a person or software; there is no recorded explicit human content decision supporting manual agency.); Parse-field admission, embedding cache/vector writes, graph derivation and access-counter updates: known — automatic [wired] (Wired conditional software mechanisms remain automatic even when dependency failure prevents a particular write.); Worker-proposed semantic admission: known — automatic [afforded] (Agency of the proposed mechanism is known; this is not a successful semantic write.); Initial or administrative semantic SQL, Qdrant and graph population: not-determinable (No admission principal or human approval/replacement decision is supplied for externally present contents.)	known coverage (The memory report inventories query retrieval, chronological/identity reload, host hooks, maintenance fetches and cache/feedback reuse. All inspected retained-material deliveries fulfill requests. Disconnected semantic writes and unknown initial provenance do not make the implemented read direction unknown; request-local candidates and result reports are excluded.); Query retrieval of episode context, semantic payload and their selection structures: known — pull [wired] (Automatic selection inside requested retrieval is not a separate push. SDK failures limit one delivery path, not the direct HTTP pull.); SQL list/get requests by direct clients, maintenance and benchmark clients: known — pull [wired] (Scheduled maintenance requests episodes rather than supplying unsolicited retained context. Nonempty SDK typed reload fails.); Embedding cache reuse and access-feedback reload within a subsequent requested operation: known — pull [wired] (The inventory includes access state as well as displayed content; no additional unsolicited delivery is recorded.)	inapplicable coverage (No included read-back unit has an unsolicited retained-material delivery operation. Signals classify push only; identifier filtering and embedding ranking on pull do not supply identifier or inferred-embedding positives.); All included query, list/get, maintenance-fetch and internal cache/feedback read-back operations: inapplicable (The completely assessed direction inventory is pull-only. Query embeddings, agent IDs and chronological limits are selectors within requested returns, not signals for a push.)	partial coverage (Wired session-derived metadata/vector learning and event-fed ranking feedback coexist with bounded local no and a claimed semantic loop. Unknown trace origin and initial semantic writes keep coverage partial; local no is retained even though the derived system union has yes. Raw logging without a derived write is outside the qualifying update inventory.); Recorded host exchanges automatically transformed into retained parse metadata and vectors with later selector/display consumers: known — yes [wired] (This chain has inspected conditional persistence and later consumers; no observed activation, truthful extraction or improvement is implied. Host SDK prompt delivery need not succeed for internal ranking reuse.); Retrieval activity automatically updating durable access counts for later ranking: known — yes [wired] (Durable behavior-shaping feedback qualifies without model training or an independent outcome oracle.); Shipped engine/helper candidate outputs and task HTTP semantic admission at the registered Go endpoints: known — no [wired] (The negative is resolved only for these shipped candidate-to-retention paths. It does not cover administrative loading, an added external writer, the declared future loop, or the separate metadata/vector and feedback positives.); Proposed automatic durable semantic extraction from episode traces: partial — yes [claimed] (Claimed trace learning is not a wired complete write-and-later-consumer chain. Correctly shaped direct inputs enable candidates, not retention.); Generic episode derivation inputs and externally initialized semantic contents: not-determinable (Generic descriptions may be traces or authored static examples; initial semantic population has no producer/input history. Neither a universal yes nor a system-wide no follows.)	partial coverage (The recorded consumers establish knowledge, learning, ranking and routing on distinct paths. No blanket instruction, enforcement or validation finding is inferred from fixed code, authentication, confidence or unconsumed flags; unknown initial semantic update consumers keep this axis partial.); Episode raw content and domain/intent labels at server context construction and conditional later prompt consumer: known — knowledge [wired], knowledge [afforded] (Context advice does not become standing instruction or enforced privilege through system-prompt placement.); Retained session descriptions at automatic metadata/vector production: known — learning [wired] (This is the update-consumer force, separate from raw text knowledge at the prompt consumer; improved capacity is not established.); Episode vectors, timestamps and retrieval counts at retrieval ranking; temporal IDs at context expansion: known — ranking [wired], routing [wired] (Fixed ranking code still gives its retained inputs ranking authority; graph input selection is not causal inference.); Retained domain/intent at maintenance grouping and raw episodes at intended durable semantic update: known — routing [afforded], learning [claimed] (The selector and proposed update are separate effects. PascalCase fetched episodes defeat intended grouping; direct compatible dictionaries do not repair persistence.); Semantic vector/payload at search ranking and server context consumption: known — ranking [wired], knowledge [wired], knowledge [afforded] (Consumer effects are inspectable despite unknown stored content. Confidence and learned-fact labels establish no independent validation authority.); Initial semantic content at unidentified prior admission/update consumers: not-determinable (Unknown administrative provenance and control prevent classifying any prior learning, instruction, validation or enforcement effect. SQL withdrawal flags and unused graph/provenance fields do not establish one.)	AAS-2026-10-14-rem-a81fdc774400-01
RSIAgent	2	Canonical and guest experience-bank trees, frozen copies, bank access inventories, journals/manifests and checked recovery checkpoints; Actor WORK LOG/fallback/restart handoffs; Curriculum notes, histories, outcomes and Actor diagnosis; Verifier retained substantive history, image attachments and private scratch. Includes unresolved initial-bank provenance, other dispatched recovery variants, opaque payloads and possible provider-side use-accumulated state. Excludes static shipped prompts/configuration/code, provider weights not evidenced as changed through use, ordinary unretained VM state, official score/report archives without an adaptive memory consumer and unspecified external API callers as consumer routes.	partial coverage (Inspected local retained objects use files and Python state. Static repository code is not use-derived repo memory, and remote weights are not evidenced as an updated memory substrate. Recovery variants and possible opaque provider-side accumulated state remain unresolved; positives establish existence, not complete enumeration.); Canonical bank, guest copies and frozen evaluation copies: known — files [wired] (Filesystem retention is known independently of opaque file content.); Bank journal, snapshots and exact-tree recovery checkpoints: known — files [wired] (Archives and manifests are retained on the host filesystem.); Actor WORK LOG, folded summary and restart handoff: known — in-memory [wired], files [wired] (Both live summary state and saved artifacts are accounted for.); Curriculum notes, role history, latest outcomes and Actor diagnosis: known — in-memory [wired], files [wired] (Inspected persistent session and artifact routes retain these separate products.); Verifier retained history, replay checkpoints and private scratch: known — files [wired], in-memory [wired] (Scratch payload opacity does not obscure its filesystem substrate.); Durable image attachment bytes and identity metadata: known — files [wired], in-memory [wired] (JSON/base64 encoding is not a separate storage substrate.); Other dispatched recovery variants: unverified practice, post-target learning, ready-target replacement and remaining wave recovery: uninspected (Variant bodies were not fully inspected; their retained parts and this axis cannot be completely classified.); Provider-side state accumulated through use, if any, behind the Actor, Verifier, Curriculum and vision endpoints: uninspected (Provider internals are inaccessible. Local history does not establish the existence, substrate, form or operations of provider-side accumulated state; no positive provider-memory value is assigned.)	partial coverage (Inspected use-derived admission is automatic. Operator launch, selecting a recovery plan or choosing an attachment path does not establish manual approval of the retained content. Recovery variants and possible opaque provider-side accumulated state remain unresolved; positives establish existence, not complete enumeration.); Terminal bank and diagnosis proposal, audit and publication: known — automatic [wired] (Human launch is not per-content approval.); Journal/checkpoint creation and exact-tree packaging: known — automatic [wired] (Host snapshot scheduling is a distinct admission mechanism.); WORK LOG folding, deterministic fallback, restart handoff and optional strategy advice: known — automatic [wired] (All documented alternatives are scheduled and admitted by software/model.); Curriculum notes nonempty replacement: known — automatic [wired] (Authorship is not the admission classification; audited model replacement is software/model control.); Verifier substantive checkpoint/replay and image archiving: known — automatic [wired] (Structural host selection is not human approval.); Verifier scratch content edits: partial — automatic [wired] (Verifier production and host retention/export are wired; initial scratch contents and all per-file content controls are not supplied.); Initial supplied bank and unspecified direct write_memory API caller: not-determinable (Initial author/approval control is unknown. Direct replacement is a storage capability with no identified caller admission mechanism; neither establishes manual or automatic content admission.); Other dispatched recovery variants: unverified practice, post-target learning, ready-target replacement and remaining wave recovery: uninspected (Variant bodies were not fully inspected; their retained parts and this axis cannot be completely classified.); Provider-side state accumulated through use, if any, behind the Actor, Verifier, Curriculum and vision endpoints: uninspected (Provider internals are inaccessible. Local history does not establish the existence, substrate, form or operations of provider-side accumulated state; no positive provider-memory value is assigned.)	partial coverage (Direction distinguishes automatic context/attachment delivery from requested file reads. Stored journals without a consumer are not counted merely for existing; classified journals have evidenced recovery/phase consumers. Recovery variants and possible opaque provider-side accumulated state remain unresolved; positives establish existence, not complete enumeration.); Bank attachment and subsequent Actor-requested content reads: known — push [wired], pull [wired] (Automatic snapshot/inventory supply and later requested file returns are separate operations in one chain.); Optional canonical bank access at Curriculum: known — push [wired], pull [wired] (Software selects evidence-bank or empty-tree mode; discretionary file inspection remains separate.); WORK LOG and restart continuation delivery: known — push [wired] (Later Actor context is automatically assembled from retained summary/handoff.); Curriculum private notes, role history, outcomes and diagnosis delivery: known — push [wired] (Repeated charters/context receive retained products without a new content request.); Verifier checkpoint, pending observations and image rehydration: known — push [wired] (Recovery state feeds the next provider invocation automatically.); Bank checkpoints/manifests at checked phase/recovery consumption: known — push [wired] (Operator plan choice is upstream selection input, not a request by the later resumed role for each retained object.); Opaque Verifier scratch content access: partial — push [wired] (Lifecycle restore is known; no actual scratch reads identify all content read-back paths.); Other dispatched recovery variants: unverified practice, post-target learning, ready-target replacement and remaining wave recovery: uninspected (Variant bodies were not fully inspected; their retained parts and this axis cannot be completely classified.); Provider-side state accumulated through use, if any, behind the Actor, Verifier, Curriculum and vision endpoints: uninspected (Provider internals are inaccessible. Local history does not establish the existence, substrate, form or operations of provider-side accumulated state; no positive provider-memory value is assigned.)	partial coverage (Inspected push selectors are structural availability/mode/checkpoint choices. No image digest, requested filename, or archive catalog alone establishes identifier-based push; model-selected compression meaning is not a later push signal. Recovery variants and possible opaque provider-side accumulated state remain unresolved; positives establish existence, not complete enumeration.); Host task bank attachment push: known — coarse [wired] (Trigger is task/attempt attachment; configured snapshot and availability select the complete bank and inventory, not relevance-ranked bodies.); Curriculum bank exposure and note/outcome reinsertion: known — coarse [wired] (Trigger is next Curriculum round; mode selects evidence bank versus empty tree and session availability selects retained notes/latest outcome.); WORK LOG summary/handoff push: known — coarse [wired] (Trigger is subsequent call or stalled restart; summary state and eligibility/budget select continuation text. Semantic selection happened at writing, not push retrieval.); Verifier substantive replay and image delivery: known — coarse [wired] (Trigger is decoder/segment recovery; parsed action category and checkpoint boundary select substantive history and pending exact observations.); Checked lineage and phase-bank push: known — coarse [wired] (Trigger is recovery/phase launch; plan/checkpoint lineage chooses the saved boundary, with hashes checking consistency rather than matching relevant content to a query.); Verifier private scratch archive restoration selection: partial — coarse [wired] (Available archive versus fresh scratch is known; contents and finer selection remain opaque.); Foreground Actor requested file-body returns only: inapplicable (This operation is pull. A requested filename and Actor relevance judgment do not establish an identifier-based or inferred-judgment push selector.); Other dispatched recovery variants: unverified practice, post-target learning, ready-target replacement and remaining wave recovery: uninspected (Variant bodies were not fully inspected; their retained parts and this axis cannot be completely classified.); Provider-side state accumulated through use, if any, behind the Actor, Verifier, Curriculum and vision endpoints: uninspected (Provider internals are inaccessible. Local history does not establish the existence, substrate, form or operations of provider-side accumulated state; no positive provider-memory value is assigned.)	partial coverage (The yes values concern durable trace-fed context/advice writes, not parameter training or demonstrated improved capacity. Raw trace retention alone is not upgraded to trace learning, and no unsupported system-wide no is asserted. Recovery variants and possible opaque provider-side accumulated state remain unresolved; positives establish existence, not complete enumeration.); Terminal bank distillation and Actor diagnosis: known — yes [wired] (Qualifying automatic write and later consumers are wired; improvement and compliant semantic criticism are not observed.); Actor folded WORK LOG, fallback and synthesized restart handoff: known — yes [wired] (Both lossy model text and deterministic digest persist for later Actor invocations.); Curriculum private-note update: partial — yes [afforded] (Iterative notes use supplied outcomes/diagnosis/practice evidence, but no generated note shows each replacement's trace dependence or exact source selection.); Verifier substantive replay, raw images and scratch versus derived behavior-shaping update: not-determinable (Replay/image retention preserve selected raw activity, not an evidenced newly extracted lesson. Opaque scratch could contain derived notes, but no artifact or qualifying write establishes it. No bounded absence record warrants a local no.); Journal snapshots and initial bank: not-determinable (Snapshots preserve existing derived memory, rather than establishing another trace-fed learning write. Initial content has unknown provenance; no bounded absence supports no.); Other dispatched recovery variants: unverified practice, post-target learning, ready-target replacement and remaining wave recovery: uninspected (Variant bodies were not fully inspected; their retained parts and this axis cannot be completely classified.); Provider-side state accumulated through use, if any, behind the Actor, Verifier, Curriculum and vision endpoints: uninspected (Provider internals are inaccessible. Local history does not establish the existence, substrate, form or operations of provider-side accumulated state; no positive provider-memory value is assigned.)	partial coverage (Update learning, downstream knowledge and host metadata controls are classified separately. No ranking value follows from foreground Actor discretionary searching, and no standing instruction value follows merely from fallible advice. Recovery variants and possible opaque provider-side accumulated state remain unresolved; positives establish existence, not complete enumeration.); Retained Actor experience and terminal evidence at same-Actor bank/diagnosis update: known — learning [wired] (Learning is the update consumer effect, separate from downstream advice and demonstrated improved capacity.); Bank advice at fresh Actor and optional Curriculum reader: known — knowledge [wired] (The inspected advice channel is fallible model-interpreted context, not an enforced lesson interpreter.); Journal manifests and saved lineage fields at recovery and phase gate: known — validation [wired], enforcement [wired], routing [wired] (This is authority of retained metadata, not of static admission code.); WORK LOG at subsequent Actor and fresh stalled attempt: known — knowledge [wired] (Model interpretation of summary/advice is distinct from code thresholds.); Retained older Actor turns/current log at summarizer or restart synthesis: known — learning [wired] (This is the update side of the summary chain.); Curriculum notes, diagnosis and accumulated outcome reports at later Curriculum: known — knowledge [wired] (Free-form content supplies decision context; it does not itself enforce READY/STALLED.); Retained Curriculum experience at private-note update: partial — learning [afforded] (The retained notes are authored over supplied outcomes and practice evidence. Particular trace dependence of each note replacement is not evidenced.); Verifier retained textual evidence, image observations and reasoning at subsequent calls: known — knowledge [wired] (Delivery is evidence/context supply, not proof of correct reliance.); Opaque bank executable/data alternatives and Verifier scratch payload effects: not-determinable (No actual payload or differentiated consumer effect establishes additional authority; potential execution is not enforcement, instruction or validation evidence.); Other dispatched recovery variants: unverified practice, post-target learning, ready-target replacement and remaining wave recovery: uninspected (Variant bodies were not fully inspected; their retained parts and this axis cannot be completely classified.); Provider-side state accumulated through use, if any, behind the Actor, Verifier, Curriculum and vision endpoints: uninspected (Provider internals are inaccessible. Local history does not establish the existence, substrate, form or operations of provider-side accumulated state; no positive provider-memory value is assigned.)	AAS-2026-10-14-rsiagent-b11225763c0b-01
SAGE	2	Included: paired rollout functions; continuing evaluation and expert libraries; predefined function/query/vector alternatives; access metadata and tensors; raw expert calls and derived supervised corpus; use-updated RL/SFT weights and recovery state; optional completion cache. Included unresolved alternatives: provider-side retention and uninspected launcher/backend memory behavior. Excluded: static shipped prompts, API docs/demonstrations as independent memory, ordinary task-local history without a derived later-consumed artifact, upstream environment storage and external training/provider implementation details. Imported static content inside qualifying retained corpora remains mixed lineage, not an independently qualifying trace.	partial coverage (Inventory covers code, access metadata/tensors, traces, corpus, learned parameters and optional response envelopes. Coverage is partial because the included provider/backend alternatives remain unresolved; wired findings are capacities, not observed use.); Paired function dictionaries: known — in-memory [wired] (The internal route retains code in process memory, not a disk skill library.); Evaluation and expert source records and query associations: known — in-memory [wired], files [wired] (Query-list output serialization and automatic restart from output are not established; these gaps do not change the known physical storage paths.); Skill/query access tensors: known — in-memory [wired], files [wired] (Dense encoding does not establish a vector database; the retained substrate is tensors and files.); Raw expert logs and derived corpus: known — files [wired] (Dataset persistence is separate from trained weights.); RL and SFT parameter retention: known — model-weights [wired], in-memory [wired], files [wired] (Backend checkpoint correctness and exact artifacts are not established.); Optional completion cache: known — files [wired] (The wrapper establishes file-backed caching, not a separately evidenced KV service.); Included provider-side retention and alternate launcher/backend-specific memory behavior: uninspected (Provider state is opaque; alternate launcher/backend retention is not exhaustively inspected. Records do not establish an additional memory inventory or its classification. These alternatives cannot be excluded by the inspected paths.)	partial coverage (Write control is classified at each admission mechanism, not from the candidate author or physical file I/O. Operator selection of an existing model path is a read choice, not a separately established content write/admission mechanism. Coverage is partial because the included provider/backend alternatives remain unresolved; wired findings are capacities, not observed use.); Generated functions, derived indexes, query appends, expert corpus and cache admission: known — automatic [wired] (Launching/configuring these routes does not make their admissions manual.); RL and configured SFT updates: known — automatic [wired] (Separate launch prerequisites do not change update agency.); Predefined-library preparation and initial weights: not-determinable (The content supplier/editor/approver and exact initial artifacts are unknown; loading alone establishes no manual admission.); Included provider-side retention and alternate launcher/backend-specific memory behavior: uninspected (Provider state is opaque; alternate launcher/backend retention is not exhaustively inspected. Records do not establish an additional memory inventory or its classification. These alternatives cannot be excluded by the inspected paths.)	partial coverage (Requests and automatic supply are separated within chains; output persistence alone is not an extra read-back route. Coverage is partial because the included provider/backend alternatives remain unresolved; wired findings are capacities, not observed use.); Task initialization library delivery: known — push [wired] (Selection precedes interpreter execution and conditional prompt rendering.); Later training generation and automatic resume selection: known — push [wired] (No task-model request selects particular learned parameter contents.); Explicitly selected SFT/checkpoint load: known — pull [wired] (The subsequent automatic inference consumption is distinct from the requested load.); Predefined constructor loading: known — push [wired] (Loading output JSONL on restart is not wired; only the separately located predefined branch is classified.); Completion cache return: known — pull [wired] (Requested return is not an additional push merely because the adapter delivers it.); Supervised corpus load: known — pull [wired] (Selection and masking inside training do not turn requested dataset consumption into unsolicited delivery.); Included provider-side retention and alternate launcher/backend-specific memory behavior: uninspected (Provider state is opaque; alternate launcher/backend retention is not exhaustively inspected. Records do not establish an additional memory inventory or its classification. These alternatives cannot be excluded by the inspected paths.)	partial coverage (Signals classify actual push selectors, not catalog identifiers, payload format or requested response delivery. Coverage is partial because the included provider/backend alternatives remain unresolved; wired findings are capacities, not observed use.); Paired and expert whole-library push: known — coarse [wired] (Trigger: next subtask/task initialization. Input: matching pair/scenario library availability. Selected part: all retained source.); Default evaluation source push: known — identifier [wired] (Trigger: task initialization. Input: scenario identifiers and retained task/name metadata. Selected part: matching function records. Matching is substring-based, not equality.); Embedding source push: known — inferred-embedding [wired] (Trigger: task initialization with the alternate retrieval mode. Inputs: current instruction and retained vectors. Selected part: ranked function source; skill and query modes use different limits.); N-gram source push: known — inferred-lexical [wired] (Trigger: n-gram initialization. Input: current instruction and retained queries. Selected part: matching task-ID functions with name deduplication.); Automatic weight and checkpoint supply: known — coarse [wired] (Trigger: subsequent generation/resume. Input: active state or latest-iteration availability. Selected part: whole parameter/checkpoint state.); Predefined whole-file supply: known — coarse [wired] (A filename does not itself establish identifier-targeted selection of individual delivered parts.); Pull-only cache and supervised dataset operations: inapplicable (Push selection signal is inapplicable to these bounded requested returns/loads; no separate push is inferred.); Included provider-side retention and alternate launcher/backend-specific memory behavior: uninspected (Provider state is opaque; alternate launcher/backend retention is not exhaustively inspected. Records do not establish an additional memory inventory or its classification. These alternatives cannot be excluded by the inspected paths.)	partial coverage (The axis uses the automatic trace-fed write definition, not demonstrated capacity improvement. Raw logs alone do not qualify. Coverage is partial because the included provider/backend alternatives remain unresolved; wired findings are capacities, not observed use.); Paired extracted functions to second-task consumers: known — yes [wired] (Conditional internal route qualifies despite per-rollout lifetime; useful reuse is unobserved.); Continuing evaluation and expert function accumulation: known — yes [wired] (Current-task failure and imperfect extraction do not erase the wired trace-to-durable-artifact chain.); RL outcome-fed parameters: known — yes [wired] (The shipped two-node launcher cannot reach this route without correcting dataset-reader selection.); Expert traces to SFT parameters: known — yes [wired] (Configured training and operator-linked later use are not evidence that the pipeline ran or improved behavior.); Predefined content, opaque initial weights and optional response replay: not-determinable (Predefined derivation is unknown. Memoizing/replaying a completion is established, but records do not establish whether that response is trace-derived behavior-shaping material; no system-wide no follows.); Included provider-side retention and alternate launcher/backend-specific memory behavior: uninspected (Provider state is opaque; alternate launcher/backend retention is not exhaustively inspected. Records do not establish an additional memory inventory or its classification. These alternatives cannot be excluded by the inspected paths.)	partial coverage (Authority is tied to concrete consumers. Transient failure-marker gates and static prompts do not become retained validation/enforcement memory merely because they gate writes. Coverage is partial because the included provider/backend alternatives remain unresolved; wired findings are capacities, not observed use.); Retained source at interpreter and task model: known — instruction [wired], knowledge [wired] (Execution and advisory model context are separate effects; neither establishes mandatory invocation or correctness.); Source and query content at retrieval scoring: known — ranking [wired] (Fixed scoring of retained content is ranking authority, even without trained ranking policy.); Task/name/index associations at selectors: known — routing [wired] (Identifier selection is not a semantic correctness check.); Raw trajectories and selected corpus at updates: known — learning [wired] (The update consumer has learning authority independently of the output source's downstream guidance.); Checkpoint tracker and saved training state at resume: known — routing [wired] (Resume selects state; it does not validate model quality.); Cached envelope at the agent: known — knowledge [wired], instruction [wired] (This optional replay route is disabled in normal evaluation; cache hits and downstream effects are not observed.); Included provider-side retention and alternate launcher/backend-specific memory behavior: uninspected (Provider state is opaque; alternate launcher/backend retention is not exhaustively inspected. Records do not establish an additional memory inventory or its classification. These alternatives cannot be excluded by the inspected paths.)	AAS-2026-10-15-sage-d0de15ae5730-01
SAGE	2	Included: accumulated governed memory text/envelopes and indexes; relationships; local evidence, verdict and operator-review state; capture units/progress; validator feedback; imported and federated copies; boot preferences and cleanup continuation; benchmark retained text/results; exact-local Messages and unresolved task-board, pipeline/history, recovery and removal alternatives. Excluded: shipped static instructions as memory objects, external model weights, host-owned conversation/compaction policy, transient current-call scores and binary installation state unrelated to this memory boundary.	partial coverage (Known retained substrates coexist with unresolved preference/run storage and included recovery/coordination alternatives. External inference weights and current-call transient scores are excluded, not classified as accumulated memory.); Canonical memory envelopes, scoped content and validator feedback state: known — kv [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Personal-node content, gate evidence/verdict/review, links and exact-local Messages: known — sqlite [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Daemon memory serving records: partial — rdbms [wired] (The daemon substrate is established, but versioned serving/projection alternatives are not exhaustively covered.); Retained dense-vector retrieval index: known — vector [wired], sqlite [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Per-thread capture progress sidecars: known — files [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Benchmark retained result output: known — files [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Boot preferences and cleanup continuation state: not-determinable (Persistence is established, but these records name a preference store and durable run state without resolving a controlled physical substrate for each.); Included task-board, pipeline/history and federation removal/recovery alternatives beyond the inspected exact-local Messages path: uninspected (Memory coverage and integration issues leave these operative families incompletely traced. Exact-local Messages resolves only its own branch; complete storage substrate classification is prevented.)	partial coverage (Automatic and explicit manual mechanisms coexist. Human workflow initiation, generic signatures and physical I/O do not resolve per-content control on every branch.); Ordinary proposals, bridge outputs, capture units, session events, vector regeneration, verdict caches and statistics admission: known — automatic [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Operator boot preference replacement and semantic review: known — manual [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Import conversion, federated copying and local Messages protocol admission: known — automatic [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Caller-authored correction, reflection, co-commit and relationship content control: not-determinable (Generic client identity/signatures do not distinguish explicit human content selection from agent authorship. Computational veto/admission is known, but manual control of these caller alternatives is not.); Initial existing content and automatic provider-ready vector repair alternative: uninspected (Deployed initialization is unavailable. The manual re-embedding lane records an automatic repair alternative without independently tracing its complete control path.); Included task-board, pipeline/history and federation removal/recovery alternatives beyond the inspected exact-local Messages path: uninspected (Memory coverage and integration issues leave these operative families incompletely traced. Exact-local Messages resolves only its own branch; complete write agency classification is prevented.)	partial coverage (Requested returns and automatic downstream prompt supply are distinguished inside chains. Exact-local Messages adds a pull witness, not complete coordination coverage.); Foreground topic/ID/list/link recall and explicit Messages inspection/result paging: known — pull [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); SessionStart recent prefetch and resumed-thread reconstruction: known — push [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Bridge prior-knowledge chain: known — pull [wired], push [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Later gate, feedback and continuation workers: known — pull [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Later inception guidance: known — pull [wired], push [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Benchmark seed-to-answer route: known — pull [wired], push [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Included task-board, pipeline/history and federation removal/recovery alternatives beyond the inspected exact-local Messages path: uninspected (Memory coverage and integration issues leave these operative families incompletely traced. Exact-local Messages resolves only its own branch; complete read-back direction classification is prevented.)	partial coverage (Positive unsolicited selectors coexist with unresolved automatic prompt-selection alternatives; known pull routes cannot make the whole axis inapplicable.); SessionStart recency prefetch: known — coarse [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Resumed-thread capture push: known — identifier [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Bridge automatic prior-excerpt supply: partial — coarse [wired] (Domain and item bounds are established. The accepted bridge record does not resolve every search mode behind query_knowledge; inferred embedding/lexical selection cannot be upgraded from another recall witness.); Inception custom boot guidance: known — coarse [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Benchmark answer-prompt supply: partial — coarse [wired] (The chain has embedding and optional query expansion, but records do not fully isolate every alternative selector in downstream prompt supply; question/reference IDs alone do not establish identifier-targeted push.); Explicit requested foreground/relationship/Message reads: inapplicable (These specifically named delivery operations fulfill consumer requests. Their embedding/lexical/identifier selectors are pull selectors, not positive push signals.); Included task-board, pipeline/history and federation removal/recovery alternatives beyond the inspected exact-local Messages path: uninspected (Memory coverage and integration issues leave these operative families incompletely traced. Exact-local Messages resolves only its own branch; complete push selection signal classification is prevented.)	partial coverage (Two qualifying automatic update chains are wired. Capture records establish raw retention but declare no bounded absence warrant for a local no; unresolved routes prevent a complete system-wide assessment.); Consensus vote feedback to durable quorum factors: known — yes [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Bridge outcome-to-advice-to-later-context chain: known — yes [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Capture selection/encoding and reconstructed trace context: not-determinable (Records establish selected conversational text re-encoding/reconstruction, not a derived recommendation or generated continuation summary. Raw retention alone establishes no yes; no canonical bounded-absence record is declared to warrant a controlled no.); Caller reflection and SessionEnd event recording: not-determinable (Reflection receives already-authored text without a traced original activity-to-generation chain. SessionEnd retains fixed event text without a derived behavior-shaping update. These records do not establish a qualifying trace-fed learning value for the broader alternatives.); Task/history, benchmark production feedback and recovery alternatives: uninspected (Messages retains raw requests/results, not an established learning update. Other included history/task and production-feedback consumers remain untraced; no system-wide negative follows.)	partial coverage (Authority follows each consumed part and effect. Textual requests and service fences are not interchangeable; unresolved consumers prevent whole-system completeness.); Recalled/captured prose and Messages results at foreground or benchmark answer models: known — knowledge [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Operator boot text at the external inception consumer: known — instruction [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Retained content, vectors, confidence and validator statistics at retrieval/quorum scoring: known — ranking [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Review enum and cached semantic verdict at Gate.Apply: known — validation [wired], enforcement [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Lifecycle/access metadata, sync suppression and message session/revision state at service checks: known — validation [wired], enforcement [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Capture cursor, cleanup pending intent and sync recovery ledger at later workers: known — routing [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Recorded vote outcomes at the statistics update consumer: known — learning [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Prior retained lessons at bridge proposal generation: known — knowledge [wired], learning [wired] (The cited records resolve this narrowly named part; operation and benefit were not observed.); Task-board/triple and benchmark metric consumers beyond inspected routes: uninspected (Triple insertion alone supplies no traced consumer. Full task-board and production feedback from benchmark metrics are unresolved; unconsumed metadata is not assigned authority.); Included task-board, pipeline/history and federation removal/recovery alternatives beyond the inspected exact-local Messages path: uninspected (Memory coverage and integration issues leave these operative families incompletely traced. Exact-local Messages resolves only its own branch; complete behavioral authority classification is prevented.)	AAS-2026-10-15-sage-d3843de6d2dd-01
sage-wiki	2	Accumulated source/capture files, summaries, articles, entries/chunks/vectors, ontology entities/relations/aliases, communities, lessons, retained answers/trust evidence, manifest and compile recovery state; changed installed prompts/configuration/skills; their admissions, maintenance and later consumers. Included unresolved alternatives: provider-native batch checkpoints, backend representations, deeper restored state, lint/evaluator retained artifacts and external lesson consumers identified in accepted records. Excludes shipped static defaults, ordinary invocation-local Q&A/map/reduce context, external provider-internal weights/state and enclosing-host history. Restore/pack routes are included only for retained material installed for later use, not all runtime storage or permissions.	partial coverage (Known file, relational and SQLite parts coexist with incomplete checkpoint, backend and ancillary-artifact coverage. External provider weights and enclosing-host history are outside the retained-memory boundary.); Readable source, summary, concept, community and answer files; source/concept manifest: known — files [wired] (Includes readable bodies and file metadata, not merely a search display.); Entry/chunk, ontology, alias, community, learning, output-trust and compile-item rows: known — rdbms [wired] (Database objects are distinct from their markdown counterparts; backend parity is not asserted.); Inspected document/chunk embedding BLOBs: known — sqlite [wired] (This unit covers the inspected SQLite vector path only.); Installed workspace prompt, skill and configuration changes: known — files [wired] (Only installed or edited retained state is included, not the shipped embedded defaults or transient registry.); Provider-native batch checkpoints; ANN/mmap/PostgreSQL alternative representations; deeper restore and lint/evaluator artifacts: uninspected (Available records identify these alternatives without a complete object/substrate inventory. Remote restore copies known state but does not establish every remote operative representation.)	partial coverage (Automatic admission is wired on specified computational routes. Caller identity and uninspected alternatives prevent a complete agency classification; no manual value is inferred from a command name.); Model capture, standard compile/curation/quality rewrite, scribe, community and automatic trust admissions: known — automatic [wired] (Includes automatic alias application and promotion when their configured conditions hold.); Program-selected metadata repairs and source/age/validity-driven maintenance: known — automatic [wired] (An explicit request to run repair/cleanup is not approval of each retained edit.); Caller-authored captures/direct writes/lessons, explicit alias/output decisions, pack/config editing and restore selection: not-determinable (Generic caller may be human or computational. Records do not establish a per-content human decision, so neither manual nor automatic admission is assigned to these caller-controlled branches.); Provider-native batch checkpoint admission and full ancillary-artifact write paths: uninspected (The inspected standard path does not classify all alternative checkpoint or evaluator/lint admissions.)	partial coverage (Requested external/answer routes are pull; automatically supplied compile and maintenance inputs are push at their own consumer boundary. Unknown alternatives prevent a pull-only or complete-direction conclusion.); Requested search/full read/ontology traversal, document answering and local/global graph answering: known — pull [wired] (Graph expansion and community selection inside requested answering do not by themselves add an unsolicited push.); Automatic compile/resume selection of retained summaries, existing articles and relevant source sections into independent completion calls: known — push [wired] (The upstream operator may start compile; the later extraction/writer completion is automatically supplied selected memory.); Automatic maintenance reuse of alias decisions, output evidence and queue/cache state: known — push [wired] (These internal software consumers count separately from the requested return to an external host.); Learning recall external use, batch checkpoint reload and incompletely traced lint/evaluator/restored consumers: not-determinable (Library recall and restoration capabilities do not resolve every actual later consumer or direction.)	partial coverage (Each positive signal names an actual push operation. Query search/graph expansion is classified under its pull request and does not establish push signals. Opaque alternatives remain unresolved, not inapplicable.); Compile/resume saved-summary and existing-article injection: known — identifier [wired] (Actual matched retained bodies are delivered, not merely an ID displayed in a catalog.); Compiler source-section supply to article-writing completion: known — inferred-lexical [wired], coarse [wired] (Trigger is article preparation; selector inputs are concept names/aliases and source text; selected parts are the source sections, separate from the whole existing article.); Alias replay and candidate arbitration using retained decisions: known — identifier [wired] (Compile-end/re-extract triggers maintenance; pair identity selects relevant retained decision state. Retained reason prose is not interpreted by the rejection selector.); Recurring-answer trust comparison: known — inferred-embedding [wired] (The retained answer is selected by question embedding similarity, not by the subsequent answer-agreement judgment.); Automatic queue/cache/state filtering: known — coarse [wired] (These are coarse lifecycle filters, not proof of semantic targeting or exact content freshness.); Batch checkpoint selection and hypothetical learning/lint/evaluator push branches: not-determinable (Accepted records do not enumerate selector inputs and delivered payloads for these alternatives.)	partial coverage (The scribe route establishes a local wired yes. Unresolved input provenance and alternate writes prevent complete coverage; no system-wide negative or improvement claim follows.); Session scribe: recorded conversation to retained ontology to later explicit ontology query: known — yes [wired] (A wired trace-fed behavior-shaping artifact suffices. No improved capacity, automatic host hook or general prose-Q&A delivery is asserted.); Conversation/text capture, arbitrary imported source compilation, caller-authored lessons and saved answers: not-determinable (Some input may record activity, but accepted records do not resolve the original content of every route or establish every qualifying trace-to-later-consumer chain. Generic Q&A prose and a learning label are not themselves trace evidence.); Batch checkpoints, full evaluator-session ingestion and lint feedback alternatives: uninspected (The records do not completely trace these alternative automatic writes through durable artifacts to later consumers. No bounded absence record licenses a local or system-wide no.)	partial coverage (Actual content and control consumers support seven local forces. Unwired lesson delivery and incomplete checkpoint/ancillary consumers prevent a complete authority inventory.); Retained prose, graph descriptions/evidence and community summaries supplied to answering, writing, traversal and review consumers: known — knowledge [wired] (Includes existing articles in later writing and stored answers at comparison consumers; delivered content need not activate real host behavior.); Entry/chunk text, vectors, entity aliases and persisted scores at retrieval/candidate-priority consumers: known — ranking [wired] (Ranking input does not require a retained learned ranking policy.); Manifest/source hashes, queue/pass/lease state, community cache metadata and applied alias identities at later work selectors: known — routing [wired] (These controls route computation rather than supply ordinary prose advice.); Alias rejection/active status and output trust/confirmation/grounding/dependency state at admission and retrieval controls: known — validation [wired], enforcement [wired] (Veto force is path-local: direct file reads and non-output admissions bypass output trust.); Recorded session text at automatic scribe update consumer: known — learning [wired] (Learning authority at the update is distinct from knowledge authority of the resulting definitions; no improved capacity is established.); Installed workspace prompt overrides at completion-provider consumers: known — instruction [wired] (Override loading and the internal completion-consumer path are wired; model compliance and observed activation are not established.); Generated or installed standing guidance at external host consumers: known — instruction [afforded] (The conditional instruction path is classified as afforded, not wired. File installation is implemented, but actual host loading, precedence, reload and compliance remain unestablished.); Stored learning entries at hypothetical external consumers; opaque batch and full lint/evaluator feedback consumers: not-determinable (No native lesson-to-article delivery is established. A template slot or retained benchmark result does not establish instruction or learning force.)	AAS-2026-10-15-sage-wiki-cc43456fb833-01
secure-llm-wiki	2	Application-accumulated claim text in pages/quarantine, eligibility/provenance and gate diagnostics, persisted embeddings/access keys, editable policy after initialization, and data-root Git history; admission, maintenance and later ingestion/query/operator consumption routes. Direct API callers and initial payload provenance remain unresolved alternatives, as does opaque external provider retention if present. Excludes shipped static prompt/code policy, transient source/completion/query output, and external weights as such: none is shown to be application-accumulated memory.	partial coverage (Inventory covers working claims, diagnostics, policy, vectors and history. Provider-side retained state remains opaque; no deployed instance is supplied.); Claim text, eligibility/provenance, gate diagnostics and retained editable policy: known — files [wired] (These application-owned objects have inspected file persistence; YAML/JSON encodings are not additional substrates.); Saved embedding coordinates and UUID access structure: known — files [wired], vector [wired] (The vector store is not a query retrieval index and is separate from model weights.); Data-root historical snapshots and commit messages: known — repo [wired] (Default reset destroys this repository; ordinary file reload does not read history.); External provider-side retention, if any, beyond application-owned objects: not-determinable (Adapters do not expose provider retention or update internals. No substrate, admission, transformation or consumer classification follows from model names; this alternative remains unresolved, not absent. External provider weights are not claimed to be application-accumulated memory.)	partial coverage (Automatic CLI admission and explicit operator mechanisms are supported separately. Generic API callers and provider admission remain unresolved.); Ordinary gated claim/metadata/diagnostic and post-commit vector admission: known — automatic [wired] (Starting ingestion or assigning source trust is not approval of each retained proposition.); Template policy initialization and normal history commits: known — automatic [wired] (Physical I/O and template authorship are not substituted for admission control.); Documented manual content/policy editing and confirmed clear/reset: known — manual [afforded], manual [wired] (Direct file-edit capability is distinguished from the wired CLI confirmation.); Generic direct storage, embedding-save and add-rule callers: not-determinable (Caller invocation can admit or replace payloads without gates, but caller identity does not resolve human content approval.); External provider-side retention, if any, beyond application-owned objects: not-determinable (Adapters do not expose provider retention or update internals. No substrate, admission, transformation or consumer classification follows from model names; this alternative remains unresolved, not absent. External provider weights are not claimed to be application-accumulated memory.)	partial coverage (All identified application consumer chains are requested pull. Unknown external/API consumers and provider retention prevent declaring the entire included boundary pull-only.); CLI future-ingestion, context/query, list and configuration reload chains: known — pull [wired] (Selection before a question does not add an unsolicited push to the requested query-startup load.); Full-file and Git-history inspection/reversal: known — pull [afforded] (The shipped query model does not receive Git history.); External consumers of alternate renderers and direct storage APIs: not-determinable (Builders afford rendering caller-supplied claims but identify no external persistence-to-consumer chain. Its request/delivery behavior is not established.); External provider-side retention, if any, beyond application-owned objects: not-determinable (Adapters do not expose provider retention or update internals. No substrate, admission, transformation or consumer classification follows from model names; this alternative remains unresolved, not absent. External provider weights are not claimed to be application-accumulated memory.)	not-determinable coverage (No controlled push-selection value is established. Inapplicability is local to inspected pull chains, not a system-wide absence of push.); Completely inspected requested CLI and operator-inspection chains: inapplicable (These chains are pull-only; trust/status predicates, UUID paths and cosine admission checks are not unsolicited push selectors.); Unspecified alternate API consumers and provider-side read-back: not-determinable (No accepted record identifies a push trigger, selector input and selected retained part for these unresolved alternatives; pull-only inapplicability cannot be extended to them.)	partial coverage (A warranted local negative counts as classified coverage, but unresolved alternatives suppress a derived system-wide no. No qualifying trace-fed positive is established in accepted records.); Local shipped query-answer and feedback-to-store update path: known — no [wired] (This negative is limited to the searched query-feedback route; it does not negate source compilation or provider adaptation.); Document-compilation inputs, direct replacements and diagnostic/history reuse: not-determinable (Compilation retains source-derived context, but accepted records do not classify original documents as activity traces. Stored outcome diagnostics/history have no automatic derived-policy consumer. Arbitrary direct input origin and downstream use remain unresolved, preventing a complete trace-fed-write inventory.); External provider-side retention, if any, beyond application-owned objects: not-determinable (Adapters do not expose provider retention or update internals. No substrate, admission, transformation or consumer classification follows from model names; this alternative remains unresolved, not absent. External provider weights are not claimed to be application-accumulated memory.)	partial coverage (Consumer-specific forces coexist. No automatic learning authority is inferred from trace retention; provider-side consumers remain unresolved.); Selected claim text and rendered metadata at answering model: known — knowledge [wired] (System-message placement does not turn evidence into instruction; actual obedience and benefit are unobserved.); Active trusted retained text at later batch security review: known — knowledge [wired] (A model receives reference text, not prior full review rationale or a factual oracle.); Trust/status and source fields at context and admission predicates: known — enforcement [wired], validation [wired] (Only consumed fields have these forces; successor pointers and other unused metadata gain no authority by serialization.); Retained vectors at Gate 5 comparison: known — ranking [wired], validation [wired], enforcement [wired] (Ranking is numerical comparison, not semantic query retrieval or proof of contradiction.); Retained rule/threshold policy at future ingestion: known — routing [wired], validation [wired], enforcement [wired] (Effects operate through fixed code, not model-readable standing instructions; edits do not re-tier existing claims.); Full diagnostics/history at operator inspection and recovery: known — knowledge [afforded] (Ordinary query/review omits full reasons and history. No automatic policy-update consumer is established.); External provider-side retention, if any, beyond application-owned objects: not-determinable (Adapters do not expose provider retention or update internals. No substrate, admission, transformation or consumer classification follows from model names; this alternative remains unresolved, not absent. External provider weights are not claimed to be application-accumulated memory.)	AAS-2026-10-19-secure-llm-wiki-df02f419bddb-01
Self-Training-LLM	2	Retained title/document cache; generated question, test, context-label and answer files (text, metadata and unresolved generation payload details); derived selection scores; persisted known subset and model responses; SFT/DPO parameter checkpoints including adapter/merge forms; judge outcome cache and aggregate reports; exposed WordNet, popularity, embedding, category and train/test topic caches. Includes unknown initialization/replacement provenance locally. Excludes shipped static prompts/examples/configuration as accumulated objects, transient current-run batches/tensors, compiled caches and external dataset/provider/trainer internals. Dependency opacity and unestablished later report consumption remain limits, not evidence of absence.	known coverage (Covers accumulated local artifacts and trained parameter products, including exposed helper caches. External pretrained/provider internals are dependencies, not retained state written by this workflow; static source and current-run tensors are excluded.); Retained document, curriculum, answer/score, evaluation and alternative topic-helper files: known — files [wired] (The accepted inventory identifies the local persistence substrate for every listed file part, including vectors and symbolic access keys.); Locally saved SFT/DPO parameters, optional adapters and merged checkpoints: known — model-weights [wired] (Checkpoint directories store weights. Transient GPU state is not a separate accumulated memory object; TGI identity uncertainty does not change this substrate classification.)	partial coverage (Known producer and helper mechanisms are automatic. Neither operator launch nor mutable-file capability establishes manual admission; unverified initialization and replacement keep coverage partial.); Implemented document admission, generated-data admission/rescoring, subset creation, optimization/save, response persistence and judge-cache admission: known — automatic [wired] (Covers the software admission mechanisms rather than treating physical file I/O or workflow launch as the authority decision.); Exposed topic-helper cache construction and topic/split selection: known — automatic [afforded] (Admission is identifiable inside helpers despite absent active caller wiring.); Initial preexisting file supply/replacement, externally supplied WordNet embedder state and independently served checkpoint replacement: not-determinable (Mutable files and an operator-configured endpoint can resolve other content, but the records do not establish how that content was admitted. Selecting a checkpoint to read is not evidence of manual checkpoint-content replacement.)	partial coverage (All established read-back operations are request-driven. Terminal reports remain included retained objects without an established later consumer, so the broader inventory is partial rather than a system-wide pull-only assertion.); Later generation/training/inference/evaluation invocations requesting retained files and trained parameters: known — pull [wired] (Includes stored documents/questions/labels/scores, known/test records, responses and judge labels, plus direct or operator-configured checkpoint consumption. Batch delivery is not a separate push operation.); Explicit topic-helper calls returning requested cached lexical/topic/count/vector material: known — pull [afforded] (Helper-level requested access is established, while a deployed downstream workflow caller is not.); Terminal aggregate reports at a later consumer: not-determinable (No later automatic reader or actual human decision is recorded. Persistence alone does not establish a read-back route or its direction.)	not-determinable coverage (No supported positive push-signal value exists in the accepted inventory. The known pull units are locally inapplicable, but the unresolved report-consumer branch prevents an axis-wide inapplicable or absent assessment.); Established requested file/checkpoint/cache chains and explicit topic-helper returns: inapplicable (These assessed direction units are pull operations. File identifiers, exact prompt matching, semantic helper queries and budgets within the requested chain do not establish a push selector.); Possible later consumption of terminal aggregate reports: not-determinable (Direction and consumer are not established, so no trigger, selector input or selected retained part can be classified as a push signal. This unresolved branch cannot be declared pull-only or inapplicable.)	partial coverage (A durable trace-fed parameter path is positively wired. Complete provenance and extra opaque payload roles remain unresolved, so this is not a complete enumeration or a blanket negative for every other stored part. Cache/report retention alone is not another qualifying learning route.); Retained generated question/answer experience and derived feedback feeding SFT/DPO parameters with later checkpoint consumption: known — yes [wired] (This qualifies as a wired trace-fed update, not mere retention and not demonstrated improvement. Imported documents and static examples are mixed inputs, not independently qualifying traces.); Additional opaque generation-payload fields and initially supplied/reused state at their possible update consumers: not-determinable (Complete payload/provenance and external initialization are unresolved. No extra qualifying route or bounded local no can be inferred from that opacity; helper encoding alone is not proof of trace-fed learning.)	partial coverage (Named retained inputs establish knowledge, learning, ranking, validation, enforcement and routing at their specific consumers. Neither generic numerical behavior nor a merely possible report reader justifies a blanket authority value; those included effects remain unresolved.); Retained document/question/answer text as context or comparison evidence at producer, answer-model, NLI and judge consumers: known — knowledge [wired] (Model-interpreted influence is distinguished from execution permission and from optimization authority.); Retained SFT labels, context-answer labels and DPO candidate answers admitted to optimization: known — learning [wired] (Learning authority is at the update consumer. It does not claim better future capacity or factual acceptance.); Stored selection scores at question eligibility, pair admissibility and chosen/rejected priority consumers: known — ranking [wired], validation [wired], enforcement [wired] (These are distinct effects of retained scores under fixed code, not a retained selection policy or independent truth oracle.); Retained identity fields and completed-context labels selecting subsequent producer/evaluation work: known — routing [wired] (Only consumed identity/state is classified; merely stored metadata has no inferred authority.); Retained popularity counts and embedding arrays at exposed topic-ranking helpers: known — ranking [afforded] (This is a helper-level affordance, not active predefined-topic generation or observed semantic retrieval; unresolved FAISS and caller wiring remain limits.); Saved parameter product at downstream generation and cached outcomes at aggregate-report consumption: not-determinable (Records establish numerical generation force and direct reuse of outcome categories, but do not establish a controlled authority label for those effects alone. Weights are not standing prose instructions; cached labels do not rank future answers or gate checkpoints.); Terminal report content at any later human decision consumer: not-determinable (Appending an operator-facing report does not establish an actual later reader, reliance decision or operational consequence; no authority value is inferred for that possible consumer.)	AAS-2026-10-15-self-training-llm-5b5ee621bde2-01
semantic-engine	2	Downloaded archive files; retained SQLite roots, chunks, provenance, linkage and numerical embeddings; their admission, clearing and later query/derivation routes; generated canvas GeoJSON and its qualified browser consumer; potential Atlas-retained exports/topics and opaque downstream consumers. Initial preexisting vector provenance and external edit/admission alternatives remain unresolved. Excludes static shipped instructions/configuration as accumulated memory, transient query/comparison outputs, unused clustering results, scaffold UI state and external embedding-model weights not updated by inspected callers.	partial coverage (Local files and SQLite cover the declared local inventory. Potential Atlas state remains opaque; numerical vectors do not imply a separate vector-store substrate.); Downloaded archive files: known — files [wired] (The download implementation identifies local file retention; operation is not observed.); Derived canvas GeoJSON files: known — files [afforded] (Current integration obstacles prevent treating the connected producer/consumer route as wired.); Root text, chunk text, offsets, provenance and chunk embeddings: known — sqlite [wired] (Vector encoding/coercion remains uncertain, not the declared database substrate.); Potential Atlas-retained payloads and requested topic outputs: uninspected (The inspected request does not establish remote persistence, storage substrate or later consumers; exporter initialization/schema defects also qualify reachability.)	partial coverage (Known software admission and explicit manual clearing coexist. Initial vectors and external GUI/provider alternatives remain unresolved rather than inheriting one route's control.); Archive download, root/chunk admission and subsequent embedding updates: known — automatic [wired] (Human source selection or starting a command does not make its subsequent admissions manual.); Dataset-wide replacement of retained vectors by null: known — manual [wired] (This is a separately evidenced maintenance decision, not the null-row read selector.); Canvas GeoJSON replacement: known — automatic [afforded] (Operator initiation remains distinct from software admission; route connection is only afforded.); Initial admission of arbitrary preexisting vectors and external database-edit alternatives: uninspected (Initial vector control is uninspected; generic external administration power does not establish a per-content operator decision.); Potential Atlas-retained payloads and requested topic outputs: uninspected (The inspected request does not establish remote persistence, write/admission control or later consumers; exporter initialization/schema defects also qualify reachability.)	partial coverage (All established local routes fulfil requests. Opaque Atlas consumers prevent a complete pull-only boundary or a system-wide absence of push.); Archive loading and embedding batches requested by operator-invoked commands: known — pull [wired] (Internal batching fulfils requested work; it is not unsolicited context injection.); Query selection, root reload and terminal delivery: known — pull [wired] (Automatic scoring within the request is not a second push route.); Canvas projection input, browser whole-file fetch and Atlas export input: known — pull [afforded] (Username/author parameters select requested inputs; current integration remains qualified.); Potential Atlas-retained payloads and requested topic outputs: uninspected (The inspected request does not establish remote persistence, later read-back direction or later consumers; exporter initialization/schema defects also qualify reachability.)	not-determinable coverage (No positive push selector can be classified. The unresolved external consumer prevents making this axis inapplicable for the entire boundary.); Established request-driven archive, embedding, query, canvas and export input reads: inapplicable (These are pull operations; push-selection signal is inapplicable within this local unit. Query cosine and map filenames do not establish a push signal.); Potential Atlas-retained payloads and requested topic outputs: uninspected (The inspected request does not establish remote persistence, push trigger, selector input and selected retained part or later consumers; exporter initialization/schema defects also qualify reachability.)	not-determinable coverage (No controlled yes/no finding is established by canonical bounded trace-learning records. Numerical durability alone does not establish trace origin; opaque provider processing is not a negative.); Inspected imported-document transformations, query, clearing and canvas/export drivers: not-determinable (The records describe imported document derivation and no retained query/feedback updates, but declare no bounded trace-learning absence record. They establish neither a qualifying trace-fed write nor a profile-wide negative; actual external input contents and initialization are unresolved.); Potential Atlas-retained payloads and requested topic outputs: uninspected (The inspected request does not establish remote persistence, automatic trace-fed updates and later consumers or later consumers; exporter initialization/schema defects also qualify reachability.)	partial coverage (Actual text, ranking and routing effects are separated by consumer. Geometric display placement alone is not a controlled authority value; no instruction, validation or learning force is inferred from mere retention.); Retained root content/provenance at terminal display and chunk content at embedding input: known — knowledge [wired] (These consumers do not grant instruction priority or factual warrant.); Stored chunk vectors at query SQL: known — ranking [wired] (No retained ranking policy or trained scorer is required for ranking-input authority.); Chunk parent IDs at query root-reload consumer: known — routing [wired] (This wired lookup effect is separate from canvas author filtering and numerical candidate ranking.); Root author fields at canvas projection-input filter: known — routing [afforded] (The canvas route's database, representation and filename mismatches retain an afforded warrant; the wired query reload does not upgrade it.); Canvas retained text and link fields at browser display: known — knowledge [afforded] (Requested and generated filenames differ; no connected display run is established.); Potential Atlas-retained payloads and requested topic outputs: uninspected (The inspected request does not establish remote persistence, behavioral authority or later consumers; exporter initialization/schema defects also qualify reachability.)	AAS-2026-10-17-semantic-engine-666c064cd8dc-01
Semiont	2	Included: accumulated resource files, resource/annotation event history and current views, annotation prose/anchors/tags/bindings and registered vocabulary, graph projection, numerical vectors and readable payload, anchored extraction text/declines, remembered indexing outcomes, Weaver and job recovery state, their admission/maintenance and later browse/search/gather/generation routes, and afforded aggregate composition/read-back. Unresolved included alternatives: direct-input human control and initial provenance, arbitrary uploaded formats, full Browser/desktop lifecycle, offline restoration, remaining backend/host paths. Excluded: static shipped prompts/configuration as memory, request-local gathered prose/scores and compiler repair state unless retained for a later consumer, external assistant conversation memory, provider-internal retention/training and external weights/corpora. No universal corpus or deployment coverage is asserted.	partial coverage (Positive store classes cover the named inspected objects, not every shipped host or backend. No trained weights are included: externally served inference parameters are excluded from this accumulated-knowledge boundary.); Resource content, JSONL record history, materialized views, extracted-text cache and Weaver sequence marks: known — files [wired] (Inspected filesystem-backed objects are covered. Staging is not a Git commit, so repo is not inferred from a working-tree location.); Retained graph projection: known — graph [wired] (Logical graph substrate is established; live backend correctness is not.); Embeddings, readable point payload and remembered extraction declines: known — vector [wired] (Classifies the operative store, not the encoding of each field.); Retained job inputs and retry/checkpoint records: known — service-object [wired] (Queue retention is finite but spans later worker attempts.); Other backend implementations and Browser/desktop retained-state alternatives: uninspected (These records bound their backend/recovery coverage; the reports do not establish all alternative host-state substrates.)	partial coverage (Automatic admission is established locally. Manual agency is not asserted merely because a UI or a generic caller exists; unresolved direct decisions remain included.); Worker document/annotation commits and software projection/extraction/checkpoint writes: known — automatic [wired] (Starting generation, detection or maintenance does not turn its subsequent admissions into manual writes.); Direct upload, clone/update, body/binding/type edits and aggregate content admission control: not-determinable (Caller identity and assistant/operator wording do not establish an explicit human supplying, editing or approving each retained item. Automatic command execution alone cannot resolve the per-content decision.); Offline restoration and full Browser/desktop write lifecycle: uninspected (Accepted reports leave these alternatives uninspected; ordinary live command gates cannot be generalized to them.)	partial coverage (Both requested knowledge return and automatic maintenance supply exist. A pull witness does not resolve other included consumers or imply pull-only operation.); Explicit browse/search/gather and requested generation-context delivery: known — pull [wired] (Includes resource and annotation focuses and narrower MCP display. Worker clipping is not a second push solely because code delivers the prompt.); Automatic maintenance selection and delivery of event-targeted retained content: known — push [wired] (Push is local maintenance supply, not independently triggered unsolicited foreground-model context.); Explicit history/cache reads within maintenance and recovery: known — pull [wired] (Request and automatic selection are distinct operations within the chain.); Documented aggregate browse/read role: known — pull [afforded] (No deployed external host behavior is inferred.); Full Browser/desktop host refresh and other retained-content consumers: uninspected (Records establish selected callers, not every host reload or unsolicited delivery policy.)	partial coverage (Signals classify only automatic maintenance push. Requested search/gather does not become push because it uses IDs, embeddings or model judgment; host alternatives prevent a complete signal inventory.); Event-triggered maintenance selection of the corresponding retained resource/annotation: known — identifier [wired] (Actual target identity selects the operative part; no universal token budget is asserted.); Startup maintenance inventory and progress/compatibility filtering: known — coarse [wired] (Availability/progress filtering is distinct from targeted event identity selection and from semantic query scoring.); Signals for untraced host push alternatives: uninspected (Known requested IDs, lexical searches and embedding queries are pull selectors, not evidence of inferred-lexical or inferred-embedding push. Other host supply is unresolved.)	partial coverage (Local qualifying yes findings coexist with unresolved other writes. No system-wide no is inferred from fixed model parameters, lack of execution evidence, or lack of demonstrated improvement.); Extraction failure outcome to remembered decline to later startup reconcile: known — yes [wired] (The wired outcome-fed behavioral artifact qualifies without new factual knowledge or measured improvement.); Applied/completed activity to persisted sequence marks or unit/cursor state to later restart/attempt: known — yes [wired] (Recovery behavior is the bounded later effect; no generalized learned strategy or exactly-once guarantee is asserted.); Generated content, graph/vector knowledge transformations and aggregate alternatives as trace-fed learning: not-determinable (Knowledge retention alone does not establish original trace input on every route. Document bytes and source annotations are not automatically sessions/tool traces; aggregate execution and other host updates are unresolved.)	partial coverage (Knowledge, ranking, routing, validation, enforcement and local learning forces coexist at different consumers. No standing instruction authority is inferred from the shipped prompt or an assessing label.); Resource/extracted/passage/assessment prose read by clients and generation: known — knowledge [wired] (Selection and delivery are wired; obedience, factual correctness and benefit are unobserved.); Retained graph descriptors, text, vectors, types and assessment prose at candidate selection and ordering: known — ranking [wired] (Fixed scoring still gives its retained inputs ranking authority; request-local scores are not learned policy.); Sequence marks, job cursors/completed units, extraction stamps and declines at maintenance/retry consumers: known — routing [wired] (These are accumulated operational controls, not static service configuration.); Recorded assignment holder/job identity and registered vocabulary at admission: known — validation [wired], enforcement [wired] (Force is bounded to these checks. Generic direct non-worker admission and offline restoration are not thereby covered.); Extraction outcome at the automatic decline-state updater: known — learning [wired] (Learning authority at the update consumer is separate from routing authority of the resulting decline; no capacity improvement is demonstrated.); Authority at untraced host, restoration and provider consumers: uninspected (Existing records do not exhaust every later consumer. Assessment prose does not by itself establish a standing instruction or a factual veto.)	AAS-2026-10-15-semiont-d516c7013d8f-01
Sentry Skills	2	Included: use-revised skill instructions/metadata and routed references; working/holdout examples, findings, SOURCES and SPEC read by later maintenance; generated project guidance; saved document and Notion Work Summary handoffs; opaque AXIS reports/baselines and unresolved reply/leaf consumer alternatives. Excluded: unchanged shipped prompts/configuration as such, ordinary current-window tool results, host frame state, provider weights, application code/status effects without an established memory consumer, and final SRED outputs without a subsequent agent route. External retention and loading remain unresolved included mechanisms, not claims to cover enclosing hosts.	partial coverage (Filesystem/repository and prescribed Notion witnesses coexist with opaque host artifacts, AXIS retention and unenumerated leaf paths. External model weights are not an accumulated-memory object established by these records.); Changed SKILL.md bodies, routed references and maintenance SOURCES/SPEC files: known — files [afforded], repo [afforded] (The inspected Git-backed files and prescribed revisions provide filesystem and repository retention; no exercised update is supplied.); Working/holdout examples and interpreted findings retained for later revision: known — files [claimed], repo [claimed] (Markdown paths on filesystem/Git are prescribed; populated evidence files are not supplied.); Generated consuming-project AGENTS.md guidance: known — files [claimed], repo [claimed] (Repository filesystem/Git is the specified destination; CLAUDE.md is an alias, not separate storage.); Work Summary consumed by the later SRED organizer: known — service-object [claimed] (A private Notion document is prescribed. Link/date/count metadata does not reveal the linked payloads or service internals.); Saved coauthored document consumed by a fresh reader: partial — files [claimed] (The host-artifact alternative is included but its physical substrate is not established; the file witness cannot classify it.); AXIS reports, raw transcripts, generated artifacts and comparison baseline: not-determinable (External AXIS owns retention; accepted annotations leave storage path and payload uninspected. Configuration wiring does not identify the result substrate.); Persisted review-thread reply with unresolved later agent receipt: not-determinable (The helper wires publication, but service retention details and a later consumer route are unresolved; remote mutation alone does not establish a scoped memory substrate.); Other included skill/reference maintenance and later-consumer paths not enumerated in the accepted records: uninspected (The recorded instruction container and revision route do not enumerate every domain leaf or admission/read path. Unread leaves remain included; complete package-memory coverage is prevented.)	partial coverage (Model admission is automatic despite user-started tasks. Explicit holdout, adoption, baseline-refresh and content-review decisions are manual. Unknown initial controls and generic callers remain unresolved.); Model admission of package guidance and evidence deltas without per-content human approval: known — automatic [claimed] (Starting the workflow is not per-change approval. The author chooses deltas/promotion or deferral under the prescribed policy, not an implemented automatic ingester.); Explicit holdout-to-working reassignment: known — manual [claimed] (The user must explicitly decide to move those examples; this is separate from ordinary model revisions.); Maintainer adoption and baseline replacement: known — manual [claimed] (Maintainer reviews results and chooses adoption or baseline refresh; configured generation/scoring does not automatically install a revision.); Project guidance model admission: known — automatic [claimed] (No per-rule human approval is required; host tools and repository review are external controls.); SRED summary grouping/drafting before human classification review: known — automatic [claimed] (Model groups and drafts content; user query initiation does not supply a per-group admission decision.); SRED human classification adjustment and reviewed project content: known — manual [claimed] (Explicit user classification changes and repeated review decide the displayed content, unlike service physical I/O.); Coauthored draft human content curation: known — manual [claimed] (The user curates content, selects changes and can finalize or decline the structured workflow.); Model-directed reader-gap repair: known — automatic [claimed] (Foreground proposes and revises problematic sections in the prescribed route; this is separate from human content selections.); AXIS retained report admission and initial baseline control beyond reviewed refresh: not-determinable (External result generation is configured, but exact retained-report admission and initial baseline mechanism are not established.); Reply publication caller content admission: not-determinable (CLI caller identity alone does not decide manual versus automatic admission. The helper performs physical publication, not content choice.); Other included skill/reference maintenance and later-consumer paths not enumerated in the accepted records: uninspected (The recorded instruction container and revision route do not enumerate every domain leaf or admission/read path. Unread leaves remain included; complete package-memory coverage is prevented.)	partial coverage (Both requested retention reads and declared/afforded host guidance pushes exist. Host implementation, AXIS alternatives, reply receipt and unread leaves prevent a complete direction inventory.); Later skill host relevance activation of changed installed guidance: known — push [claimed] (Automatic relevance activation supplies changed guidance without a separate request for that retained text. Host activation is declared, not observed.); Future project host loading of generated root/nested guidance: known — push [afforded] (Standing project defaults are intended to enter later host guidance automatically; actual loading and precedence remain external.); Later maintenance intake of prior SPEC/SOURCES/examples and requested step references: known — pull [afforded] (Synthesis/revision agents request relevant retained material by maintenance task/current step; returning those requested files is not another push.); User-identified Work Summary requested by organizer: known — pull [afforded] (The later organizer asks for and reads the prior summary link and related source content.); Requested saved-draft reader test, including human fresh-chat transfer: known — pull [afforded] (Invoker requests comprehension of the selected document/question. Delivery in the requested test does not become unsolicited push; the manual fresh-chat alternative remains included.); Later maintainer request for AXIS reports and baseline comparison: partial — pull [claimed] (Actual harness payload delivery and additional automatic report routes are not inspected; the requested review witness does not resolve them.); Later receipt of stored review-thread replies: not-determinable (Future caller is not established; feedback first-comment projection and self-comment omission prevent inferring a reply read-back route from persistence.); Other included skill/reference maintenance and later-consumer paths not enumerated in the accepted records: uninspected (The recorded instruction container and revision route do not enumerate every domain leaf or admission/read path. Unread leaves remain included; complete package-memory coverage is prevented.)	not-determinable coverage (No controlled push selector is established in accepted records. Host relevance/scope declarations do not reveal selection rules; unresolved alternatives prevent a pull-only or absent boundary assessment.); Automatic relevance selection of changed installed skill guidance: not-determinable (Trigger is a relevant task; selected retained part is updated SKILL.md/references. The records do not establish selector implementation/input matching, so neither inferred-judgment nor identifier can be assigned.); Automatic later host loading of generated project guidance: not-determinable (Trigger is later project work, selected part is root/nested AGENTS.md. Scope names and a closest-file policy do not establish an actual identity-match selector or its host wiring.); Requested maintenance, SRED and reader-test delivery operations only: inapplicable (These scoped operations are pull: a consumer/invoker requests retained files, summary or draft. Push-signal classification is inapplicable to these operations, not to the entire package.); Opaque AXIS automatic delivery alternatives and unresolved review-reply/leaf read-back: not-determinable (Read direction/selection for these alternatives is unresolved. Known pull review or a thread identifier cannot make their push signals inapplicable.)	partial coverage (The prescribed skill-evidence chain supports a local claimed yes independently of observed capacity improvement. No bounded absence record supports no; other write chains and initial origins remain unresolved.); Model admission of behavior-example-derived findings and durable skill rules for later invocation: partial — yes [claimed] (Origin tags include human-verified, mixed and synthetic inputs; actual trace provenance and populated promotion/adoption are unknown. The qualifying route does not imply every example or revision is trace learning.); AXIS transcript feedback through human adoption and opaque report writes: not-determinable (Retained transcripts can guide revision, but the recorded adoption gate is manual and the precise automatic trace-fed admitted artifact chain is not established. Do not infer an automatic installed update from scores.); Project guidance, SRED summaries and reader-gap document repairs: not-determinable (Durable model writes are prescribed, but records do not establish the original input as qualifying activity traces for every such update. Repository facts/shared documents and reader feedback cannot be relabelled wholesale as traces.); Review replies and unenumerated domain-leaf writes: not-determinable (Reply caller origin and later consumption are unresolved; other leaf writes are not enumerated. Neither supports a system-wide no.)	partial coverage (Authority follows each retained part and consumer. Knowledge, instruction, learning, validation and routing are locally supported; helper structure checks and triage approval are not reclassified as retained-memory enforcement.); Revised skill rules and project guidance at later task-agent consumers: known — instruction [afforded] (Later hosts deliver standing directions; delivery affords instruction authority, not compliance.); Retained behavior examples and findings at the instruction-revision consumer: known — learning [claimed] (Experience and interpreted deltas are prescribed inputs to durable guidance updates. Learning authority here does not establish improved future capacity.); SOURCES and evidence explanations at future synthesis/maintenance agents: known — knowledge [afforded] (These supply provenance, context and suspected causes. Ordinary task consumers need not receive the reasons behind promoted rules.); SPEC scope/output/validation contract and held-out expected behavior at maintenance/replay judgment: known — validation [claimed] (Retained contracts and example-specific reference outcomes guide conformity judgments; they are not a machine-enforced repository admission gate.); Changed skill metadata at external host selection: known — routing [afforded] (Host-consumed invocation metadata and descriptions select guidance or an invocation; host selection details and actual triggering are not inspected.); Retained Work Summary at organizer and saved draft at fresh reader: known — knowledge [afforded] (Summary supplies evidence for project analysis; draft supplies the content being understood and checked. Neither is a package-level standing policy update.); AXIS retained reports/baseline at later maintainer comparison/adoption: partial — knowledge [claimed], learning [claimed] (Opaque baseline/report parts and actual payload consumption prevent assigning every local effect. A score is not itself a deployed enforcing control.); Published review reply at a later repair-agent consumer: not-determinable (Advisory remote publication is wired but future receipt is unestablished; do not infer knowledge authority at a hypothetical reader.); Other included skill/reference maintenance and later-consumer paths not enumerated in the accepted records: uninspected (The recorded instruction container and revision route do not enumerate every domain leaf or admission/read path. Unread leaves remain included; complete package-memory coverage is prevented.)	AAS-2026-10-14-skills-2719567f917c-01
Siftly	2	Retained imported bookmarks/rawJson, extracted entities, image and semantic metadata, category guidance and memberships, processing markers, FTS5, hydrated search cache, updated configuration and derived exports; their write, maintenance and later-consumer routes. External CLI context/checkpoint and provider-retention alternatives remain included but uninspected at Siftly adapters. Excludes static instructions as independent memory, ordinary current-run queues/progress, executor weights without established use-dependent updates, remote service internals beyond the adapter boundary and hypothetical external helper callers.	partial coverage (The collection, access index, process cache and filesystem export have established substrates. Included external retained state is opaque; remote model weights are executors, not an established use-accumulated memory store.); SQLite imported content, entities, image/semantic metadata, category guidance, memberships, processing markers and FTS5: known — sqlite [wired] (Each named local retained part is stored in SQLite. FTS is a virtual table, not a vector or graph store.); AI search hydrated-response cache: known — in-memory [wired] (This Map persists across requests within one process, not restarts.); Obsidian Markdown notes and indexes: known — files [wired] (The exporter writes independent filesystem artifacts; caller persistence of download responses is not assumed.); Downloaded CSV/JSON/ZIP response copies after HTTP delivery: not-determinable (Response formats are inspected, but whether and where a caller retains downloaded files is not supplied.); Included external Claude/Codex retained context, checkpoints and provider-owned state at the adapter boundary: uninspected (No external CLI checkpoint or provider retention substrate was inspected; endpoint execution alone does not establish a service-object memory store.)	partial coverage (Automatic admission is wired for transformations and copies. Explicit human content decisions are afforded in file supply and category selection/editing, but generic HTTP principals and initial deployment content remain unresolved.); Pipeline extraction, vision/semantic/category writes, default seeding, FTS/cache writes and export formatting: known — automatic [wired] (An operator starting processing/export is not per-content approval of generated additions or formatted copies.); Browser file selection and category suggestion/manual-assignment UI: known — manual [afforded] (The UI affords an explicit human content-supply or selection decision; this is not evidence of exercised human review.); Direct generic import/category/suggestion/assignment/settings callers and initial settings: not-determinable (An HTTP caller may be human or software. isAiGenerated=false and confidence=1.0 do not establish an explicit human decision; the additional direct category writer remains incompletely assessed.); Included external Claude/Codex retained context, checkpoints and provider-owned state at the adapter boundary: uninspected (No admission mechanism for external retained context/checkpoints was inspected.)	partial coverage (Request-directed query delivery and automatic context supply inside processing are separate operations. Opaque host retention prevents declaring a complete local-plus-host direction inventory.); Requested search/cache hit, browse/graph, CLI detail/list and export return: known — pull [wired] (Automatic scoring inside a requested retrieval does not make the requested return another push.); Application-selected retained tags/entities/category guidance delivered into enrichment and categorization calls: known — push [wired] (The pipeline trigger starts work; application code then selects retained context for models without those models requesting its retrieval. This is not an unsolicited external-agent injection.); External reading of exported response files or vault artifacts: partial — pull [afforded] (Export affords later reader access, but no deployed vault or agent read trace establishes the full external delivery chain.); Included external Claude/Codex retained context, checkpoints and provider-owned state at the adapter boundary: uninspected (Host context/checkpoint delivery is not inspected; local requested routes cannot resolve it.)	partial coverage (Classify only the pipeline push selector, not a requested query return. Known pull-only local consumers are inapplicable units; opaque host alternatives remain included.); Pipeline automatic loading of retained per-bookmark media/tags/entities into worker and category batches: known — identifier [wired] (Trigger: pipeline worker or queued category flush. Selector inputs: chosen database bookmark IDs and linked media rows. Selected parts: retained text/entities/imageTags/semanticTags for those matching rows. Field truncation limits delivery after the identity selection.); Pipeline available-category context and incremental eligibility: known — coarse [wired] (Trigger: pipeline start/work selection. Inputs: available categories, explicit/force versus missing enrichedAt, null/non-null tag markers. Parts: category descriptions/slugs supplied as the available set and eligible bookmark content. These broad predicates are distinct from matching chosen bookmark identities.); Requested AI search/cache, browser, CLI and export returns: inapplicable (These inspected delivery operations fulfill consumer requests. Lexical and model relevance selectors inside that pull chain are not classified as push signals.); Included external Claude/Codex retained context, checkpoints and provider-owned state at the adapter boundary: uninspected (No trigger, selector input or delivered retained host part is known; cannot assign lexical, embedding or judgment push signals from external execution alone.)	not-determinable coverage (No accepted bounded evidenced-absence record licenses a local or system-wide no. Operational markers have explicit future selectors but do not establish a qualifying trace-fed durable behavior-shaping artifact; imported posts and generated annotations do not alone establish trace learning. Opaque hosts remain unresolved.); Local imported-content transformations, category guidance, markers, caches and export routes: not-determinable (The records distinguish imported source content from Siftly execution traces and report only a bounded feedback search, explicitly not an absence record. Marker/assignment persistence and topic generation do not by themselves establish qualifying experience learning. No yes or no is warranted.); Included external Claude/Codex retained context, checkpoints and provider-owned state at the adapter boundary: uninspected (External conversations, checkpoints and parameter-update internals are uninspected; no automatic trace-fed write/later-consumer chain can be classified.)	partial coverage (Local consumed parts establish context, standing classification guidance, retrieval priority, routing and admissibility constraints. External retained-state effects and downstream vault/agent behavior remain unresolved.); Retained source text, derived metadata and response explanations delivered to models or query callers: known — knowledge [wired] (Evidence-like context is a consumption role, not verified truth or general tool authority.); Category descriptions delivered to subsequent categorization models: known — instruction [wired] (Standing organizational guidance is distinct from the bookmark content it classifies.); Retained text/metadata, FTS tokens and membership confidence at search and graph ordering: known — ranking [wired] (Fixed-code scoring and model relevance interpretation both consume retained inputs; ranking does not imply learned ranking policy.); Category slugs at assignment parser and retained candidate IDs at result hydration: known — validation [wired], enforcement [wired] (These retained identities are consumed as allowed-domain criteria and enforced admission restrictions.); enrichedAt, non-null/empty markers and retained configuration at future work selection and dispatch: known — routing [wired] (Markers can suppress retry without establishing annotation quality. Configuration chooses executor/location rather than teaching a model.); Exported content at an external vault or agent consumer: partial — knowledge [afforded] (The delivered artifact affords reference use, but the external consumer and any stronger authority are not observed or fully inspected.); Included external Claude/Codex retained context, checkpoints and provider-owned state at the adapter boundary: uninspected (No retained host payload, update consumer or downstream force was inspected; learning authority cannot be inferred from model execution.)	AAS-2026-10-18-siftly-b81c94cf0f77-01
Signet	2	Accumulated captured/caller/imported evidence and access structures; semantic graph and curated head/projection; maintenance history/runbook; host compaction/recovery and hierarchy/thread-head alternatives; edited identity/authored working memory; retrieval vectors and feedback state; retained benchmark phase/results and unresolved optimization, provider/native-source, coordination and backup/recovery routes. Excludes unchanged shipped instructions, current-invocation prompts/staging buffers, external model weights/host internals, and secrets merely stored as opaque nonsemantic payloads.	partial coverage (Named SQLite, graph, vector and file witnesses do not enumerate all retained provider, artifact, checkpoint and coordination substrates. Transient prompts and staging buffers are excluded.); Remembered rows and semantic aggregate notes: known — sqlite [wired] (SQLite backing is established for these retained rows.); Canonical session transcripts and completed session content: known — files [wired], sqlite [wired] (Files and database content are separate retained parts.); Curated head, revisions, entries and filesystem projection: known — sqlite [wired], files [wired] (Database retention and file publication have different failure boundaries.); Entity/aspect/attribute state and projections: known — graph [wired], sqlite [wired] (Graph is the operative access structure; SQLite is its backing substrate.); Dense retrieval representations: partial — sqlite [wired], vector [wired] (Stored vectors are known; every alternative provider/index implementation is not.); Maintenance history and feedback events/weights: known — sqlite [wired] (Both named persistent stores are database-backed.); Edited identity and authored working-memory files: known — files [wired] (Only accumulated edits and authored working memory are included, not unchanged installation prompts.); Host compaction summary rows and alternative artifacts/checkpoints: partial — sqlite [wired] (Summary insertion into the memory store establishes SQLite for the row; artifact, snapshot and hierarchy backing remain incompletely recorded.); Imported artifacts and benchmark checkpoints: not-determinable (Logical retention is established, but complete physical storage/serialization for these parts is not recorded.); Other provider/native-source retained state, hierarchy/thread heads, backup and coordination state: uninspected (Included alternatives are mapped but storage and recovery details remain unread. Secret storage alone is excluded from semantic memory; accumulated coordination state remains unresolved.)	partial coverage (Automatic admission has concrete witnesses. Arbitrary caller control, identity editing and recovery alternatives remain unresolved; no explicit human per-content decision is recorded strongly enough to add manual.); Terminal capture, history compression, Dreaming proposals/head finalization and aggregate-note admission: known — automatic [wired] (Software/model admission does not await human per-content approval.); Document importer and host-summary persistence: partial — automatic [wired] (Inspected daemon persistence is automatic; upstream host summary production and all importer alternatives remain unresolved.); Direct remember, modify/withdrawal and caller-rating submission: not-determinable (Generic HTTP caller origin does not resolve human supply/approval; downstream feedback propagation has its own mechanism.); Feedback weight/statistic propagation: known — automatic [wired] (This is distinct from unresolved upstream rating authorship.); Benchmark checkpoint phase update: known — automatic [wired] (Orchestrator, not a per-result human decision, writes completion/failure state.); Identity initialization/editing, source synchronization, backup replacement, other adapters and hierarchy writes: uninspected (Initial content approval and alternative admission controls remain uninspected; no manual finding is warranted from authored filenames alone.)	partial coverage (Both requested read-back and unsolicited selection exist. Pi/history witnesses do not resolve every shipped adapter or provider route.); Explicit Pi/MCP/CLI recall and maintenance zoom/evidence requests: known — pull [wired] (The consumer requests retained material; returning it is not a second unsolicited push.); Pi startup automatic background selection and hidden context delivery: known — push [wired] (A lifecycle request triggers automatic selection of accumulated context, not an item-specific memory request.); Scoped history injection into later maintenance prompts: known — push [wired] (Initial history is automatic; zoom is separately requested.); Saved aggregate notes at later ordinary recall: known — pull [wired] (Immediate unsaved synthesis is not accumulated-memory read-back.); Benchmark checkpoint phase reload on resume: known — pull [wired] (The resumed orchestrator reads retained phase status rather than receiving independent background injection.); Other hosts, provider/native-source consumers, hierarchy/thread heads and backup/coordination consumers: uninspected (These included read-back routes remain untraced; object storage alone supplies no direction.)	partial coverage (Recorded push selection supports coarse gating and scope-identifier history matching. Explicit recall query selectors do not classify startup push. Missing startup predictor and alternative-selector facts prevent complete signal classification.); Startup ranked memories and graph constraints: partial — coarse [wired] (Agent/project policy, scoring, focal entities and budgets are recorded, but all ranking/predictor internals are not. No lexical/embedding startup finding is inferred from explicit-recall capability alone.); Current working-memory/identity sections selected at startup: known — coarse [wired] (Configured inclusion, freshness and budget supply whole retained sections; a path alone does not establish targeted identity matching.); Maintenance history stream selection and folding: known — identifier [wired], coarse [wired] (Trigger is new pass construction; allowed stream scopes select history nodes, then level/byte-target folding limits the delivered lines.); Checkpoint recovery, continuity, inherited context and predicted startup context: not-determinable (Startup includes these alternatives, but complete selectors/input lineage are not recorded. Session labels alone do not establish an identifier-selection finding.); Explicit recall/zoom and benchmark checkpoint reload: inapplicable (These named operations are pull; push-signal classification does not apply to their requested return. Maintenance initial injection remains a separate applicable unit.); Other adapter/source/hierarchy automatic deliveries: uninspected (Untraced alternatives cannot be declared pull-only or signal-inapplicable.)	partial coverage (Four wired trace-fed update/continuation chains establish yes locally, including host compaction summaries. No bounded absence record warrants no for unresolved routes, and no observed benefit or learned provider parameters is claimed.); Captured session evidence to semantic/head maintenance to later context: partial — yes [wired] (Automatic evidence-fed semantic/head writes have later context consumers; actual improvement and every alternate evidence source remain unresolved.); Pass/tool activity to history leaf/parent lines to later maintenance prompt: known — yes [wired] (This generated continuation-context chain qualifies without proving new knowledge or improved decisions.); Session retrieval records plus ratings to aspect weights to later traversal: partial — yes [wired] (Automatic propagation is known even though upstream rating agency/origin and other statistics consumers remain unresolved.); Host compaction summary and recovery/hierarchical alternatives: partial — yes [wired] (Host compaction produces a retained continuation summary and automatic daemon ingestion makes it available to later memory/head consumers. Original turn types, surviving tail and hierarchy alternatives remain unresolved.); Aggregate recall from mixed evidence, caller memory, imported sources and benchmark/optimization alternatives: not-determinable (Automatic aggregate admission is known, but a qualifying original trace input is not fixed for all saved notes. Benchmark retention alone does not establish a behavior-shaping update; optimization alternatives remain unread.); Other hierarchy, identity-maintenance, source-provider and coordination update routes: uninspected (Neither qualifying writes nor bounded absence were established for these included alternatives.)	partial coverage (Authority is consumer-relative: knowledge/instruction, ranking, learning inputs, enforcing validation and checkpoint routing have separate witnesses. Untraced consumers prevent a whole-boundary union from being complete.); Retained content at foreground lookup/startup and maintenance evidence consumers: known — knowledge [wired] (Delivered text supplies reference/context, not host action permission.); Accumulated identity edits at foreground guidance consumer: partial — instruction [wired], knowledge [wired] (Only edited/maintained material is scoped; static shipped templates are excluded.); Evidence text and access structures at retrieval selectors: partial — ranking [wired] (Text, vectors, graph features and localized weights influence selection; complete alternate ranking policies remain unresolved.); Eligibility/provenance/scope state at head and current-memory admission/read consumers: known — validation [wired], enforcement [wired] (Checks have bounded correspondence and eligibility force; successful NOOP is not fresh support validation.); Trace-derived evidence/history and feedback at durable update consumers: known — learning [wired] (Learning authority names the update input, separately from downstream knowledge/ranking and without demonstrated improvement.); History and recovery material at subsequent model invocations: partial — knowledge [wired] (Selected summaries supply context; no blanket standing-instruction classification follows.); Retained checkpoint phase status at benchmark resume consumer: known — routing [wired] (Completed evaluations are skipped; judge explanations are not established product-memory validation criteria.); Untraced provider, hierarchy, identity-propagation, backup and coordination consumers: uninspected (Force cannot be inferred from storage or opaque payload labels.)	AAS-2026-10-14-signetai-64f703d83203-01
SkillLift	2	Use-accumulated skill bodies and auxiliary content, candidate reconstruction records, search history, receipt rules and rationale, revision diagnostics, verifier scores, trial recovery records, category and cluster checkpoints, trained cluster artifacts and fused final guides. Includes unresolved raw-trajectory consumers and other shipped baseline/entry-point alternatives; their missing records prevent complete coverage. Excludes static shipped prompts, transient request views as independent stores, external benchmark/host memory and opaque provider weights: no accepted record establishes those weights as use-accumulated SkillLift memory.	partial coverage (Files are wired across the assessed routes; remaining shipped memory alternatives prevent a complete substrate inventory.); Task skill trees, candidate recipes, search state and trial recovery records: known — files [wired] (Filesystem retention is established for these parts; a Git dependency used to apply patches does not make the store a repository substrate.); Receipt rules, rationale, revision attempts and persisted verifier scores: known — files [wired] (Python schema objects are working copies of these retained files, not an independently established in-memory persistence system.); Category snapshots/final trees, domain checkpoints/cluster artifacts/master bundle and behavior diagnostics: known — files [wired] (Full snapshots differ from task replay but share its file substrate.); Raw trajectory payloads and remaining shipped baseline/standalone memory alternatives: uninspected (Raw trajectory serialization/readers and wider algorithm stores are not fully established by accepted records; inspected file witnesses do not enumerate their operative parts.)	partial coverage (Automatic subsequent admission is wired, while initialization and alternatives remain unresolved.); Task proposal retention, patch/successor admission, search-state update and receipt validation/alignment admission: known — automatic [wired] (Starting/configuring the run is not manual content admission; saving failed proposals is separate from promoting a successor.); Revision diagnostics, behavior reducers and trial/score recording: known — automatic [wired] (These automatic diagnostic writes do not all qualify as behavior-shaping learning.); Category/domain updates, snapshot admission and final fusion/materialization: known — automatic [wired] (A human domain/configuration choice is not approval of each rule or snapshot.); Initial supplied seed/checkpoint content and alternate/custom admission paths: not-determinable (Interfaces do not establish whether an explicit operator decision supplied/approved each initial object; constructor injection and remaining shipped routes lack complete admission evidence. No manual value is inferred.)	partial coverage (Assessed operations are requested pull chains; unresolved consumers prevent known or absent push at system scope.); Task replay, planning, receipt scoring/revision, editor inputs and logical-cell recovery: known — pull [wired] (Task IDs, accepted IDs and model file selection do not make requested read-back push.); Category resume/fusion/test and domain checkpoint/fusion/held-out consumption: known — pull [wired] (Category uses winners; domain loads all current cluster skills. Neither establishes unsolicited foreground retrieval.); Raw trajectory/archive consumers, auxiliary host loading branches and remaining shipped alternatives: uninspected (No complete accepted consumer inventory resolves these alternatives; pull witnesses cannot establish a repository-wide pull-only boundary.)	uninspected coverage (No positive push signal is established; unresolved alternatives prevent making the entire axis inapplicable. Identifiers in requested lookup are not identifier-based push.); Inspected requested task/category/domain read-back chains: inapplicable (These operations fulfill explicit requests; push-signal classification does not apply to their IDs, budgets or optional file selector.); Unresolved raw trajectory, auxiliary loading and other shipped read-back alternatives: uninspected (Missing consumer/selector facts prevent determining whether a push exists and what signal selects retained parts.)	partial coverage (Several qualifying automatic trace-fed durable-text/control updates are wired. Remaining consumers are unresolved; no global negative or capacity-improvement conclusion follows.); Task scalar evaluation/proposal experience to retained search history/constraints and later planner: known — yes [wired] (A task-scoped durable behavior-shaping artifact qualifies without new knowledge, provider weight change or observed benefit.); Recorded oracle/verifier contrast to revised receipt and later scoring/refinement: known — yes [wired] (Scalar experience is sufficient; trajectory exclusion does not negate this route.); Category/domain evaluation-responsive updates, snapshots and later training/final consumption: known — yes [wired] (Category and domain gates differ; no observed improvement or causal transfer is claimed. Fusion consumes retained training products, not established raw trajectories.); tau2 behavior-summary to policy, raw trajectory readers, diagnostic archive reuse and other shipped alternatives: not-determinable (Shared revision input excludes summary metadata, but unresolved other consumers and raw readers prevent a bounded no for the included alternative inventory. Diagnostic retention alone is insufficient.)	partial coverage (Instruction, context, update, ranking, routing and veto effects have distinct actual consumers. Unresolved diagnostic and alternative routes prevent a complete authority inventory.); Retained skill bodies at later benchmark agents: known — instruction [wired] (Wiring establishes instruction delivery, not actual agent compliance.); Retained skill bodies and plans/rationales at planning, editing and revision consumers: known — knowledge [wired] (Knowledge means reference/context force here, not accepted factual truth.); Receipt criterion text and signed points at verifier and score/selection consumers: known — instruction [wired], ranking [wired] (Language interpretation and fixed arithmetic are separate mechanisms; neither certifies task quality.); Persisted score/history/direction experience at planning and receipt-update consumers: known — learning [wired], ranking [wired] (Learning authority is the update-input effect, independently of downstream instruction authority and without demonstrated capacity improvement.); Accepted IDs, checkpoints, trial results and attempt markers at continuation/recovery consumers: known — routing [wired] (Checkpoint contents route continuation; merely supplying a run-root path does not constitute manual admission.); Retained hashes, validity, scopes, novelty exclusions and identity fields at replay/admission consumers: known — validation [wired], enforcement [wired] (These are retained controls, not attribution of every static validator rule to memory. Domain checkpoint loading lacks category's configuration check.); Behavior summaries, full revision archives beyond repair, auxiliary host payloads and wider alternative consumers: not-determinable (Diagnostic summary metadata is excluded from shared receipt input; archive rereading and other consumers are unresolved. Retention alone establishes no authority, and no bounded global absence is warranted.)	AAS-2026-10-14-skilllift-0f350bf64418-01
SkillNote	2	Accumulated current skill rows, content history and staged inbound bodies, independently published archives, installed project/user copies, browser-local fallback, comments, ratings, usage/call events and watcher state; their admission, maintenance, selection and later-consumer routes. Includes opaque archive payloads, arbitrary extra frontmatter and usage metadata as unresolved parts. Excludes shipped static authoring/adapter policies as memory objects, ordinary transient hook/bundle responses, and external-host weights and compaction state. Static policies remain evidence of memory routes.	partial coverage (Database and filesystem witnesses coexist with browser/provider storage uncertainty. A Git origin does not make the accumulated registry a repo substrate.); Registry current rows, snapshots/staging, release metadata, comments, ratings and usage/call events: known — rdbms [wired] (The declared stores cover the named database parts, not every analytics consumer.); Published archive bytes, installed skill copies/manifests and watcher cursor state: known — files [wired] (Unknown archive content does not obscure the inspected archive storage substrate.); Browser-local editor fallback and remote claude.ai copies: not-determinable (localStorage persistence and remote upload are established, but records do not resolve their implementation to a controlled substrate. Browser storage is not classified as files or kv by assumption; remote provider storage is uninspected.)	partial coverage (Manual affordances and automatic admission coexist; generic caller control remains unresolved rather than silently assigned to either.); Human editor supply and approval-following skill-push publication: known — manual [afforded] (This does not classify all POST/PATCH callers as human-controlled.); Snapshots, hook sync, code-governed imports/staging/adoption, adapter updates, installed-copy replacement and trace/feedback software admission: known — automatic [wired] (Automatic admission is independent of who originally authored the body or report. Ask-policy resolution has a separate available human surface, not a universal requirement.); Generic initial direct writes, caller archive proposal and explicit conflict-resolution caller control: not-determinable (Caller-supplied content or resolution choice alone does not establish a human decision. Human UI resolution is available, but no mandatory human decision is established for every such call.)	partial coverage (Pull and push are both supported on different operations, not by counting a requested response twice.); Requested registry/history/editor reads, MCP skill calls, archive install/update, conflict resolution and context-bundle requests: known — pull [wired] (HTTP/code return routes are wired; downstream host use and deployed resolver wiring remain separate.); Hook-triggered project synchronization toward later host discovery: known — push [wired] (The fetch inside this chain is requested, but the automatic selection/supply toward host discovery is push; it does not prove model-window injection.); Foreground OpenClaw body loading and symbolic scanner/maintenance state reads: known — pull [claimed], pull [wired] (Request-driven state access is separated from any subsequent automatic distribution.); PostCompact reminder delivery, remote-host later consumption and untraced analytics readers: not-determinable (Catalog fetch/output is inspected but model-visible PostCompact delivery is unresolved; remote discovery and full reporting consumers are uninspected. These alternatives prevent complete direction coverage.)	partial coverage (Known push selectors are retained; neither a known pull route nor the supplied foreground policy resolves other alternatives.); Automatic project-sync collection selection and unfiltered wildcard supply: known — identifier [wired], coarse [wired] (Selection is toward host discovery; subsequent model choice is not established by the sync selector.); Requested MCP/archive/bundle retrieval and foreground body reads: inapplicable (These are pull operations. Slug lookup, bundle ranking and task-description judgment here do not establish an additional push signal.); Unresolved reminder/remote/analytics delivery alternatives: not-determinable (Missing downstream delivery/consumer facts prevent determining whether further push operations and signals exist.)	not-determinable coverage (No accepted record establishes a qualifying automatic trace-fed durable artifact update with its later consumer. Uninspected alternatives and lack of bounded absence records prevent a system negative; learning benefit is independently unobserved.); Watcher/hook analytics, evaluation reports and caller usage/comment/rating chains: not-determinable (Automatic trace extraction and retained reports are established. Watcher events serve analytics/recovery; bundle aggregates are recomputed rather than written as durable derived behavior-shaping artifacts. Evaluation production/consumers and broader reporting paths remain uninspected, so these records establish neither qualifying yes nor a bounded complete no.); Conversation-to-procedure authoring and opaque imported/published payload production: not-determinable (Repeated-convention authoring can involve session material but the supported human-approval path is not automatic trace learning. Generic callers and independent archive producers have unresolved derivation/control; no trace-fed automatic artifact update is established for them.)	partial coverage (Instruction, knowledge, ranking and symbolic control forces coexist. Feedback ranking alone does not establish a durable learning update.); Retained procedural body at the external task-performing agent: known — instruction [afforded] (The consumer force is instruction, not SkillNote enforcement of task actions.); Retained comments and reports at reporting readers and the documented bundle consumer: known — knowledge [afforded] (Information availability does not verify reported outcomes or establish a deployed resolver.); Usage IDs/count inputs and comment ratings at bundle ordering; skill descriptions at foreground selection: known — ranking [wired], ranking [claimed] (Distinct code and host-policy consumers retain distinct evidence bases.); Collections, active-version/pointer/manifest fields and watcher cursor state at fixed-rule selectors: known — routing [wired], enforcement [wired], validation [wired] (Covers the named control consumers, not arbitrary frontmatter semantics or a universal access-control guarantee.); Opaque archive files, arbitrary frontmatter/metadata and incompletely traced analytics/evaluation consumers: not-determinable (Missing payload/consumer facts prevent complete force classification, including a learning-authority conclusion for evaluations.)	AAS-2026-10-14-skillnote-0ccfe85cda4a-01
SkillRL	2	Included: banks accumulated or changed by offline and online use, derived experience records, their access metadata and dense indexes, retained SFT transformation products, fitted actor parameters/checkpoints and their write, maintenance and later-consumer operations. Initial content provenance and direct API admission alternatives remain unresolved. Excluded: unchanged shipped instructions as read-back, ordinary episode history/current rollout transport as standalone memory, fixed remote producer/encoder parameters and opaque provider internals not shown to accumulate target memory, unrelated runtime storage and external SFT fitting. Raw trajectories remain included as inputs to qualifying updates.	partial coverage (Includes bank, derived experiences, access caches, SFT products and changed actor state. Alternate worker/checkpoint implementations remain incompletely assessed; no whole-runtime storage census is asserted.); Changed skill bodies, access metadata and derived experience records: known — files [wired], in-memory [wired] (The recorded bank and experience objects identify both persistence and live storage.); Derived skills-only and experience access indexes: known — in-memory [wired], files [wired], vector [wired] (Vector is the operative access substrate, not a classification of JSON instruction encoding.); Updated actor parameters and local recovery checkpoints: partial — model-weights [wired], in-memory [wired], files [wired] (The inspected FSDP policy path records live state and checkpoint persistence; alternate serialization details are not fully audited.); Retained SFT intermediate traces, reasoning conversations and final pairs: known — files [wired] (Recorded artifact persistence is known; external SFT fitting is outside this scope.)	partial coverage (Automatic inspected transformations coexist with unresolved generic caller and initial-content control; no manual finding is invented from file loading.); Validation and training skill admission: known — automatic [wired] (Launching/configuring the run does not make later admission manual.); Original offline generation and inspected SFT writes: partial — automatic [wired] (Known admission control on inspected routes; environment-specific repair and generation stages remain incompletely inspected.); Actor optimizer and periodic checkpoint production: known — automatic [wired] (Operator-selected initial model and restart configuration are separate from each optimizer admission.); Derived vector cache production: known — automatic [wired] (No per-vector operator approval is recorded.); Direct bank addition/removal and heuristic experience save callers: not-determinable (Exposed methods do not identify whether their caller makes a human content decision or computational admission; physical I/O alone cannot resolve agency.); Initial bank, experience and actor-content admission: not-determinable (Configured paths identify reads, not the human/computational decision that supplied or replaced retained content.)	partial coverage (Both requested reads and automatic context supply exist at distinct operations. Alternate worker and environment-specific consumption details prevent a complete route census.); Manager retrieval requests and requested structured returns: known — pull [wired] (This internal request operation is distinct from unsolicited delivery to the foreground policy.); Host-selected bank and experience supply to policy: known — push [wired] (Delivery is gated and may omit fields or rules; no foreground arbitrary-memory pull tool is established.); SFT aggregation and distillation file consumers: known — pull [wired] (Requested file loads are pull, not identifier-targeted push merely because paths are named.); Actor parameters for subsequent generation and local resume: partial — pull [wired] (Inspected numerical read routes are established; alternate workers and exact synchronization remain incompletely audited.); Heuristic save future reload capability: known — pull [afforded] (No live saver-to-selector loop or same-instance index update is established.)	partial coverage (Known push selectors retain lexical, embedding and coarse signals. Other environment context handling is incompletely recorded; no identity-match finding follows merely from dyn_ prefixes or file paths.); Skills-only template-mode push at manager reset: known — coarse [wired], inferred-lexical [wired] (Trigger is reset; task text and bank categories select cached bodies for later observations. Search query-type categories are not delivered by the online shared schema.); Skills-only embedding-mode push: known — inferred-embedding [wired] (Dynamic instructions can be omitted; ranking is not evidence of lesson relevance or compliance.); Alternate experience/skill push: known — inferred-embedding [wired], inferred-lexical [wired], coarse [wired] (Reset triggers selection; task description and retained goal vectors/category lists select displayed rules, examples and mistakes.); Other environment-specific context delivery and filtering: uninspected (The accepted rollout and alternate-selector records leave other manager context handling and full WebShop length filtering incompletely inspected; the signals governing any additional selected retained parts cannot be exhaustively classified.); Requested SFT file and actor/checkpoint read operations: inapplicable (These are requested reads, not push selectors, so no push signal applies to these particular operations.)	partial coverage (Positive trace-fed routes do not justify system-wide no, nor resolve unknown initial and caller-driven admissions. Ordinary episode suffix buffers are excluded, not classified as negative trace learning.); Validation and training failure-to-skill writes: known — yes [wired] (Qualifies without observed benefit or a complete reconstructed episode transcript.); Offline experience-to-bank and inspected SFT abstraction/distillation: partial — yes [wired] (External final fitting is not wired here; bank and intermediate consumers are recorded. Other environment transformation semantics remain incompletely inspected.); Reward-driven actor parameter writes: known — yes [wired] (Capacity wiring does not establish actual gains or exact deployed initial weights.); Heuristic experience save API: not-determinable (Generic caller control and the missing saver/index-refresh connection prevent an automatic trace-fed write-to-later-consumer classification for this capability.); Initial content and direct mutation alternatives: not-determinable (No recorded provenance/control establishes whether these alternative admissions are automatic trace-fed learning or nonqualifying supplied content.)	partial coverage (All positive authorities are consumer-specific. Downstream actor-weight force has no supported controlled mapping; incomplete alternative SFT repairs cannot establish further memory authorities.); Skill bodies delivered to action policy: known — instruction [wired] (Standing behavioral guidance, not simulator enforcement; applicability may be omitted.); Experience display fields delivered to policy: known — knowledge [wired] (Only display fields reaching the later policy are classified here.); Bank text and goal vectors at similarity selectors: known — ranking [wired] (A fixed selector can give retained content ranking authority without a learned ranking policy.); Skill access metadata at category/order selection: known — routing [wired], ranking [wired], validation [wired], enforcement [wired] (Fixed initial ID checks are not within-call semantic deduplication or a universal uniqueness guarantee.); Derived experiences, cleaned traces and bank text at automatic SFT transformations: known — learning [wired] (Update consumers differ from downstream instruction consumers; external parameter fitting is not classified as wired.); Rollout examples and rewards at actor fitting: known — learning [wired] (The update input has learning authority; benefit is unobserved.); Changed actor weights at rollout inference: not-determinable (Weights determine generation, but that downstream parametric force is not one of the controlled authority values; learning applies to the experience at its update consumer, not automatically to the output weights.); Existing skill titles at failure proposal model: known — knowledge [wired] (A requested avoidance instruction does not establish actual deduplication.); Caller-directed removal of retained skill IDs: known — enforcement [afforded] (Capability is not wired automatic withdrawal; cached episodes and saved copies may retain the content.)	AAS-2026-10-14-skillrl-0dba0e1c895e-01
SkillWeaver	2	Includes live and checkpointed Python skill bodies, documentation/signatures, test/version/reference metadata, loaded semantic text, raw browser/action artifacts, selected trajectories, synthesis/success diagnostics, retained benchmark results and optional embedding access cache; includes their initial/imported and subsequent write/read paths. Excludes static shipped instructions as memory, provider parameters/internal retention, website/authentication state and external Browser-Use context management. Shipped skill content is an initial input with unresolved provenance, not observed learning.	known coverage (The recorded inventory covers local retained content, diagnostics, scores and optional derived access cache. External model parameters and website state are not learned-library objects.); Skill bodies, embedded documentation, function metadata and semantic text: known — files [wired], in-memory [wired] (Checkpoint files and live strings/dictionaries are separately supported substrates.); Raw observations/actions, selected episode, synthesis/success diagnostics and benchmark result artifacts: known — files [wired], in-memory [wired] (Disk artifacts and transient structures are recorded; evaluator entry defects do not erase local output wiring.); Optional exact-text embedding cache: known — in-memory [wired] (This is a process-local cache, not a persisted vector database.)	partial coverage (Automatic builder admissions and artifact capture are supported. Loading or invoking a generic public method does not establish a human per-content decision.); Model synthesis plus software merge admission: known — automatic [wired] (Success judgment triggers bounded repair and unconditional final merge without per-candidate human approval.); Builder test-counter increments, merge resets and reference-discard initialization: known — automatic [wired] (These software operations alter retained metadata independently of initial file provenance.); Raw artifact/diagnostic recording and benchmark result retention: known — automatic [wired] (Runtime code records results when the relevant paths execute; benchmark enclosing-entry limits remain.); Optional embedding-cache insertion: known — automatic [afforded] (On API invocation software caches returned vectors.); Initial code/documentation/metadata/semantic content, blanket test marking and separately supplied replacement checkpoints: not-determinable (Generic caller/file provenance does not identify the admission principal or prove manual approval. No manual finding follows from selecting a prefix at read time.)	partial coverage (Established native, builder, adapter and benchmark consumer chains support pull. Optional selection/export/reconstruction capabilities lack a recorded later consumer invocation, leaving included direction coverage unresolved. External host retention is excluded, not classified.); Native tasks, later builder iterations and Browser-Use adapter library reads: known — pull [wired] (Requests precede selection and delivery; assembly of requested library context is not an independent push.); Benchmark score aggregation and later prefix reload: known — pull [afforded] (Local score-read/reload code exists; enclosing benchmark remains afforded because of entry defects.); Optional embedding selection, schema exports and artifact reconstruction capabilities: not-determinable (Records establish request/return capabilities but no shipped later consumer invocation. That missing route prevents classification as pull or push; capability shape alone does not establish read-back.)	not-determinable coverage (Established pull units have locally inapplicable push signals. Included capability-only alternatives lack later-consumer routes, so direction coverage is incomplete and axis-wide signal inapplicability is not established. No controlled signal value is supported for the unresolved alternatives.); Established native, builder, adapter and benchmark library/context/result consumer chains: inapplicable (These established consumer chains are request-directed, so push-signal classification is locally inapplicable. Relevance judgment and name matching within a request do not establish push.); Optional embedding selection, schema exports and artifact reconstruction capabilities: not-determinable (No later consumer route is established. Its direction and any push trigger, selector input or selected retained part are unresolved, preventing either a signal value or inapplicability.)	partial coverage (The automatic episode-to-skill route qualifies architecturally without observed benefit. Automatic trace-conditioned metadata also shapes later selection. Arbitrary initial content and unknown external caller mutations do not inherit these findings.); Selected browser episode to synthesized skills/documentation to later native/controller consumers: known — yes [wired] (The durable behavior-shaping chain is wired; actual exercise and improvement are not observed.); Last-action result predicate to persistent test metadata to practice/visibility consumers: known — yes [wired] (Qualifies as durable selection-policy adaptation, not evidence that the named skill executed.); Imported initial content and externally invoked blanket marking/replacement: not-determinable (Missing origin and admission control prevent deciding whether these alternative writes are automatic trace-fed learning routes.)	partial coverage (Actual recorded consumers establish local forces. Executable procedures have operational effects not fully captured by this vocabulary; external planner internals and export-only consumers remain unresolved.); Documentation at relevance/action models and registered controller descriptions: known — knowledge [wired], ranking [wired] (This is advisory content, not an enforced obligation to invoke a skill.); Existing skill source/documentation and retained episode at synthesis: known — learning [wired] (Input experience and existing procedures shape a durable library update, separately from downstream execution.); Semantic text at task proposal and procedural synthesis: known — knowledge [wired], learning [wired] (Read-only semantic context is neither an automatically learned semantic store nor direct native action guidance.); Test/version metadata at visibility, synthesis inclusion and practice scheduler: known — enforcement [wired], routing [wired], ranking [wired] (These are real policy effects even when the stored test label lacks execution coverage.); Selected episode at success judgment and benchmark scores at reporting: known — knowledge [wired] (Evidence/reporting paths do not grant generalized skill correctness.); Executable bodies at native and controller execution consumers: not-determinable (Records establish direct symbolic execution and website/host effects, but not standing instruction, action-permission control or another controlled authority solely from executable availability.); Optional embedding selection over retained description vectors: known — ranking [afforded] (This API is afforded, with no shipped downstream agent caller established.); Export-only tool schemas, unused diagnostics and external planner continuation: not-determinable (No recorded later consumer establishes further authority; external planning is outside inspected adapter internals.)	AAS-2026-10-14-skillweaver-c921cc7a5ed0-01
SkillX	2	Includes accumulated experience inputs, derived plans and functional/atomic skill parts with access metadata, provisional extractor proposals, copied feedback summaries, derived retrieval vectors/cache, and synthetic task/checkpoint state; includes unresolved initialization and remaining bundled library/historical artifact alternatives. Excludes shipped static prompts/tool schemas, ordinary rollout message accumulation, statistics without a later consumer, immediate parser effects, and external base-model weights not established as SkillX memory updates.	partial coverage (Local stores and derived access structures are resolved; remaining included bundled payloads prevent complete inventory coverage. External model weights are not asserted as accumulated SkillX memory.); Unified library content, metadata and imported experience files: known — files [wired], in-memory [wired] (Inspected serializers and input loaders establish files and loaded objects; encoding does not add a database substrate.); Provisional proposals and copied feedback: known — in-memory [wired] (Lifetime is extractor-instance or returned-copy retention, not a persisted file.); Derived retrieval arrays and exact-text cache: known — in-memory [wired], vector [wired] (Dense index is held in NumPy arrays; FAISS-HNSW is only configuration.); Synthetic task results and optional recovery checkpoint: known — files [wired], in-memory [wired] (Manager results and checkpoint JSON are inspected.); Remaining bundled library and historical-result payloads: uninspected (Accepted records inspect only iteration-one shapes and no historical-log update consumer; their remaining operative inventory is not classified.)	partial coverage (Automatic transformations are known; initial and direct-caller admission control prevents a complete automatic/manual classification.); Pipeline and provisional model-proposal admission: known — automatic [wired] (Starting/configuring construction is not per-content human approval.); Copied summary admission: known — automatic [wired] (Tag parsing chooses replacement without a human content decision.); Embedding cache/index admission: known — automatic [wired] (Software stores computed encodings, including zero-vector fallback.); Task generation/checkpoint admission: known — automatic [wired] (Models propose and code parses/filters/saves; operator configuration does not approve objectives.); Input initialization, prior content and direct library/export admissions: not-determinable (Generic caller control does not reveal an explicit human content decision; blocked checked-export preflight supplies no successful gate.)	partial coverage (Pull library service and push internal context/recovery routes coexist; the absence of background injection on the library service is not a system-wide pull-only finding.); Saved-library service task/step requests: known — pull [wired] (Requested selection and returned prompt/material are pull; subsidiary semantic retrieval is not counted as another push. This unit excludes the alternate public rewrite helper.); Alternative public PlanRewriteService.get_pseudo_plan invocation: known — pull [afforded] (Internal progression is wired, but no shipped caller or later delivery is established; this alternative invocation retains afforded rather than service-route wired read-back evidence.); Provisional extraction context across invocations: known — push [wired] (Code supplies proposals to a later extraction model without that consumer requesting stored proposals.); Copied feedback to later extractors: known — push [wired] (Substitution is automatically included in subsequent invocation input.); Task-manager checkpoint recovery: known — push [wired] (Code resumes matching checkpoint state on generation startup; no caller request for its retained entries is recorded.); Remaining sample/historical artifacts and unspecified public API consumers: not-determinable (Storage and selection capability do not establish later delivery for these alternatives.)	partial coverage (Push selectors have coarse and identifier signals; embedding and optional model judgment on requested retrieval do not establish push signals.); Functional proposal-list supply: known — coarse [wired] (Trigger is next plan step; selector supplies the currently available augmented list rather than a targeted semantic match.); Atomic/hybrid proposal-map supply: known — identifier [wired] (Trigger is tool extraction; actual tool-name lookup selects a retained proposal for delivery.); Copied summary supply: known — coarse [wired] (Long-feedback role/length selects replacement; later assembled history carries that replacement, not embedding-ranked memory.); Checkpoint recovery supply: known — coarse [wired] (Generation startup and seed-hash validity select whole checkpoint; processed indices control continuation, not identity-targeted delivered content.); Requested library retrieval: inapplicable (Push signal does not apply to fulfilling explicit task/step requests; cosine retrieval remains a pull selector.); Unestablished remaining artifact/API delivery: not-determinable (No complete direction account licenses a signal or inapplicability for these alternatives.)	partial coverage (Wired yes routes establish trace-fed memory capacity, not observed improvement. Unresolved alternatives remain local; no system-wide no is warranted.); Experience-to-durable library write and later inference consumer: known — yes [wired] (Automatic trace-fed instructional artifacts persist and have a later requested context route; successful learning is not asserted.); Feedback summary and subsequent extraction: known — yes [wired] (Returned-copy retention and later invocation suffice even without a separately saved summary file.); Provisional proposal update and subsequent model invocation: known — yes [wired] (Instance/list lifetime crosses consumer calls; final-library rejection does not erase this path.); Synthetic objectives/checkpoint recovery: not-determinable (Trace-fed objectives persist, but accepted records do not establish task execution or exactly how recovered objective content shapes a later model; symbolic resume progress alone does not settle that relation.); Arbitrary initialization, remaining historical artifacts and opaque service changes: not-determinable (Unknown input origin and absent recorded provider-update internals cannot establish qualifying writes or bounded absence.)	partial coverage (Consumer-specific knowledge, learning, ranking and control coexist. Static schemas/prompts and parser eval are not accumulated memory criteria or authority paths in this scope.); History and summaries at plan/skill update models: known — knowledge [wired], learning [wired] (The same material supplies evidence and feeds durable instruction updates; authority does not certify truth.); Plan and skill text at later extraction and prompt consumers: known — knowledge [wired], ranking [wired] (Reference implementation advice and task-specific plans are knowledge, not standing host policy or callable enforcement.); Skill access fields at eligibility, coverage and replacement consumers: known — routing [wired], enforcement [wired] (Tools and types constrain actual inclusion; source-task metadata without a reader gets no authority.); Task confidence and checkpoint controls at generation/recovery consumers: known — validation [wired], enforcement [wired], routing [wired] (Threshold and hash checks are consequential but do not validate task solvability.); Dense vectors at similarity selector: known — ranking [wired] (Numerical access state influences relative priority.); Initial/remaining sample material at unestablished later consumers: not-determinable (No recorded consumer chain covers all remaining sample content or arbitrary initial content; exercised authority is not inferred.)	AAS-2026-10-14-skillx-61e3fc8f9aa5-01
smriti-mcp	2	Use-updated Markdown bodies and YAML metadata, NDJSON observations and operation reports, generated Markdown and YAML indexes, and their transient lexical/alias/link access structures; includes caller writes, external-file admission, trace review/consolidation, access reinforcement, lifecycle changes, maintenance and requested reload/delivery. Excludes static server instructions, executable package code, request-local diagnostics, external host/model state and dependency internals. External caller authorship, initial file provenance and eventual host interpretation remain unresolved, not silently excluded alternatives.	known coverage (The object inventory and write/reload/maintenance containers cover the shipped memory surfaces; no included opaque provider store is established. Files and in-memory access structures exhaust this named inventory, not all storage in an enclosing host.); Durable note bodies and metadata: known — files [wired] (The records identify both operative parts; unknown initial authorship does not obscure their substrate.); Raw observations, remembered-content copies and consolidation operation reports: known — files [wired] (The recorded population includes both supplied content and generated bookkeeping; JSON encoding is not a separate substrate.); Saved Markdown index and YAML machine-index export: known — files [wired] (Export storage is established independently of later consumption.); Freshly derived lexical, alias and adjacency access structures: known — in-memory [wired] (Includes the access structure, not ordinary executor state; directed links are fields in dictionaries, not a separate graph storage service.)	partial coverage (Automatic code-controlled admissions are wired, while content-supplying callers and initial file editors retain unknown human control. No manual finding is established merely from caller initiation.); Caller-selected content writes, explicit mutations, trace recording and consolidated-prose submission: not-determinable (A generic caller may be human, model or application. No recorded per-content human decision establishes manual agency, and lack of an approval gate alone does not settle who supplied or approved the content.); Raw filesystem admission of initial content: not-determinable (Reload is inspected but initial admission control is unresolved; API-mediated control cannot be transferred to file editors.); Remember auto successor selection and generated description/default metadata: known — automatic [wired] (Local automatic branch/field admission does not resolve approval of the caller's prose or explicit-mode alternatives.); Selected-ID access reinforcement: known — automatic [wired] (A caller can disable marking, but does not individually determine each numeric update.); Invoked repair, link rewriting, index generation and consolidation bookkeeping: known — automatic [wired] (Dry-run/skip options withhold changes; triggering a batch does not make its individual admissions manual.)	known coverage (The records cover the shipped note, trace and index delivery inventory, all explicitly requested. External host scheduling is excluded, not an unresolved shipped push route; unknown eventual prompt use does not change request direction.); Get/list/search/recall of retained notes and fresh access structures: known — pull [wired] (Automatic ranking, neighbor expansion and access side effects inside requested recall do not create an unsolicited push.); Raw-event reload/group suggestions: known — pull [wired] (No automatic continuation invokes consolidation or delivers a later recall after the review response.); Saved Markdown index reload: known — pull [wired] (Stale saved displays remain request-delivered, not a push channel.); Saved YAML export: inapplicable (No shipped YAML reload consumer is established; persistence alone is not read-back. Fresh machine structures used in requested retrieval are included above.)	inapplicable coverage (The completely assessed shipped delivery inventory is pull-only. No identifier, lexical or coarse push is inferred from catalog entries, query responses or budget filtering.); Requested note retrieval, trace review and Markdown index load; YAML export without read-back: inapplicable (All established read-back is pull, and the YAML export lacks read-back. This axis classifies push selectors only; lexical query scoring and requested IDs therefore supply no signal value.)	not-determinable coverage (Inspection establishes retained experience, caller-led review and automatic numerical adaptation, but no fully classified automatic trace-fed write. Missing caller control and source-to-output linkage prevent yes; no scoped absence warrant establishes no. This is not a claim that learning or adaptation is absent.); Raw trace review followed by caller-supplied durable prose and later retrieval: not-determinable (The experience-to-prose affordance is recorded, but generic human/model/application identity leaves automatic semantic admission unresolved and trace IDs need not support the submitted content. Neither an unconditional yes nor a bounded no follows for this included alternative.); Access timestamp/count/salience update and later ranking: not-determinable (Automatic durable use reinforcement is wired, but the records identify selected IDs and prior metadata as inputs, not a recorded trace-fed derivation. The update writes no raw trace. This supports behavior shaping without establishing a controlled trace-learning value.); Ordinary content admission, event persistence and deterministic maintenance/indexing: not-determinable (Raw event retention, note acquisition and note-derived repair/export do not by themselves establish automatic trace-fed durable behavior shaping. The accepted set supplies no bounded absence record licensing a system-wide no, and externally supplied content origin remains unresolved.)	partial coverage (Fixed retrieval effects and afforded reference/update roles are supported. External content interpretation and review control prevent a complete authority classification; static tool guidance is outside the memory boundary.); Retrieved prose, display descriptions and Markdown index at external readers: partial — knowledge [afforded] (Host interpretation is not supplied. Standing instruction force or other content-specific effects cannot be classified from delivery alone.); Note text, lexical metadata, lifecycle, expiry, counters and derived term/graph features at retrieval selectors: known — ranking [wired], routing [wired], enforcement [wired] (Fixed selector consumers are inspected. Confidence and verification labels are not truth validation, and unconsumed related/source metadata has no inferred force.); Retained trace fields/content at grouping and review: partial — routing [wired], ranking [wired], knowledge [afforded], learning [afforded] (The raw-to-review-to-durable-note route is supported as an affordance; actual semantic update control, resulting content and later host effects remain unresolved. Learning authority here does not establish improved capacity or automatic trace learning.); Persisted YAML machine-index export: inapplicable (The inspected export has no shipped reload consumer, so no authority is assigned to the saved YAML; the freshly derived selector representation is classified separately.)	AAS-2026-10-14-smriti-mcp-14f6928f9817-01
SoL-Pi	2	Included: accumulated ObservationPack text archives, replay metadata and ledger; reducer raw-log archives, retained receipts and journals; OCC retained plan/progress/control estimates and native replacement summary/recent tail; Skill host transcripts, persisted exported prompts, child event/session records and result handbacks; Visual retained crop PNGs and receipts, including same-input reuse. Includes unresolved original host-summary/image admission and opaque native persistence/selection where material. Excluded: static shipped prompts/policies, ordinary untransformed tool returns, general configuration, arbitrary workspace edits without a demonstrated memory consumer, and external model weights not established as changing through use.	partial coverage (Local archives, exported prompts, run records and crop caches are files; replay counters are in memory. Pi custom entries and native replacement context have an external persistence contract not fully specified by these records.); Observation text archives, reducer raw-log files, Skill prompt/events/session/result files, and Visual PNG/receipt cache: known — files [wired] (The records identify concrete local file paths, including session JSONL in the Skill launcher. Temporary/no-session archive roots still contain files.); Observation replay-count access map and OCC pending selection state: known — in-memory [wired] (Root/id replay counts and pending selection/continuation objects reside in process memory; retained session estimates are separately classified.); Core Pi custom entries, persisted receipt history, and native summary/recent-tail backing store: not-determinable (The records establish session retention and later consumption, but not a complete physical substrate inventory outside the shipped Skill file route. No database or service substrate is inferred from the host API.)	partial coverage (Inspected ongoing admissions are software/model controlled. Workflow initiation and manual compaction initiation are not per-content approval; original host checkpoint/transcript/image admission remains unresolved.); Observation/archive admission, reducer receipt/journal admission, OCC plan/estimate updates, Skill export/result creation, Visual crop/cache admission: known — automatic [wired] (Code and models select, validate and admit content without an explicit human decision approving each retained item.); Requested native compaction commit through Pi callback and continuation contract: known — automatic [wired] (The extension automatically requests summarization and consumes commit/refusal; a manual/native trigger is an alternate initiation path, not evidence of per-summary human approval.); Initial host transcript/summary and source-image content admission before the inspected transformations: not-determinable (Generic caller identity and selected file paths do not identify a human content decision or software-only admission for every original input. No manual value is asserted from an operator starting the workflow.)	partial coverage (Both automatic supply and requested readback exist. External native cut/assembly and unestablished progress/ledger consumers remain included limitations rather than being silently treated as pull-only.); Provider-context observation projection and native summary/tail delivery to continuation: known — push [wired] (Code supplies selected retained context without the later foreground invocation requesting each item. The hidden continuation message is a trigger, not a task-specific checkpoint.); Observation id/offset recall, Skill CLI status/result/wait/list, and Visual same-input cache read: known — pull [wired] (The consumer explicitly requests retained pages, a run result or a crop; fulfillment is pull even when code reads bytes and delivers them.); Reducer archive ranged shell read: known — pull [afforded] (Receipt guidance offers an explicit foreground consumer and range-read route; no observed shell reread or separate wired recall executor is established.); OCC branch restore and automatic diagnostic-receipt replay: known — push [wired] (Restore supplies the latest valid state to extension code; retained substituted receipts enter later foreground context without a specific archive request.); Skill transcript request and later automatic child-prefix/parent-notification supply: known — pull [wired], push [wired] (A caller requests transcript export or CLI result; the selected prefix is supplied to the child and a correlated completion report is supplied to the parent without those model invocations requesting individual historical parts.); Independent pendingProgress/pack-ledger consumption and opaque native context-selection alternatives: not-determinable (No dedicated prompt route for pendingProgress or later pack-ledger policy consumer is established; native internal assembly is external. This does not erase the supported summary delivery.)	partial coverage (Push selectors are classified separately from identifiers used only on requested reads. Known coarse and judgment-based selection do not resolve the host summarizer or an unestablished checkpoint channel.); Observation projection at provider-context construction: known — coarse [wired] (Trigger: context event. Inputs: success/pure-text/size/receipt exclusion and replay count. Selected part: full retained message or literal head/tail placeholder. Root/id keys counts but no targeted relevance match selects evidence.); Reducer evidence selection for substituted receipt later replay: known — inferred-judgment [wired] (Trigger: eligible diagnostic result. Selector: reducer judgment over the selected log and metadata. Selected part: bounded quotations and kind/uncertainty fields delivered in the admitted receipt.); Reducer candidate gate before receipt selection: known — coarse [wired] (Trigger: tool_result. Command regex, tool class, error metadata, byte limits and secret precaution filter select the candidate retained body; this is coarse eligibility, not lexical relevance search over an archive.); OCC active-branch restore: known — identifier [wired], coarse [wired] (Trigger: session start/tree restore. Actual customType identity match selects online-state entries; parser validity and newest-active-branch order choose the retained state supplied to code.); Skill child-prefix supply: known — coarse [wired] (Trigger: requested export during handoff. Host compaction point, roles, current-handoff cutoff, recency and character budget select retained transcript blocks. No semantic ranker is established.); Skill asynchronous parent notification: known — identifier [wired], coarse [wired] (Trigger: correlated submit/notify hook. The recorded run id selects the matching retained result, terminal status permits reporting, and length limits trim final text. Identity targeting and coarse status/budget filtering remain distinct findings.); Native summary/tail delivery: partial — coarse [wired] (Chronological native cut and recent-tail selection are described, but host internal selector inputs and summary content decisions are uninspected.); Requested archive pages, CLI result and crop-cache returns: inapplicable (These are pull operations. Their requested ids/paths/hash keys do not establish an additional identifier-based push signal.); Independent pendingProgress or pack-ledger push: not-determinable (No implemented later supply channel is established for these retained parts, so no selector value can be assigned.)	partial coverage (Qualifying automatic writes produce retained context artifacts or control estimates for later consumers. This is the profile definition of trace learning, not demonstrated capacity improvement. Raw storage and image origins remain separately unresolved.); Diagnostic tool log to admitted retained receipt to later foreground context: known — yes [wired] (Automatic model/code selection turns a tool trace into retained behavior-shaping evidence; the later consumer chain is wired, not observed.); Request/context/boundary traces to retained OCC estimates to later compaction decisions: known — yes [wired] (The memory annotation establishes automatic trace-fed control-state adaptation and an actual later code decision consumer.); Accumulated history to native summary to later foreground invocation: known — yes [wired] (Automatic summary production, host retention/commit and continuation form a qualifying wired chain even though summary-generation internals and fidelity are uninspected.); Host transcript to persisted prefix to child; child events/session entries to persisted result to parent: known — yes [wired] (Automatic filtering and aggregation produce retained reference/handback artifacts consumed on later invocations. Starting delegation is not per-content human approval.); Raw observation archive/replay ledger, pendingProgress as independent checkpoint, and crop cache: not-determinable (Raw trace retention alone does not qualify; the pack placeholder is a context projection rather than an established durable generated artifact, and progress lacks a dedicated prompt readback. Crop input activity origin is not established. These records do not warrant a bounded system-wide no or an extra yes for every part.)	partial coverage (Knowledge, learning, routing, validation and enforcement are supported on named consumption paths. Opaque native selection and unconsumed progress/ledger parts prevent a complete authority inventory; static shipped policies are excluded.); Observation excerpts/pages, reducer quotations/labels, Skill prefix/final report and crop pixels at later foreground consumers: known — knowledge [wired] (Delivered retained content supplies advisory evidence/reference context. It neither certifies task correctness nor becomes an enforced system policy.); Native summary at later foreground invocation: known — knowledge [wired] (Replacement summary supplies task context after compaction; fidelity and compliance remain unobserved.); Retained raw log at reducer; history at native summarizer; host transcript at exporter; child events/session entries at result summarizer: known — learning [wired] (Retained experience feeds automatic construction of durable behavior-shaping receipts, summaries, prefixes and handbacks. This learning authority belongs to each update consumer, not automatically to all raw-log storage.); Trace-derived OCC estimates and plan state at later turn-end selection; replay counts at context projection: known — routing [wired] (Retained state selects whether/when to compact or which replay representation to supply. It changes operation/input-window choice, not a verified task outcome.); Active-branch session-entry metadata at restore and run-status metadata at notification/status selection: known — routing [wired] (Custom-entry type/validity/recency selects restored state; run identity/status selects the stored result and notification timing.); Existing observation/archive content and existing Visual receipt/output hashes at integrity admission: known — validation [wired], enforcement [wired] (Existing retained bytes supply conformance references for reuse; mismatches reject reuse or throw rather than silently overwriting. The same checks provide a veto as well as a criterion.); Reducer decision journal at Skill accounting consumer: known — knowledge [wired] (Session entries supply evidence for reported usage, counts and costs; no learned reducer policy or task-correctness authority follows.); PendingProgress as independently delivered checkpoint; ObservationPack ledger; opaque native summary-selection metadata: not-determinable (Parsing progress is not a dedicated task-content prompt consumer. No later policy consumer of the pack ledger is established. Native internal selection is not inspected; additional authority cannot be inferred.)	AAS-2026-10-14-sol-pi-42f04fbbc6bb-01
Spacebot	2	Use-accumulated or revised agent claims, human anchors, association/selectors and dense indexes; working events and temporal summaries; live/durable conversation and worker continuation records; chronicle checkpoints; mutable skills/support files and custom identity snapshots; wiki pages/versions; retained task/goal/project/wake state and attachment alternatives where inventoried. Includes known routes and locally unresolved initial contents, org refresh, attachment payloads and restored-worker/auxiliary-state paths. Excludes shipped static presets/prompts/built-in procedures as such, transient current-run accounting, debug captures without established later consumers, provider/model-weight internals, remote MCP and external OpenCode internal retention. Spacebot-owned OpenCode snapshots remain included.	partial coverage (Known SQL, graph, vector, file and live-memory objects coexist. Attachment payloads and full auxiliary-state stores prevent a complete substrate inventory.); Agent claim text, selectors and human-anchor mapping: known — sqlite [wired] (Text and selector parts have the same evidenced persistence substrate.); Typed weighted memory associations: known — graph [wired], sqlite [wired] (Graph organization and SQL persistence coexist.); Working events, temporal narratives, durable conversation and chronicle checkpoints: known — sqlite [wired] (Storage is known for these objects, not all runtime state.); Wiki current pages and version history: known — sqlite [wired] (Separate instance-level retained object.); Memory and chronicle retrieval tables: known — vector [wired] (Underlying Lance implementation and startup repair coverage are not independently audited; vector storage is explicit.); Mutable skill procedures, support files and custom identity bodies: known — files [wired] (Embedded static seeds are not use-derived memory.); Live history summaries, loaded skill/identity snapshots and participant cache: known — in-memory [wired] (Does not assert that rolling summaries survive restart.); Worker/OpenCode retained snapshots, full task/goal/project/wake stores and attachment payloads: uninspected (Records establish retention or included routes, but do not classify every operative payload/store. Exact substrate inventory remains unresolved. External OpenCode internals are excluded.)	partial coverage (Automatic admission is wired on named routes. No manual value is inferred from administrative naming; unresolved initial and caller control prevents a complete agency inventory.); Model memory saves/import extraction, merges, wiki edits and reflected skill mutations: known — automatic [wired] (Conversation initiation, file upload and User origin do not convert subsequent model-selected admission to manual.); Runtime log/event/selector/index writes, temporal/chronicle production and periodic maintenance: known — automatic [wired] (Physical I/O is not the warrant; the decision mechanism is software/model control.); Administrative/API/CLI edits, external skill installation requests and generic file/shell changes: not-determinable (Generic caller identity and available operator surfaces do not establish a human exact-content decision for each mechanism. Automatic reflection remains separately known.); Initial custom contents, org seed replacement and untraced attachment/auxiliary admission: not-determinable (Initial control and complete alternative write paths are unresolved; no manual classification or bounded automatic-only claim is licensed.)	partial coverage (Both directions are wired; missing alternative-route coverage prevents known rather than cancelling either positive.); Explicit memory/channel/chronicle/wiki reads and optional skill body requests: known — pull [wired] (Hybrid vector/lexical selection in requested recall is not a separate unsolicited push.); Automatic store/temporal/participant context and chronicle view/hydration: known — push [wired] (Budgets may omit content or collapse it to a first line/title.); Loaded skill index and task-required skill bodies: known — push [wired] (Required selection may originate in an upstream task decision; automatic body supply is distinct from optional pull.); Live rolling summary and worker durable completion/resumption at later inference: known — push [wired] (Continuation context is memory across invocations without requiring cross-restart persistence.); Mutable identity snapshots in later system prompts: known — push [wired] (Only custom/revised retained text is in memory scope; static preset loading alone is excluded.); Attachment, auxiliary-state and all restored-worker alternatives: uninspected (Known push and pull routes cannot determine all included delivery alternatives.)	partial coverage (Coarse and identifier push have local witnesses. Embedding/lexical selectors in requested recall are pull, not automatic push signals; unresolved alternatives prevent an exhaustive signal assessment.); Automatic global store and temporal/activity blocks: known — coarse [wired] (No query embedding or lexical relevance selector is established on this push branch.); Participant-scoped anchor supply: known — identifier [wired], coarse [wired] (Actual key matching, not the presence of an ID label, establishes targeted selection; cache refresh is unresolved.); Chronicle view, durable channel hydration and rolling history continuation: known — coarse [wired], identifier [wired] (Requested sequence/cursor expansion is pull and does not supply an additional push signal.); Skill index and required procedure body supply: known — coarse [wired], identifier [wired] (Suggested labels are not themselves full-body delivery; optional named reads remain pull.); Worker completion/result retrigger: known — identifier [wired] (The consumer receives result text, not every private tool message.); Custom identity prompt rendering: known — coarse [wired] (No relevance query is evidenced on this path.); Uninspected attachment, auxiliary-state and restored-worker push alternatives: uninspected (Unresolved delivery directions also leave their push selector signals unresolved; pull witnesses cannot make this unit inapplicable.)	partial coverage (Several qualifying automatic trace-fed routes are wired. Unresolved writes remain included; this yes union is not an efficacy, exercised-learning or self-improvement finding.); Conversation extraction to standing memories: known — yes [wired] (No new claim, weight update or improved task performance is required for this route classification.); Working-event intraday and daily summaries: known — yes [wired] (Raw event retention alone is not the qualifying write; the derived narratives and consumers are.); Chronicle summaries/rollups and rolling continuation summaries: known — yes [wired] (Durable chronicles and session-live rolling summaries differ in persistence horizon; fallback notices do not demonstrate semantic summary production.); Procedure reflection: known — yes [wired] (Prompted abstention and local structural rails are not an observed beneficial-learning result.); Generic wiki/identity/API/file writes, attachment and auxiliary-state alternatives: not-determinable (Inputs and admission are not completely traced. No bounded absence record supports a local no or a system-wide negative.)	partial coverage (All seven forces have local wired witnesses; remaining consumers prevent a complete authority assessment. Shipped static criteria and transient completion bookkeeping are not counted as accumulated memory.); Claims, temporal narratives, checkpoints, wiki and restored result text at later work models: known — knowledge [wired] (A system-prompt channel does not by itself turn claim text into standing instruction.); Retained procedures and custom identity at later model consumers: known — instruction [wired] (Required delivery is distinct from compliance.); Retained session/worker/event inputs at extraction, summarization and reflection writers: known — learning [wired] (Update-consumer authority is separate from downstream knowledge/instruction; no improved capacity is asserted.); Claim text, embeddings, associations and importance at retrieval/render rankers: known — ranking [wired] (Ranking inputs need not be a learned ranking policy.); Forgotten flags and archive metadata at ordinary discovery: known — enforcement [wired] (Wiki explicit slug reads and previously delivered copies remain outside that exclusion.); Skill pin/origin/read state and durable worker/task metadata at admission checks: known — validation [wired], enforcement [wired] (These local vetoes do not establish global filesystem protection or work-product correctness.); Chronicle boundaries, task plans and durable wake/run state at continuation/control consumers: known — routing [wired] (Routing is supported at these consumers without resolving every auxiliary state machine.); Uninspected attachment, profile-card and auxiliary-state consumers: uninspected (Generated profile-card production has no established behavioral consumer; full attachment and auxiliary paths cannot be assigned authority from display summaries.)	AAS-2026-10-14-spacebot-6a954b400f95-01
Sparks	2	Accumulated inbox/raw/archive files; wiki bodies and Revision text; operation log; live Tasks; index descriptions/catalog; all seven regenerated collections; SQLite file/frontmatter/link projections and ingest history; optional Git history; their write, maintenance and later-consumer routes. Excludes shipped static contract/instruction copies, product builds, external harness conversation/checkpoints and provider/model state. Unknown initial content and generic caller/direct-edit admission remain included limitations, not excluded alternatives.	known coverage (The accepted object inventory covers retained files, SQLite access/session state and optional Git history. Static shipped instructions, build products and excluded provider/harness state are not accumulated vault memory.); Raw capture/archive, wiki bodies and revisions, log, Tasks, index descriptions/catalog and seven regenerated collections: known — files [wired] (The retained file inventory distinguishes live Tasks from the seven generated collections and prose from manifest projections.); Manifest file/frontmatter/link projections and ingest lifecycle rows: known — sqlite [wired] (Includes both symbolic access projections and the separately declared session-state rows.); Optional Git history of accumulated vault files: known — repo [wired] (Git is optional and not a semantic acceptance store; filesystem and SQLite storage remain operative without a commit.)	partial coverage (Explicit human capture and model/software admission are supported separately; initial/direct/generic-caller control remains unresolved, so the axis is partial.); Human capture and direct human source provision: known — manual [claimed] (Manual admission is doctrinal here; no actual operator-content instance was supplied.); Protocol-directed external-agent wiki/index/log/task writing: known — automatic [claimed] (Classification concerns the prescribed admission mechanism, not observed model obedience or the unknown author of an initial file.); Archive/clear, scan/session transitions, collection regeneration and index preservation/rebuild: known — automatic [wired] (Known deterministic admission covers subsequent writes; it does not identify upstream content authors.); TaskAdd/TaskDone deterministic append and selected-marker admission: known — automatic [wired] (Automatic target/write admission is separate from who supplied the task text or matching request.); Initial vault content, arbitrary direct edits and generic caller-supplied task text: not-determinable (No authenticated initial-content decision or generic caller principal establishes human versus model admission for these alternatives. This prevents complete write-control coverage, not classification of inspected subsequent transformations.)	known coverage (The accepted dispatch/consumer inventory describes caller-requested chains throughout the accumulated-vault boundary. Static harness contract loading and external context scheduling are excluded; no included unsolicited payload branch is left unresolved.); Query/Brief/Viewer request, selection and delivery; prescribed external page reads: known — pull [wired], pull [afforded] (Automatic selection within fulfillment is not unsolicited push; actual agent body-read/context inclusion is not observed.); Requested collection/index/task/scan maintenance and affected advice consume earlier retained state: known — pull [wired] (Includes machine later-consumers, not only model readers. Affected advice is not retained, but its scan/session inputs are read back.)	inapplicable coverage (No push operation is established within the completely inventoried requested vault-consumer boundary; signal classification applies to push only. This does not classify excluded harness scheduling.); All included caller-requested vault content and metadata read-back chains: inapplicable (All covered read-back is pull. Query identity matches, brief date/budget rules and requested URLs are selectors within fulfillment, not push signals.)	not-determinable coverage (The accepted records support retention/curation and historical read-back, but neither a qualifying trace-fed write nor an evidenced-absence record supplying a controlled negative. This is not a system-wide no and does not assess excluded provider/harness learning.); Archive, scan/lifecycle state, collections, index and task writes with later consumers: not-determinable (Inspected writes retain captures, access state, projections or task status; accepted records establish no qualifying original trace-to-durable-behavior-shaping update. No bounded absence record warrants a controlled no finding for the whole family.); Agent-authored operation log and semantic wiki revision; Brief log consumption: not-determinable (Log summaries are requested authored history, not recorded automatic trace extraction. Brief delivers history but saves no derived checkpoint/policy. Unavailable actual agent traces/content prevent establishing a qualifying automatic write; no supported yes or bounded no is asserted.)	known coverage (The inventory resolves actual reader, selector, scoring and lifecycle consumers. Static contract instruction authority is excluded; unconsumed metadata gets no authority. Missing observed activation does not erase wired/afforded force.); Retained raw/wiki/Revision/log/task/collection prose and index descriptions at readers or answering/briefing agents: known — knowledge [afforded], knowledge [wired] (These routes supply context/advice, not standing harness instructions; actual compliance or answer dependence is unobserved.); Manifest fields/edges and ingest history at Query, index, Projects, affected and viewer navigation selectors: known — routing [wired] (Metadata has fixed selection force even when no source body is loaded. Advisory affected names do not automatically execute regeneration.); Retained open/done task markers at TaskDone and Brief: known — routing [wired], enforcement [wired] (The exclusion concerns delivery/selection, not enforcement of task execution or real-world success.); Retained task text at fuzzy target matching: known — ranking [wired] (Fixed scoring establishes ranking input authority without retained learned ranking policy.); Retained ingest active state at prepare/finalize admission: known — validation [wired], enforcement [wired] (This is a best-effort lifecycle interlock, not a semantic-content veto or race-free filesystem lock.)	AAS-2026-10-15-sparks-563865f8b26f-01
Stash	2	Use-accumulated PostgreSQL episodes, facts, dense embeddings, entity relationships, causal links, abstraction patterns, contradictions, hypotheses, goals, failures, namespace focus and descriptions, maintenance cursors, embedding cache and the recorded dimension/model settings; caller admissions, all maintenance transformations and later client/model/operational reads. Includes unresolved provenance, human control and storage-only consumer alternatives. Excludes compiled server/model prompts and release products, current request-local state, external client histories/decisions, Stash Cloud and provider model weights not updated through these routes. Further settings consumers remain unresolved rather than silently excluded.	known coverage (All inventoried use-accumulated objects reside in PostgreSQL; vector columns and cache are included explicitly. Provider weights and shipped static products are outside this memory inventory.); Episode/fact text and metadata; relationships, causal links, patterns, contradictions, hypotheses, goals, failures, focus, namespaces, cursors, settings and embedding cache: known — rdbms [wired] (The accepted object inventory identifies PostgreSQL for each retained part; graph-shaped edges are relational rows, not evidence of a graph storage engine.); Episode/fact embeddings and cached embedding vectors: known — vector [wired] (Dense numerical parts have an explicit vector store within the relational database.)	partial coverage (Automatic admissions are established locally. Generic caller identity cannot resolve manual versus automatic initial content/edit/withdrawal decisions.); Maintenance model proposals admitted by code, confidence/validity updates, cursor saves, embedding memoization and namespace scaffold ancestors: known — automatic [wired] (Starting consolidation is not per-content human approval.); Caller remember/focus, namespace descriptive writes, goals, failures, hypotheses, causal assertions, explicit edits/status changes and withdrawal: not-determinable (The API accepts the caller decision but does not identify human approval. Physical SQL writes and possible agent authorship do not resolve admission control.)	partial coverage (Requested external reload coexists with automatic maintenance supply. Pattern and bare getter alternatives are capabilities with unresolved actual later consumers; they remain in scope.); Recall, typed tools and exact-namespace focus reload: known — pull [wired] (A consumer requests the returned material; serialization is not another push.); Embedding cache lookup and operational namespace/checkpoint/settings reads: known — pull [wired] (These are requested operational reloads, not client knowledge delivery.); Automatic retained batch selection and message assembly for maintenance reasoner: known — push [wired] (Manual or timer initiation does not make each model input batch a consumer request for that retained material.); Pattern QueryPatterns route: partial — pull [afforded] (Requested API read is afforded, but no shipped later consumer is found, preventing a complete wired read-back account.); Bare GetFact/GetEpisode alternatives without a found shipped consumer: not-determinable (Storage getters alone do not establish a concrete later-consumer route.)	partial coverage (Push findings apply to maintenance selectors, not query embeddings on requested recall. Unknown actual use of storage-only alternatives prevents completely assessing every direction.); Maintenance trigger -> namespace/watermark/status/availability/row-limit selection -> model evidence batches: known — coarse [wired] (Every maintenance invocation uses bounded availability/window filters; row limits are not token budgets.); Episode-to-fact trigger -> stored episode vector similarity -> clustered episode texts delivered to reasoner: known — inferred-embedding [wired] (The cluster selector operates before model input; this is distinct from requested semantic recall.); Inline contradiction trigger -> retained entity/property identity match with differing values -> old/new values delivered to comparator model: known — identifier [wired] (The actual equality match chooses retained old fact values, not merely a catalog or requested ID.); External requested recall/typed/context routes and operational cache/control lookups: inapplicable (These local routes fulfill a consumer request; their filters do not classify push signals.); Unestablished consumers of patterns and bare getters: not-determinable (A hypothetical actual consumer could select/deliver differently. The records do not establish that chain.)	partial coverage (Experience-fed durable derivation is supported without weight training or demonstrated improvement. Generic episode provenance and caller-controlled admissions prevent exhaustive route coverage.); Failure/activity reports plus recent episodes -> automatic repeat evidence and failure-pattern facts -> later consolidation/recall: known — yes [wired] (Reports of experience are trace inputs; no authentication or measured benefit is implied.); Session activity summarized into episodes -> automatic fact consolidation -> later recall/maintenance: partial — yes [wired] (The write/retention/later-consumer chain is wired for experience-bearing episodes. Arbitrary caller assertions are also admitted, so not every consolidation input is known to be a trace.); Caller initial writes/focus replacement and downstream knowledge-only updates with unresolved original episode/fact provenance: not-determinable (Caller control and original input content are unresolved. Neither raw storage nor an unknown upstream derivation warrants a blanket no or yes.); Cache, namespaces, settings and cursor bookkeeping as operational metadata rather than trace-derived behavior-shaping artifacts: inapplicable (These recorded mechanisms memoize values or route work; they are not the qualifying trace-to-durable-content update.)	partial coverage (Positive forces are consumer-specific. Pattern downstream use and further settings consumers remain unresolved; delivery does not establish client obedience or benefit.); Recalled episodes/facts, typed retained knowledge, focus and descriptive namespace text at requesting external client: known — knowledge [wired] (Knowledge/advice is delivered, without enforcing the external agent's answer.); Retained preferences or standing self-advice at external agent following the session protocol: known — instruction [afforded] (The documented standing-guidance role is afforded; stored instances, activation and host instruction precedence are uninspected.); Episode/fact/relationship, hypothesis/goal and failure content at maintenance model evidence consumer: known — knowledge [wired] (Retained content is evidence for bounded model judgements, not a truth certificate.); Reported episode/failure experience at automatic durable-derivation consumers: known — learning [wired] (Learning authority here means experience feeding durable artifact updates, not demonstrated improved capacity.); Stored embeddings at recall ranking and clustering/deduplication; goal priority at list ordering: known — ranking [wired] (The retained numerical state and priority influence relative ordering/grouping under fixed rules.); Namespace IDs/slugs and checkpoint watermarks at storage/maintenance selectors; retained hypothesis/goal state at next scan: known — routing [wired] (These fields choose input scope/window and eligible operation, not caller authorization.); Deleted/resolved/terminal state at read/scan selectors and dimension-lock setting at bootstrap: known — enforcement [wired] (Only consumers consulting a field enforce it; valid_until is not enforced by ordinary recall.); Retained hypothesis/goal status at transition admission and dimension lock at vector setup: known — validation [wired] (Retained criteria constrain admissible transitions/dimensions, separately from generated model judgement.); Pattern rows beyond QueryPatterns capability and unspecified additional settings consumers: not-determinable (No concrete shipped pattern consumer was found; settings were not exhaustively audited. These gaps prevent a complete authority inventory.)	AAS-2026-10-15-stash-7ac5612837a0-01
supermemory	2	Included: accumulated remote source documents/chunks and embeddings, extracted graph facts and eligibility/relations, profiles/aggregates, Claude exact-file documents, continuation display reuse, retained namespace/upload access state, their writes, maintenance and later consumers, including opaque engine and incompletely traced adapter/connector alternatives. Excluded: unchanged shipped prompts/package code, transient current-call payloads without later reuse, generation/extraction model weights not shown changed through memory use, external host machinery and uncaptured benchmark/plugin internals.	partial coverage (Remote objects, graph/vector declarations, local reuse and durable access state are classified. Opaque engine physical storage prevents complete substrate coverage; local file paths do not establish files or repo storage.); Remote documents, Claude file bodies and service-owned profile representation: known — service-object [wired], service-object [claimed] (The service-object classification identifies the operative API-owned object, not its physical database.); Declared temporal graph and indexed dense representations: partial — graph [claimed], vector [claimed] (Logical graph/vector organization is declared; physical stores and the remaining engine state are inaccessible.); Wrapper-local retained display strings: known — in-memory [wired] (The LRU retains formatted memory across continuation invocations, not canonical durable facts.); Durable active namespace and upload capability state: known — service-object [wired] (This includes retained access structures, not all runtime credentials or permissions.)	partial coverage (Automatic wired and claimed paths coexist with an afforded manual path and unresolved generic callers. Neither human authorship nor physical API I/O decides agency.); Automatic conversation submissions, voice capture and inspected hook submissions: known — automatic [wired] (Flags permit disabling these paths; starting a host request is not per-content approval.); Model-selected framework saves and Claude file commands: known — automatic [wired] (The shipped model-driven routes admit proposed content through code/service checks.); Remote fact/profile extraction and connector sync: known — automatic [claimed] (Authorization of acquisition scope is not approval of each resulting document or fact.); Interactive guided-save human editing/approval surface: known — manual [afforded] (The memory annotation records an afforded human content decision, not an exercised or mandatory gate.); Generic MCP/HTTP saves, uploads, review and document maintenance callers; initial backend population: not-determinable (Generic caller identity does not determine human versus software admission; initialization control is unknown.); Other adapter admission predicates: uninspected (Not all alternative write mechanisms are traced.)	partial coverage (Both directions occur at different operations in a chain. Explicit MCP prompt requests are pull; automatic pre-model insertion is push. The complete read inventory remains unresolved.); Explicit framework/MCP search, profile, document, context-prompt and Claude file requests: known — pull [wired] (Selection inside fulfillment does not turn a requested result into unsolicited supply.); Middleware and lifecycle context injection and continuation reuse: known — push [wired], pull [wired] (A host model invocation triggers automatic memory supply without a model request for that material.); Connector downstream consumer contract: known — pull [afforded] (Sync writes automatically but does not independently deliver context to a model.); Other adapter and opaque engine delivery alternatives: uninspected (Untraced alternatives prevent complete direction coverage; no independent connector push is inferred.)	partial coverage (Signals classify automatic push selection, not requested file IDs or every API return. Coarse profile supply and actual identity matches are supported; inferred query-signal taxonomy remains unresolved.); Automatic profile mode before model invocation: known — coarse [wired], identifier [wired] (Trigger is host generation/lifecycle; selector uses configured mode and namespace to supply static/dynamic profile content.); Vercel continuation cache hit before generation: known — identifier [wired] (Trigger is a non-user continuation; namespace, conversation ID, mode and normalized latest-user text match a retained string.); Query-driven automatic recall: partial — identifier [wired] (Latest user text and namespace select retained query results, but remote semantic scoring cannot be decomposed into embedding, lexical or judgment mechanisms from these records.); Other push selectors and opaque ranking internals: not-determinable (No accepted record resolves all alternative selector mechanisms; do not infer embedding selection merely from embedding storage.)	partial coverage (A declared qualifying session route supports yes without upgrading the opaque transformation to wired. Unresolved alternatives keep coverage partial; no negative is asserted.); Conversation-to-fact/profile automatic extraction and later context consumption: known — yes [claimed] (Public capture/read wiring is inspected; the intervening durable derivative production is declared, so the complete qualifying chain remains claimed.); Model-authored exact files and generic remember calls: not-determinable (Automatic write and later reload are wired, but a qualifying original trace-fed content derivation is not established for arbitrary saved text.); Connector and mixed document inputs to dreaming: not-determinable (Documents may contain activity records, but the accepted records do not resolve qualifying trace content for these alternatives.); Other adapter routes, initialization, opaque engine updates and evaluation feedback: uninspected (Alternative trace-fed updates are not exhaustively assessed; no bounded absence record licenses a local or system-wide no.)	partial coverage (Authority follows each actual consumer. Recorded knowledge, instruction, learning, ranking, routing and capability/eligibility controls are not a claim of observed dependence or improvement.); Fact/profile/document text and exact file prose at later answering models: known — knowledge [wired] (Delivery supplies context, not reliable compliance or factual warrant.); Retained file guidance and preferences interpreted as directions: partial — instruction [wired] (Instructional force is supported for the content/guidance path, not every arbitrary file or factual assertion.); Retained conversation experience at the extraction/profile update consumer: known — learning [claimed] (The declared trace-to-durable-context update supplies learning authority independently of downstream knowledge authority.); Retained fact text and dense features at remote retrieval ranking: partial — ranking [claimed] (Search ranking consumes retained material, but exact scores/features and all backend rankers are opaque.); Active namespace at the default-space resolver: known — routing [wired] (Explicit namespace can override this routing default.); Retained eligibility flags at search/review selection: known — enforcement [claimed] (This is a documented search veto, not demonstrated service enforcement.); Retained upload capability at upload admission: known — validation [wired], enforcement [wired] (The stored token hash and expiry are admission criteria and a code-level veto, not content-truth criteria.); Opaque engine and incompletely traced alternative consumer effects: not-determinable (Exact internal consumers and alternative effects cannot be exhaustively classified.)	AAS-2026-10-19-supermemory-1cc18c778c48-01
Swamp	2	Retained material accumulated or revised through Swamp use: versioned resource/file content and access metadata; issue plans, feedback, findings, phase and verification state; method/verification reports; command audit; workflow checkpoints and evaluated replay forms; task briefs; editable/reloaded extension capabilities; retained evaluation outcomes and attestations. Includes unresolved acquisition, custom-report, remote-backend, cache, maintenance and later-consumer alternatives identified in the accepted records. Excludes static shipped skills/prompts as memory objects, ephemeral scratch state without a later invocation, external harness conversation/compaction and model weights, and external service internals. Shipped skills remain evidence of instructed consumption of retained objects.	partial coverage (Local retained objects have file-backed witnesses. Backend adapters, hosted persistence and some alternative paths remain unresolved; JSON/YAML encoding does not establish SQLite or relational storage.); Persistent resource/file content and its access metadata: known — files [wired] (Content files, YAML metadata and latest/version directories establish files; ephemeral-only working data is excluded from retained memory.); Issue plan-main: known — files [wired] (Plan JSON is persisted through the versioned file-backed data repository.); Issue feedback-main: known — files [wired] (Feedback JSON uses the same inspected persistent resource path.); Adversarial and conformance findings: known — files [wired] (Both resource families use versioned persistent JSON files.); Issue phase and verificationResult resources: known — files [wired] (Persistent resource storage is established for phase and verification flags.); Method report Markdown and JSON artifacts: known — files [wired] (Separate persisted report files retain both representations.); Verification summary Markdown and JSON artifacts: known — files [wired] (The report persistence wrapper saves file artifacts.); Command audit records: known — files [wired] (Date-partitioned JSONL files retain selected command events.); Workflow run checkpoint: known — files [wired] (Atomic YAML run records establish files; evaluated plan repositories are distinct.); Task brief: known — files [wired] (The retained workspace Markdown brief is a file.); Editable extension sources, bundles and catalog: partial — files [wired] (Code and cached bundle files are established, but the catalog/registry persistence substrate is not fully characterized.); Attestation hosted state and fetched CI representation: partial — files [wired] (A fetched JSON file is consumed by CI; server-side persistence and provenance remain outside inspected implementation.); Retained skill-evaluation outcomes: partial — files [wired] (Shell/run outputs retain selected outcomes; external evaluator storage and optional Actions-summary backing storage are not inspected.); Evaluated plan/cache repositories, remote data backends and remaining retained acquisition/notification outputs: uninspected (Records retain these alternatives without inspecting their complete persistence; neither repo, SQLite, service-object nor another substrate can be supplied from labels.)	partial coverage (Automatic admission is supported for machine-generated accounts and checkpoints. Caller-authored issue content and initial acquisition retain unresolved human control; starting a command is not a per-content human decision.); Automatic method and verification report derivation/persistence: known — automatic [wired] (Software derives and admits nonempty artifacts without per-report human approval.); Hook audit append and scheduler run/checkpoint saves: known — automatic [wired] (Hooks and scheduler save activity records as part of execution.); Attestation projection and evaluation result recording: known — automatic [wired] (Scripts compute projections/results and record outcomes without per-result human approval; hosted acquisition alternatives remain separate.); Skill-prescribed human approval of the issue plan: known — manual [claimed] (The intended human explicitly decides whether the plan proceeds. This policy does not authenticate the actual approve caller.); Actual plan, feedback, findings, verification flags and explicit approval submissions: not-determinable (Caller identity does not resolve whether each retained content item was supplied, edited or approved by a human. Empty approve arguments do not supply conversational consent.); Initial brief/definition content, local extension edits and other acquisition/replacement alternatives: not-determinable (Authors and computational validators are identified, but per-content human decision versus software/model admission is not established on every route.)	partial coverage (Named request-driven routes establish pull. No complete inventory establishes pull-only behavior; internal requested loads are not relabelled push because software performs delivery.); Requested data query/get and method resource loads: known — pull [wired] (Caller predicates, resource names and optional versions request retained material; concrete issue readers establish later consumption.); Requested plan review/iteration/checks and instructed phase reload: known — pull [wired] (Method consumers request exact resources, and a later agent is instructed to request state-main.); Requested stored report, audit timeline and activity summary: known — pull [wired] (A report/history request selects surviving retained artifacts or source records; computed displays need not themselves be persisted.); Explicit workflow/task continuation and CI attestation fetch: known — pull [wired] (Resume loads stored state and CI requests the commit-associated attestation. External conversation restoration is excluded.); Automatic capability resolution and broader custom/serve/dashboard consumers: uninspected (Records establish later consumption but do not inventory every request/selection/delivery operation or alternative injection route. No global absence of push follows.)	uninspected coverage (The classified routes are requested pulls; they require no push-signal findings. Remaining included consumers prevent a boundary-wide inapplicable assessment.); Explicit data, issue, report, audit/activity and continuation requests: inapplicable (These operations fulfill requests. Resource IDs, report tags, versions and time windows on a pull do not establish an identifier/coarse push signal.); Remaining capability/custom/serve/dashboard delivery alternatives: uninspected (No accepted record fully specifies an unsolicited selector trigger, its input and delivered retained part for these alternatives. Push existence and its signal cannot be determined.)	partial coverage (Automatic trace-fed retained reports with later reference/context consumers establish a local yes. Raw logs, caller revisions and unresolved adoption alternatives are not thereby classified; no system-wide no or improved capacity is inferred.); Automatic method-summary derivation and later report/data retrieval: known — yes [wired] (Execution context and handles feed software-produced durable summaries; later report/data consumers receive selected reference context. This qualifies without parameter updates, automatic push or observed benefit.); Automatic verification-summary derivation and later report retrieval: known — yes [wired] (Recorded workflow step feedback feeds persisted summaries exposed to later report search/get consumers as execution context. This is not authenticated verification or demonstrated improvement.); Raw audit and run checkpoints, requested activity projections: not-determinable (Raw logging and unchanged execution continuation alone do not establish the qualifying transformation. Activity display is not persisted as a replacement summary. No absence record covers every included alternative.); Issue revision and attestation/evaluation-to-adoption alternatives: not-determinable (Issue content comes from callers with unresolved control and trace origin; attestation changes operational admission rather than a demonstrated durable behavioral update. Evaluation records establish no automatic successor-skill selection, while broader adoption alternatives remain uninspected.)	partial coverage (Values name actual retained parts and consumers, not a global force. Uninspected search/ranking, custom consumers and adoption routes prevent complete classification.); Issue plan/feedback/finding prose at requested review and revision readers: known — knowledge [wired] (Requested reads provide reference and criticism; retrieval does not prove obedience.); Approved plan steps at the later implementing agent: known — instruction [claimed] (The procedure says to follow the approved plan step by step; this instruction path is doctrine, not observed compliance.); Retained criticism at the instructed plan-revision consumer: known — learning [claimed] (The skill directs renewed criticism and revision of a durable plan; that update-consumer force is separate from the revised plan downstream. No improved capacity is established.); Plan/review versions, severity/resolution flags and verification flags at configured preflight checks: known — validation [wired], enforcement [wired] (These retained criteria determine admissibility and can veto approval or PR linking on the selected local paths.); Issue phase at an external agent following the reload procedure: known — routing [claimed] (The skill asks the later agent to query phase and choose the corresponding next procedure.); Workflow checkpoint/progress and captured plan at scheduler continuation: known — routing [wired], enforcement [wired] (Stored status/dependencies select remaining work; waiting gates refuse ordinary continuation.); Reloaded extension code and type catalog at execution/type resolution: known — routing [wired], enforcement [wired] (Retained type identity selects an implementation; registered code determines executable effects.); Method/verification reports and audit history at requested display/query consumers: known — knowledge [wired] (These accounts supply selected reference about execution and command history; they do not automatically become host instructions.); Attestation subject/gate/hash fields at CI admission: known — validation [wired], enforcement [wired] (Retained attestation fields are checked for admissibility; error predicates fail the job. Warning-only freshness does not veto.); Remaining data search/ranking, evaluation/adoption and custom payload consumers: uninspected (Search scoring and downstream adoption/custom behavior were not fully traced. No ranking positive, general learning negative or universal authority conclusion follows.)	AAS-2026-10-15-swamp-440d19cc0bcc-01
SwarmWorld	2	Use-generated private evidence and research state, retrieval metadata, empirical cells, public archive, spatial insights, message/request ledger, research mission history, program registry/skills and installed retained instructions; recorded traces, replay/playback and included but uninspected offline analysis/dossier transformations. Excludes shipped static prompts/catalogs, pre-execution queues, ordinary physical world arrays and external provider state/weights.	partial coverage (Live memory and its file-backed recorded representation are separately inventoried. External provider state, static packages and ordinary physical arrays are excluded. Offline alternatives remain included and unresolved; positives do not establish a complete system union.); Live agent-memory: known — in-memory [wired] (The record identifies the operative collections; serialization is classified separately.); Live research-state: known — in-memory [wired] (The record identifies the operative collections; serialization is classified separately.); Live retrieval-stats: known — in-memory [wired] (The record identifies the operative collections; serialization is classified separately.); Live empirical-knowledge: known — in-memory [wired] (The record identifies the operative collections; serialization is classified separately.); Live cultural-archive: known — in-memory [wired] (The record identifies the operative collections; serialization is classified separately.); Live deposited-insights: known — in-memory [wired] (The record identifies the operative collections; serialization is classified separately.); Live message-records: known — in-memory [wired] (The record identifies the operative collections; serialization is classified separately.); Live research-mission: known — in-memory [wired] (The record identifies the operative collections; serialization is classified separately.); Live program-library: known — in-memory [wired] (The record identifies the operative collections; serialization is classified separately.); Live program-skills: known — in-memory [wired] (The record identifies the operative collections; serialization is classified separately.); Recorded events, model decisions, prompt templates and snapshots: known — files [wired] (Files are the retained substrate, not the JSON encoding.); Included offline analysis/dossier transformations and downstream consumers of recorded research state: uninspected (Available offline transformation and consumer internals were not inspected. Their additional classifications cannot be enumerated from accepted records.)	partial coverage (Human control is local to explicit content-bearing decisions, not inherited by all writes in an interactive episode. Offline alternatives remain included and unresolved; positives do not establish a complete system union.); Sensing, measurement, retrieval feedback, recording, reconstruction and inheritance admission: known — automatic [wired] (Starting the episode or replay is not a content approval decision.); Model-generated research-state admission: known — automatic [wired] (Optional outage rollback is a separate software control.); Model/script social content and program proposals: known — automatic [wired] (Different controller routes are not homogenized as manual.); Human-play content-bearing action proposals: known — manual [wired] (Only the human content decision is manual; resulting sensing, receipts and metadata updates remain automatic.); Included offline analysis/dossier transformations and downstream consumers of recorded research state: uninspected (Available offline transformation and consumer internals were not inspected. Their additional classifications cannot be enumerated from accepted records.)	partial coverage (Request, automatic selection and later delivery are distinguished within each chain. Offline alternatives remain included and unresolved; positives do not establish a complete system union.); Scheduled macroturn context selection and delivery: known — push [wired] (This includes latest research state, private/public retrieval and empirical, spatial, skill and request views.); Controller-requested teaching/resolution and recorded reconstruction: known — pull [wired] (Requested return is not counted again as push merely because code delivers it.); Automatic request-result receipt and replacement inheritance: known — push [wired] (Receipt routing differs from the controller request to teach a named record.); Included offline analysis/dossier transformations and downstream consumers of recorded research state: uninspected (Available offline transformation and consumer internals were not inspected. Their additional classifications cannot be enumerated from accepted records.)	partial coverage (Signals classify automatic push selection only; source-ID resolution on requested pull routes is not added as another identifier push. Offline alternatives remain included and unresolved; positives do not establish a complete system union.); Macroturn private retrieval and compact archive selection: known — inferred-lexical [wired] (Scored retained text is the selected part; optional feedback modifies ranking.); Latest state, ordinary archive, nearby insights, empirical summaries and context budget selection: known — coarse [wired] (These predicates are not an identity match merely because records have IDs.); Program focus and recipient-specific pending requests: known — identifier [wired] (Program ranking also uses verified status and recency, a separate coarse selection input.); Replacement skill selection: known — coarse [wired] (Selection is not a content-understanding or embedding inference.); Reply-result receipt routing: known — identifier [wired] (This is an actual match controlling unsolicited delivery, not merely an identifier on a requested file.); Included offline analysis/dossier transformations and downstream consumers of recorded research state: uninspected (Available offline transformation and consumer internals were not inspected. Their additional classifications cannot be enumerated from accepted records.)	partial coverage (The axis tests automatic trace-fed durable artifacts and later consumers, independently of demonstrated learning benefit. No bounded system-wide negative is inferred from excluded provider weights. Offline alternatives remain included and unresolved; positives do not establish a complete system union.); Automatic model research-state continuation write: known — yes [wired] (A generated continuation artifact qualifies without observed improvement or verified explanatory novelty.); Citation/action feedback to durable selector metadata: known — yes [wired] (This is a behavior-shaping access artifact, not a weight update.); Empirical/experiment summaries and measured program indices: known — yes [wired] (Raw log retention alone is not an additional qualifying route.); Included offline analysis/dossier transformations and downstream consumers of recorded research state: uninspected (Available offline transformation and consumer internals were not inspected. Their additional classifications cannot be enumerated from accepted records.)	partial coverage (Actual consumers, not labels such as verified or publication, determine authority. Offline alternatives remain included and unresolved; positives do not establish a complete system union.); Retained evidence and research prose at later model/viewer context: known — knowledge [wired] (Delivery is not demonstrated reliance; research-state goals are not automatically standing host instructions.); Private/archive text, salience and feedback counters at retrieval; skill evidence at ordering: known — ranking [wired] (Ranking authority does not require a learned ranking model.); Retained resource evidence, publication provenance and program membership at action validators: known — validation [wired], enforcement [wired] (Criteria and enforcing veto coexist at this consumer.); Request ledger and committed trace fields at routing consumers: known — routing [wired] (Routing is separate from interpreting request prose.); Retained selections, citations, outcomes and measurements at automatic update consumers: known — learning [wired] (This update-consumer force is separate from the downstream knowledge and ranking forces; benefit is unobserved.); Installed retained artifact instructions at repeated VM execution: known — enforcement [wired] (This covers the executable memory part, not arbitrary prose or host permissions.); Included offline analysis/dossier transformations and downstream consumers of recorded research state: uninspected (Available offline transformation and consumer internals were not inspected. Their additional classifications cannot be enumerated from accepted records.)	AAS-2026-10-15-swarmworld-e39fbc17e3fb-01
Synapptic	2	Includes accumulated observation content/envelopes, weighted profile text/metadata/history, saved and published narratives, benchmark cases/results/reduced verdicts, opaque resumed evaluator context, original transcript read-back, relay request displays, transcript indexes and restored session bookkeeping, with their write, maintenance and later-consumer routes. Excludes static shipped patterns/skills/configuration as memory, ordinary transient current-call state, external host instruction admission/compaction and model weights/provider internals other than the included opaque resumed context.	known coverage (The recorded inventory covers local files, relay SQLite and reconstructed tracker state, and the external resumed-session object. Service-object describes the external interface, not an assertion about its hidden physical medium.); Observations/envelopes, profiles/history, narratives/published copies, evaluator cases/results and original transcript files: known — files [wired] (All these retained parts have recorded file-backed persistence; JSON/YAML/Markdown are encodings, not extra substrates.); Relay request displays, transcript index and session bookkeeping: known — sqlite [wired], in-memory [wired] (The records distinguish database retention from reconstructed operative tracker objects.); Claude evaluator condition-session retention: known — service-object [wired] (The session is externally retained; the handle alone is not its substrate. Internal storage technology is not classified.)	partial coverage (Known automatic admissions coexist with explicit human-controlled exclusions/replacement. Initial external transcript admission and hidden provider checkpoint writes are not resolved by a generic caller or a resume read.); Extraction, merge/promotion, synthesis/publication, evaluator cases/results/verdicts, relay capture/index/bookkeeping: known — automatic [wired] (Humans start/install machinery; models/software admit this content without per-item human decisions.); Global exclusion confirmation, explicit re-inclusion and rollback replacement: known — manual [wired] (Human decisions approve exclusion changes or request restoration of the previous global checkpoint; software performs selection/I/O.); Original external transcript production and provider-internal resumed-context/checkpoint admission: not-determinable (External logs are available and handles are chosen by code; the records do not establish per-content admission control for every original/hidden write.)	partial coverage (Explicit requested returns and automatic model-context supply are separate operations. Hidden provider selection and incompletely inspected live UI paths retain local uncertainty.); Requested merge/synthesis/integration/results/history/browser retrieval: known — pull [wired] (Drivers/operators request retained files, cases, snapshots or UI material. Fulfilling these requests is pull, not a second push merely because code delivers bytes.); Profile context automatically supplied at extraction and eligible profile/case treatment supplied to model calls: known — push [wired] (The extraction, synthesis and benchmark triggers cause code to select and attach retained context without a model retrieval request.); Published guidance at external host: known — push [afforded] (Static shipped instructions are excluded; these files contain accumulated generated memory. Host loading is documented/afforded, not observed.); Condition-session resume request and hidden payload selection; full live UI delivery: partial — pull [wired] (Code explicitly requests session reuse, establishing pull at the interface. Internal context selection/delivery and full live-update routes are not inspected.)	partial coverage (Known push selectors include coarse eligibility/treatment, identifier matching and inferred-lexical matching of guard text to retained narrative blocks in per-guard benchmarks. Embedding or judgment retrieval is not established. Requested FTS/browser search is pull and does not add a push signal. Hidden provider and live UI alternatives remain unresolved.); Extraction and synthesis automatic context selection: known — coarse [wired], identifier [wired] (Trigger: session extraction or synthesis. Inputs: global availability, weight/count/dimension order, exclusions and optional target-model name. Parts: profile observations and associated fields.); Benchmark model prompt treatment and cached case reuse: known — coarse [wired], inferred-lexical [wired], identifier [wired] (Trigger: explicit benchmark. Inputs: mode, project/model/seed/hash cache identity and chosen guard text. Parts: retained cases and saved narrative treatment, including lexical selection of matching archetype blocks in per-guard mode. Coarse treatment choice, lexical block selection and identifier-based case reuse are distinct operations.); External host loading of published generated guidance: known — coarse [afforded] (Trigger: host instruction/memory loading. Inputs: destination/discovered project and output conventions. Parts: combined published narrative; current task content does not select passages.); Opaque resumed-session payload and uninspected live UI push selection: not-determinable (A requested resume identifier is not by itself a push selector. The actual hidden delivery operation and full live UI selection are not established.)	partial coverage (Automatic session-derived durable updates and benchmark feedback qualify without observed benefit or parameter changes. Hidden resumed-session transformations and the complete observability write/consumer inventory are unresolved, not system-wide negatives.); Session conversation to observations, profiles and saved/published narrative: known — yes [wired] (Automatic trace-fed extraction produces durable artifacts reused by future extraction/synthesis and afforded host guidance. Hook maintenance alone omits external publication.); Benchmark execution feedback to durable profile eligibility: known — yes [wired] (Automatic per-model verdicts change future explicit-target synthesis inputs; human global exclusions remain a separate mechanism.); Opaque evaluator context maintenance and complete relay/live UI alternatives: not-determinable (Resume establishes retained consumption, not any particular automatic derived checkpoint write. Relay storage/operator read-back is established but no qualifying profile-update route or bounded system-complete absence record is supplied.)	partial coverage (Authority is assigned at actual consumers, separately from write agency. Opaque resumed payloads and uninspected live display paths prevent a complete inventory.); Observations, profiles and evaluation experience at durable update consumers: known — learning [wired], ranking [wired] (Retained experience changes durable profiles, narratives and future eligibility; this is learning authority at update consumers, not demonstrated improvement.); Selected profile context at later extraction and saved descriptive narratives at evaluators: known — knowledge [wired] (These consumers receive reference context. Preventive imperatives have a separately scoped instruction path.); Published imperative guards at external assistant: known — instruction [afforded] (The documented host-loading route is afforded; loading, precedence and compliance are not observed.); Observation envelope and profile selectors at processed-selection, merge and synthesis: known — routing [wired], ranking [wired], validation [wired], enforcement [wired] (IDs select work/scope, weights/counts rank or filter, and exclusion/verdict fields veto source eligibility. Exclusions are not read by the extraction summary.); Benchmark cache identities and cached criteria at evaluator consumers: known — routing [wired], validation [wired] (Cache fields choose reuse; retained rule/criteria content defines the model compliance check, without an independent factual oracle.); Detailed results, relay displays and transcript index at operator/UI and software consumers: known — knowledge [wired], ranking [wired], routing [wired] (Detailed results inform operators, not synthesis directly; only reduced profile verdicts reach eligibility filtering.); Opaque resumed condition history and incompletely inspected live UI consumers: not-determinable (Later consumption exists, but the complete consumed payload/effects and live display paths are not recorded well enough to enumerate their authority.)	AAS-2026-10-15-synapptic-9dce97c66546-01
Synthadoc	2	Accumulated wiki/candidate bodies and frontmatter; extracted sidecars; stored chat turns and overflow summaries; lifecycle/fix snapshots; computation/answer and faithfulness caches; generated overview, purpose, index, routing and agent guidance; page-vector and BM25 access structures; exports and backup/restore copies; cross-wiki retained page/control consumption; their admission, maintenance and later-consumer routes. Included uncertainties: initial content/registry control, embedding population/refresh, unestablished summary/overview delivery, other workflow/client feedback, peer backing stores, hook-produced state and opaque coding-agent checkpoints at the integration boundary. Excluded: fixed shipped package instructions/protocols, model weights and provider internals, arbitrary external host implementations, current-run-only tool messages, transient request strings and raw source files except as inputs to retained derivations. No deployed operation or causal benefit is asserted.	partial coverage (File, SQLite, vector and in-memory witnesses classify inspected retained parts and access structures. Peer storage and opaque integration state prevent a complete substrate inventory. Configured model weights and shipped static prompts are outside this accumulated-memory scope.); Wiki and candidate Markdown bodies/frontmatter; extracted text/pagemaps; faithfulness JSON; overview; generated purpose/index/routing/instructions; exported bundles and backup archives: known — files [wired] (The records identify the file-backed surfaces and separate live pages, staged pages, sidecars and snapshots. Raw originals are source inputs, not an additional inferred memory substrate.); Stored chat turns/summary, lifecycle snapshots/events, fix history in AuditDB, response/query caches and page-vector rows: known — sqlite [wired] (The accepted object records identify each durable store; page fix pointers also reside in file frontmatter.); Persisted page embeddings used to rerank BM25 candidates: known — vector [wired] (Vector is the operative access substrate alongside its SQLite persistence, not a claim that page vectors are trained model weights.); Reloadable process-cached BM25 corpus/token lists as access structures: known — in-memory [wired] (Includes access structures built over retained pages; excludes current tool-loop messages and transient request strings.); Cross-wiki peer physical stores and registry backing: not-determinable (The coordinator reads peer pages and registry purpose summaries, but peer deployments and registry initialization are not audited. HTTP delivery does not establish service-object or any specific underlying peer storage.); Included opaque coding-agent checkpoints and hook-produced retained state: uninspected (The accepted records leave native checkpoints and hook-produced state uninspected. Their storage substrates cannot be classified from subprocess invocation or executable capability alone. External implementations are excluded; the unresolved retained-state alternative at the integration boundary is not silently excluded.)	partial coverage (Automatic content admission and an explicit operator-confirmed CLI branch are supported. Generic API callers and session confirmation do not establish manual agency. Unknown initial, client-mediated and opaque write mechanisms keep this axis partial.); Ingest/annotation, overflow-summary/cache admission, overview/scaffold regeneration, lint/automatic lifecycle/warning changes, verdict caching and enabled server masking: known — automatic [wired] (Includes known subsequent transformations regardless of unresolved initialization. Default/force/staging alternatives are not merged into a human-review guarantee.); CLI sensitive masking on the confirmed --apply path: known — manual [wired] (Only the confirmed CLI branch is classified manual. The --yes bypass does not establish a fresh per-content human decision.); Candidate promotion/discard, direct MCP replacement, workflow client-gated fixes/rollback and HTTP snapshot replacement: not-determinable (These callers supply content, movement decisions or approval booleans, but accepted records explicitly leave human identity/control unresolved. A client confirmation is not automatically a human decision; software I/O is not automatic admission independent of the caller either.); Initial content and purpose/routing/registry admission, embedding refresh/population and backup restore caller control: not-determinable (Subsequent scaffold software adoption is known; exact deployed initialization and restore/refresh control are not. Engine capability alone cannot classify operator approval.); Included opaque coding-agent checkpoints and hook-produced retained state: uninspected (The accepted records leave native checkpoints and hook-produced state uninspected. Their write/admission control cannot be classified from subprocess invocation or executable capability alone. External implementations are excluded; the unresolved retained-state alternative at the integration boundary is not silently excluded.)	partial coverage (Both requested pull and automatic context/control supply are supported at separate operations. Internal ranking in a requested query is not re-counted as push. Unknown host, summary/overview and recovery consumption prevent complete direction coverage.); Requested local/cross-wiki query, context pack, direct read, export/bundle answering and cached requested-answer return: known — pull [wired] (Pull includes internal selectors serving the request. Actual external host use beyond returned material remains unobserved.); Requested repair page reads, audit displays and historical snapshot/fix-record recovery: known — pull [wired] (A requested slug/index on rollback is not itself targeted push.); Automatic recent-session history supply to a later streaming query: known — push [wired] (Separates history enrichment from the requested wiki answer. The overflow summary text is not included in this push.); Standing purpose/routing and retained control state read automatically by later ingest, retrieval, auditing, scanning and overflow decisions: known — push [wired] (Control consumption is distinct from independent foreground prose delivery. Background writes alone are not push.); Overview/overflow-summary semantic read-back, backup-restored payload reload and external generated-instruction loading: not-determinable (Summary/overview production does not establish semantic read-back; restoration makes bytes available but actual reload is unobserved; the host loading route is only afforded and does not fix direction.); Included opaque coding-agent checkpoints and hook-produced retained state: uninspected (The accepted records leave native checkpoints and hook-produced state uninspected. Their later read-back chains cannot be classified from subprocess invocation or executable capability alone. External implementations are excluded; the unresolved retained-state alternative at the integration boundary is not silently excluded.)	partial coverage (Identifier and coarse signals describe supported automatic supply operations. Lexical and embedding relevance in requested knowledge retrieval do not automatically become push signals. Incomplete direction coverage prevents a pull-only/inapplicable assessment.); Automatic session-history push on /query/stream: known — identifier [wired], coarse [wired] (Actual row identity selection and subsequent coarse tail selection are separate operations. Stored summary text is not selected for synthesis.); Automatic standing purpose/routing and global controls supply: known — coarse [wired] (Downstream routing keyword tests can select work/scope without establishing that retained routing-file fragments were themselves selected by lexical matching for delivery.); Automatic cached decision reuse, per-page provenance/control and stale-audit state: known — identifier [wired] (The selected retained control record has an identity consumer; this differs from merely naming a requested file. Does not classify requested cached-answer delivery as unsolicited push.); Other included host/opaque, overview/summary and restored-payload selection: not-determinable (Direction/selection is unresolved for these alternatives. A known pull route cannot make their push signal inapplicable; a display summary cannot identify an opaque selector.)	partial coverage (A wired yes is established for imported conversational trace compilation. Other retained traces, summaries and control histories do not supply bounded system-wide negatives. Unresolved routes remain included, so this axis is partial rather than known or absent.); Imported JSONL conversational trace to admitted durable wiki page to later eligible query: known — yes [wired] (Manual workflow initiation is not per-page approval. The positive is architectural wiring, not proof that a transcript was ingested or improved future answers.); Overflow summary production and raw-tail reload: not-determinable (Summary writing is automatic and trace-fed, but semantic summary delivery to a later answering consumer is not established. Raw-tail reload alone is not a learned transformation. Summary/count control reads do not establish use of the summary content as behavior-shaping guidance.); Fix/recovery history, faithfulness evaluation cache, scan metadata, ordinary cached computations and benchmark score retention: not-determinable (These records establish retention and some later control/review consumers, but do not completely establish an automatic trace-fed transformation into a durable behavior-shaping artifact on each route. Snapshotting bodies, storing raw evaluation outcomes or replaying cached text is not by itself qualifying trace learning. No local no is inferred without a bounded absence warrant.); Other workflow/client feedback and initialized guidance alternatives: not-determinable (The representative workflow changes products without revising shipped guidance; that is not a complete absence search for every trace-fed route. Initial guidance provenance and other feedback paths remain unresolved.); Included opaque coding-agent checkpoints and hook-produced retained state: uninspected (The accepted records leave native checkpoints and hook-produced state uninspected. Their automatic trace-fed writes and later consumers cannot be classified from subprocess invocation or executable capability alone. External implementations are excluded; the unresolved retained-state alternative at the integration boundary is not silently excluded.)	partial coverage (Knowledge, instruction, ranking, routing, enforcement, validation and trace-update learning have distinct local consumers. Known generation or permission cannot upgrade opaque alternatives. Coverage remains partial because semantic summary/overview delivery and external state consumers are unresolved.); Page/sidecar bodies, raw recent session turns, cached answers and restored bodies at answering, maintenance, source-review and requested export consumers: known — knowledge [wired] (The named consumers are wired. External direct readers receive text, but their subsequent model use is only afforded.); Generated purpose at later ingest and generated instruction files at external host loading: known — instruction [wired], instruction [afforded] (The wired purpose path and afforded external instruction path keep their different warrants; actual host delivery/compliance is not claimed.); Page text/titles/tags/aliases and stored vectors at lexical scoring and optional vector reranking: known — ranking [wired] (Both lexical inputs and dense ranking features are included, without implying learned scoring policy.); Lifecycle eligibility and retained warning count at ordinary search and auto-resolution gates: known — enforcement [wired], validation [wired] (The veto is consumer-specific. It does not cover direct reads/default OKF export or certify factual acceptance.); Routing/index/registry state, provenance/dedup fields, cached action fields, summary coverage counters, faithfulness freshness keys, scan-cycle timestamps and recovery snapshot/fix pointers at later software decisions: known — routing [wired] (Names actual selector effects. Semantic summary text and faithfulness reasons do not become answering instructions merely because adjacent counters/keys are consumed.); Retained imported session text at the compilation/update consumer: known — learning [wired] (This path is distinct from the derived page knowledge authority and does not entail weight training or revised general methods.); Cached faithfulness reasons/verdicts at reviewing clients and CLI result consumers: known — knowledge [wired], enforcement [wired] (CLI status has a bounded software effect. Unknown downstream shell policy is not a new enforced page-admission gate.); Overview, semantic overflow summary and opaque backup payloads at possible downstream semantic consumers: not-determinable (Creation/storage is established, but a later model consumer of overview or overflow-summary text is not. Backup manifest inspection does not resolve every restored payload consumer. No authority is assigned to unconsumed metadata or text.); Included opaque coding-agent checkpoints and hook-produced retained state: uninspected (The accepted records leave native checkpoints and hook-produced state uninspected. Their retained parts, actual consumers and effects cannot be classified from subprocess invocation or executable capability alone. External implementations are excluded; the unresolved retained-state alternative at the integration boundary is not silently excluded.)	AAS-2026-10-15-synthadoc-a3cae9152fff-01
Synto	2	Retained source documents/segments and raw notes; derived analyses, summaries, articles/drafts, saved Q&A/synthesis; identities, denials, relations, routing indexes and rebuild seeds; rejection guidance; checkpoints and response cache; metric/audit and comparison reports; exported content and vault-derived pack guidance; optional injected RAG write payload and unresolved store/consumer. Static shipped prompts, current-call-only answers/context and external provider weights or mutation are excluded: no use-updated provider memory is established within Synto. External hosts are excluded implementations, but shipped pack consumer affordances remain included.	partial coverage (Files, SQLite and graph organization are supported. The included injected-store alternative prevents complete substrate coverage; external inference weights are not asserted to be use-updated Synto memory.); Vault and pack files: originals, raw Markdown, articles/drafts, summaries, saved answers, indexes, identity seed, guidance and comparison reports: known — files [wired] (These retained filesystem objects have inspected write and later-reader paths.); SQLite analysis, source segments, identities/denials, relations, checkpoints, response cache, rejection and metrics state: known — sqlite [wired] (The accepted inventory names the database objects separately from exported representations.); Retained relation edges and occurrence associations: known — graph [wired] (Graph-organized memory coexists with its SQLite persistence.); Injected RAG store persistence: not-determinable (add_document arguments do not disclose the store substrate or durability; dense encoding alone does not establish a vector storage backend.)	partial coverage (Manual content decisions and automatic subsequent admission coexist. Unknown initial control and injected-store policy prevent complete write coverage.); Explicit import selection/replacement, draft editing/approval/rejection reasons and identity curation: known — manual [wired] (These paths require a content-directed operator decision, not merely workflow launch.); Analysis, candidate filtering, summaries, drafts, automatic publication, answer saves, checkpoints/cache, seed restoration, export and metric retention: known — automatic [wired] (Starting a run, save flag, clean or export does not approve each subsequently generated item.); Source import extraction and segment admission after file selection: known — automatic [wired] (Separate from the manual decision supplying the original document.); Visible optional embedding submission and unknown injected-store admission/initialization: partial — automatic [wired] (The API write is software driven; internal acceptance/replacement and initial state are uninspected.); Initial control of externally supplied raw notes and preexisting seed material: not-determinable (External material may be supplied or restored; records do not establish each original admission principal.)	partial coverage (Request-led query/source consumption and automatic auxiliary/recovery selection coexist. Pack host behavior and optional store recall remain unresolved.); Question-driven vault/MCP answer context, explicit article/source reads and operator report inspection: known — pull [wired] (A consumer requests retained material or an answer using it; automatic delivery within that request is not a second unsolicited push.); Compiler/ingestion automatic auxiliary context: concept hints and recent rejection feedback: known — push [wired] (A run/compile trigger causes selectors to supply prior hints/feedback without a separate request for those retained parts.); Automatic checkpoint and exact response reuse on later ingestion/inference attempts: known — push [wired] (The caller requests work or inference; automatic selectors substitute compatible retained prior results.); Identity seed restore on later ingest: known — push [wired] (Code selects preserved identities and constraints into reconstruction without a per-row read request.); Exported pack consumption under generated guidance: partial — pull [afforded] (Files permit later requested inspection; host startup injection is not established.); Injected RAG later consumer: not-determinable (Only the write is established; neither requested recall nor automatic supply is classified.)	partial coverage (Push signals classify actual auxiliary/resume/cache selectors, not IDs on requested source files. Query model relevance and BM25 belong to request-led pull routes, not inferred push signals.); Compilation/ingestion trigger: concept hints selected from retained concepts using current body/title matches and caps: known — inferred-lexical [wired], coarse [wired] (Targeted lexical matching is separate from the coarse list budget.); Recompile trigger: latest feedback strings for the current concept: known — identifier [wired], coarse [wired] (The concept key selects the retained rejection records; recency and the three-record bound are separate.); Ingest restart trigger: compatible saved analysis units: known — identifier [wired] (A compound identity match actually selects reused payloads.); Inference request trigger: exact-key response substitution: known — identifier [wired] (No similarity lookup is implemented in this cache.); Later ingest trigger: preserved identity seed restoration: known — coarse [wired], identifier [wired] (Selection reads available seed entries and restores identities while respecting existing entities.); Opaque injected-store selection and external pack host injection: not-determinable (No selector/input/delivered-part chain establishes a push signal on these alternatives.)	partial coverage (A wired review-feedback revision chain supports yes locally. Checkpoint/cache qualification, external source origins and optional-store behavior prevent a complete trace-learning inventory; no system-wide no is inferred.); Retained human draft criticism to automatic revised article to later answer/export consumers: known — yes [wired] (The per-content feedback write is manual, but subsequent feedback-fed article generation/retention is automatic and behavior-shaping across invocations.); Checkpoint analysis and response reuse: not-determinable (These are automatic trace-fed retention routes with later consumers. Records describe replay/merge, but do not fully resolve whether each retained result and subsequent derived artifact qualifies as a behavior-shaping update rather than unchanged result reuse.); Metrics/audit and comparison reporting: not-determinable (Known traces reach operator reports; no automatic production-update consumer is established. Records do not declare a bounded absence across every downstream human-mediated use.); Source-derived knowledge and optional RAG inputs: not-determinable (External raw-note contents/origins and injected-store later consumers are unresolved. Source import alone is not evidence of trace learning, nor of its absence.)	partial coverage (Distinct retained parts support knowledge, instruction, learning, ranking, routing, validation and enforcement. Host-dependent guidance stays afforded; injected-store effects remain unresolved.); Raw prose, summaries, published/saved article slices and source passages at writing, answering or external reading consumers: known — knowledge [wired] (These objects supply context/evidence, not a factual acceptance guarantee.); Rejection feedback at heavy writer: known — instruction [wired], learning [wired] (The same experience has separate update and downstream effects.); Generated pack guidance at an external agent host: known — instruction [afforded], routing [afforded] (Rendering and disk delivery are wired, but later host loading and interpretation are only afforded.); Retained aliases/denials and identity seed at normalization, query expansion and restore code: known — routing [wired], validation [wired], enforcement [wired] (Fixed-rule identity decisions have distinct force from article prose.); Retained status/hash/confidence/block state at scheduler and admission/overwrite consumers: known — routing [wired], validation [wired], enforcement [wired] (These values control eligibility rather than certify truth.); Retained catalogues/text/aliases and graph occurrence/confidence fields at retrieval selectors: known — ranking [wired], routing [wired] (Ranking inputs need not be learned ranking policies.); Checkpoint and cache identity/payload at resume and response consumers: known — routing [wired], knowledge [wired] (Payloads supply prior results, while keys and compatibility fields determine reuse.); Metrics/audit and comparison reports at examining operator: known — knowledge [wired] (The recorded later consumer is human inspection, not an automatic winner installer.); Injected RAG payload at unknown later consumer: not-determinable (No consumer is established; storage/add_document alone cannot assign authority.)	AAS-2026-10-15-synto-eabcdcce3da2-01
talorys	2	Application-retained personal facts and lifecycle/review state; conversation messages, activity and generated summary/cutoff; saved tasks, notes and automation instructions/dispatch state; deferred action payloads/status; memory/item vectors and their metadata/caches; their admission, maintenance and later model/owner/update consumers. Excludes shipped static policy/tool definitions, general runtime settings/auth/budget counters, transient active-run controllers, deployment artifacts, external provider parameters/private retention and ordinary current-turn tool accumulation. Notifications/receipts are included only as owner-readable retained results and recorded action feedback, not assumed recurrent model-training inputs.	known coverage (The scoped application inventory has SQLite source stores and vector access stores with decoded in-memory caches. External service weights, provider-private state, deployment files and runtime-only active controllers are excluded, not classified as application memory.); Personal facts and lifecycle metadata; conflict-review queue: known — sqlite [wired] (Both content and its access metadata have inspected app-level stores.); Conversation messages, activity display records, summary and summary cutoff: known — sqlite [wired] (The summary is retained by value; no alternative app checkpoint is shown.); Memory and task/note/automation semantic indexes: known — sqlite [wired], vector [wired], in-memory [wired] (Vector describes the operative retrieval store, not an inference-model weight update; decoded caches are separate from durable rows.); Saved tasks, notes, routine prompts/dispatch state, and deferred approval objects: known — sqlite [wired] (Shipped edit, scheduled, approval and later-read routes cover these additional retained surfaces.)	known coverage (The included admission mechanisms separate human content decisions from software/model decisions and from physical I/O.); Immediate model memory/task/note/routine writes and model corrections: known — automatic [wired] (Ordinary model correction can retire a predecessor without an approval card.); Owner manual supply/edit/correct/status/review and exact destructive proposal approval: known — manual [wired] (The decision is content-specific, even though code performs persistence; actual owner decisions are unobserved.); Automatic summary, derived index, rule/Clef review and lifecycle maintenance writes: known — automatic [wired] (Configuration opt-in does not change subsequent automatic admission.); Backup selection and per-field merge: known — manual [wired], automatic [wired] (File-level human selection and row-level merge are separate mechanisms, not evidence of manual factual review of every row.); Raw turn/activity persistence and scheduled receipts: known — automatic [wired] (User authorship of a message is separate from automatic host logging admission.)	known coverage (Both requested and unsolicited later-consumer routes are wired. The inventory includes owner readers and automatic maintenance as well as foreground delivery.); Automatic memory and item selection at new chat/scheduled invocations: known — push [wired] (Memory push covers keyword/semantic branches; item push is semantic foreground-only.); Conversation summary/history and refreshed approval outcomes: known — push [wired] (A request to start chat is not a model request for each retained context part.); Stored routine at alarm dispatch: known — push [wired] (Reminder/digest deterministic branches remain distinct from the model prompt branch.); Explicit tool search/read and owner list/history/search/review reads: known — pull [wired] (Requested query/ID returns are not an additional push because code delivers them.); Automatic maintenance consumption of saved content and indexes: known — push [wired] (These are update consumers, not additional foreground retrieval requests.)	known coverage (Each push signal is attached to a trigger, selector input and retained selected part. Tool and owner-requested IDs belong to pull, not independent identifier push.); Keyword/hybrid personal-memory push: known — inferred-lexical [wired], inferred-embedding [wired], coarse [wired], identifier [wired] (Trigger is a new chat or scheduled prompt; selectors use current text, vectors, lifecycle links, recency and limits. Identity-based additions are separate from coarse filtering.); Foreground item push: known — inferred-embedding [wired], coarse [wired] (Trigger is semantic-enabled foreground context assembly; delivered text/excerpts are loaded after numerical selection.); Conversation history/summary and approval-outcome push: known — identifier [wired], coarse [wired] (A conversation identity actually selects unsolicited context; this is not merely an ID displayed in a catalog.); Stored routine push: known — identifier [wired], coarse [wired] (Trigger is SDK callback; retained matching routine supplies the later prompt.); Automatic diagnosis, repair and compaction input selection: known — inferred-lexical [wired], inferred-embedding [wired], identifier [wired], coarse [wired] (Clef judges nominated pairs after selection; its relationship judgment is not itself a read-back selector delivering fact text to the foreground model.)	partial coverage (Qualifying summary and remembered-fact chains are wired. Other retained prose/import origins are not completely resolved, so positives do not become a system-wide complete trace-learning classification.); Automatic conversation compaction to retained later summary: known — yes [wired] (This qualifies without improved capacity, parameter training, verified fidelity or novel propositions.); Conversational remember to durable personal facts: known — yes [wired] (Explicit remember requests do not create per-content human approval of the generated wording.); Other model-authored notes/routines, imported provenance and opaque upstream derivations: not-determinable (Records establish retained authoring/import and later consumption but not every input derivation or qualifying trace-fed update for these alternatives; raw result logging alone is insufficient.)	known coverage (Authority follows each actual consumer. Information, routine guidance, trace-fed update inputs, ranking features, eligibility vetoes, approval criteria and routing metadata have different forces.); Selected fact/note/task text, conversation history/summary and review explanations at information consumers: known — knowledge [wired] (Data-not-instructions is intended consumer force, not demonstrated provider compliance.); Stored automation message at scheduled model: known — instruction [wired] (This is across-run guidance, unlike the informational recall section.); Retained session text and prior summary at automatic fact/summary update consumers: known — learning [wired] (Learning authority names the update input force; it does not assert improved capacity or parameter training.); Retained fact text, dates/preferences and semantic payloads at rankers: known — ranking [wired] (Fixed scoring code still consumes retained ranking inputs.); Lifecycle/status and existence controls at recall; pending status/expiry/hash at approval: known — enforcement [wired], validation [wired] (Eligibility veto is not a universal truth gate or historical-data erasure. Snapshot validation applies to approval, not every write.); Summary cutoff, successor links, vector metadata and stored routine dispatch fields: known — routing [wired] (Only fields consumed by actual operations are assigned force; inert provenance labels receive none.)	AAS-2026-10-15-talorys-7bce41ba9401-01
Tardigrade	2	Included: through-use event/conversation content, current and legacy derived summaries and control state, retained continuation payloads and attachment refs/bytes, effect checkpoints and adopted initial state, current named-text workspace, legacy spills/ref manifests, retained permission and output-rejection decisions, later-consumed installed actor revisions, and prescribed optimization candidate/source/evaluation lineage. Custom backends, opaque payload content, arbitrary caller provenance and uninspected host/maintenance alternatives remain included uncertainties. Excluded: external model weights/provider internals, remote dependency implementations, shipped unchanged static prompts/instructions as memory, ordinary transient work without a later retained consumer, and scratch SQL databases not established as a retained memory corpus. Optimization instructions are evidence of a prescribed memory route, not themselves accumulated memory.	partial coverage (Known local memory, SQLite, KV and file witnesses do not enumerate caller-selected backends or uninspected host alternatives.); Current event-backed trajectory, summaries/control state, continuation and content refs, permission/rejection decisions, recorded code outcomes, adopted state and effect checkpoints; configured SQL journal: known — in-memory [wired], sqlite [wired] (Classification covers the inspected configured journal and materialized projections, not every custom journal.); Legacy event-backed conversation and summary checkpoint in the shipped Bun server: known — sqlite [wired] (Legacy storage and summary consumption are separately supported; current checkpoint machinery is not substituted for this path.); Current named-text workspace in the shipped example: known — in-memory [wired] (The example backend is inspected and has only layer-lifetime retention.); Legacy spilled JSON values and ref manifest in the inspected Bun binding: known — kv [wired], sqlite [wired] (The KV interface and its inspected Bun database backing coexist; this does not classify all platform bindings.); Installed actor modules/manifests and prescribed optimization source/evaluation artifacts: known — files [wired], files [claimed] (The two evidence layers remain separate; no optimizer artifacts or exercised promotion were supplied.); Attachment byte backends, custom journals/workspaces, remaining legacy backup and Cloudflare spill bindings: uninspected (Backend choice and host-specific retention are unresolved; ObjectRef encoding does not establish files, service-object or another substrate.)	partial coverage (Automatic admission is supported on specific routes. No explicit per-content human admission mechanism is recorded, so manual is not asserted; that is uncertainty, not a system-wide absence.); Journal/reducer projection, current and legacy summary admission, checkpoint production and inspected final-code spill threshold: known — automatic [wired] (Human initiation of a message does not make subsequent journal/summary admission manual; only the inspected spill producer is covered.); Current model-tool workspace write path and output-rejection state admission: known — automatic [wired] (This is the evidenced software/model path, not a classification of every direct workspace caller.); Prescribed optimizer candidate editing and selection: known — automatic [claimed] (User scope and thresholds do not establish manual admission of each revision. Exercise remains unobserved.); Generic direct workspace callers, initial-state conversion/adoption, artifact submission and attachment ingress: not-determinable (Code performs structural admission, but the record does not establish whether a human or software decides the submitted content. Generic caller identity cannot support manual or automatic content control for these alternative mechanisms.); Remaining package-spill writers, custom backends and maintenance automation: uninspected (Not all admission mechanisms are traced; inspected automatic summaries do not classify these alternatives.)	partial coverage (Both push and pull are supported. Their inventory is incomplete for custom and advertised-but-unwired alternatives; requested return transport is not counted as another push.); Later inference context and compatible provider/file replay after automatic context selection: known — push [wired] (A user message triggers inference but does not request retrieval of each retained context part.); Model-requested current workspace reads and legacy spill reads/grep/SQL; explicit operator event inspection: known — pull [wired] (Substring/identifier selectors on requested retrieval do not by themselves establish push signals.); Runtime checkpoint/seed and recorded code-result recovery: known — push [wired] (The automatic resumption selector is distinct from foreground workspace queries; hydration is not a second independent consumer request.); Prescribed optimization feedback-window consumption: known — push [claimed] (Upstream optimizing-agent choices may select the window; teacher receipt remains a prescribed route, not observed delivery.); Custom model/storage callers, advertised logs.events and untraced host/maintenance consumers: uninspected (Known CLI inspection does not establish advertised model-facing full-log reload. Alternative automatic versus requested consumers are not completely recorded.)	partial coverage (Known push selectors use coarse windows and actual identity matches. Pull grep does not establish inferred-lexical push. Unresolved consumers prevent a complete selector inventory.); Automatic context/summary selection at later inference and summarizer input preparation: known — coarse [wired] (This coarse signal does not establish judgment-based or lexical retrieval. Model generation of summary text is not the selector of retained context.); Continuation substitution and attachment resolution on a selected inference request: known — identifier [wired] (Actual matching and object resolution affect delivered parts. Merely having a filename or ref in a catalog would not establish this signal.); Runtime recovery and resumed code-result selection: known — coarse [wired], identifier [wired] (Position-based windows and identity-based state/result matching are distinct selectors within automatic recovery.); Prescribed teacher feedback window: known — coarse [claimed] (The record establishes a prescribed recent-window supply, not embedding or judgment-based trace retrieval.); Requested workspace/event retrieval: inapplicable (Exact key/ref, substring grep and event filters operate on pull requests. Signal values here classify push selection, not requested queries.); Custom consumers and advertised but unestablished model log reload: uninspected (The automatic trigger, selector and delivered retained parts are not established; this alternative cannot be made inapplicable by a known pull route.)	partial coverage (Wired summary and correction routes and claimed optimizer revision support yes locally. No bounded absence records support no for arbitrary alternatives. Raw event retention and unchanged recovery snapshots alone are not qualifying learning writes.); Automatic current and legacy conversation-summary writes and later inference consumers: known — yes [wired] (This qualifies independently of whether claims are novel, fidelity is verified or future capacity improves.); Current durable output-rejection state to later corrective inference guidance: known — yes [wired] (This is bounded within the correction route, not an automatic rewrite of the standing actor prompt. No exercised correction or benefit is supplied.); Prescribed NPO/GEPA trace-to-prompt/harness revisions: known — yes [claimed] (The qualifying mechanism is requested in instructions, not an observed or wired optimizer service.); Arbitrary workspace replacement, initial-state conversion, attachment ingress, opaque continuation contents and general pushed actor source: not-determinable (The actual trace input and durable behavior-shaping derivation are caller-specific and not supplied. These writes cannot be classified no merely because the generic API does not demand traces.); Remaining package-spill producers and maintenance automation: uninspected (Raw spill retention alone does not establish trace learning; untraced transformations or maintenance writers prevent a complete negative inventory.)	partial coverage (Knowledge, update-learning, corrective/standing instruction, routing, replay validation/enforcement and prescribed score ranking have separate consumers and evidence layers. Opaque arbitrary parts prevent complete coverage.); Conversation/summary/file and compatible continuation content delivered to inference; workspace/spill pages and operator event rows: known — knowledge [wired] (Summary text is user-role context, not automatically a standing system instruction. Opaque payload semantics are considered separately.); Compaction boundaries, recovery positions/seed state, and replayed method results at runtime consumers: known — routing [wired], validation [wired], enforcement [wired] (Checkpoint data is not itself a schema criterion merely because a decoder validates it; recorded replay arguments really are comparison criteria.); Retained permissions/resolved decisions and current output-rejection state: known — enforcement [wired], routing [wired], instruction [wired], learning [wired] (Only the composed tool gate and conditional output-contract route are covered, not all runtime effects.); Retained conversation and prior summaries at automatic summarization consumers: known — learning [wired] (Learning authority concerns the trace-fed artifact update, not improved capacity or factual acceptance.); Prescribed optimization feedback, candidate scores, retained prompt versions and promoted executable actor source: known — learning [claimed], ranking [claimed], validation [claimed], instruction [claimed], routing [wired] (Executable artifact consumption is wired; the optimizer-to-promoted-prompt path and its criteria remain prescribed, not observed.); Semantic force of opaque continuation/file contents and arbitrary adopted/workspace or pushed artifact content: not-determinable (Known delivery and execution channels do not resolve every internal payload role, standing instruction, or arbitrary code effect. No universal authority inventory follows.)	AAS-2026-10-15-tardigrade-455e1aaccd4d-01
Tendril	2	Capability selection text, symbolic index metadata and TypeScript source accumulated or replaced through registration, human editing or afforded generated-program writes; their listing, execution and human-inspection routes. Includes unresolved preexisting-content provenance, generated workspace-memory alternatives and SDK-owned retained summaries/checkpoints if any. Excludes shipped static instructions and their display copy, runtime configuration/credentials, ordinary current-run conversation/display state, transient stdout and temporary scripts, provider weights, and developer scaffolding. Arbitrary application files have no established shipped later-consumer route; their possible memory mechanisms remain unresolved rather than classified from the workspace grant alone.	partial coverage (File-backed capability memory is established; opaque alternatives prevent complete inventory coverage.); Capability selection text, index metadata and executable source on disk: known — files [wired] (All three operative registry parts have explicit filesystem storage.); SDK retained context/checkpoint alternatives and possible generated workspace-memory payloads: uninspected (Dependency internals and actual generated payloads/readers are unavailable in the accepted account. A workspace write grant does not enumerate a memory substrate or establish a later consumer.)	partial coverage (Automatic wired registration, automatic afforded program mutation and manual wired editing are distinct mechanisms; unknown initialization remains included.); Model registration admission: known — automatic [wired] (Node I/O and model authorship are distinct from the automatic admission policy.); Generated program writes under workspace grant: known — automatic [afforded] (Afforded permission capacity remains separate from wired registration.); Human editor replacement: known — manual [wired] (The confirmation applies to the shipped editor action, not arbitrary backend callers.); Admission of preexisting content and unspecified direct backend writers: not-determinable (Workspace initialization preserves old files; their control is not recorded. Generic backend identity does not establish a human content decision.); SDK retained context/checkpoint admission: uninspected (Dependency-owned writes and any checkpoint replacement control are uninspected.)	partial coverage (Known request-driven routes do not establish a completely pull-only system boundary while opaque alternatives remain.); Model-requested index listing and named source execution: known — pull [wired] (Listing is exhaustive; execution is name-addressed. Neither route independently pushes selected memory.); Human capability browsing and file inspection: known — pull [wired] (Human inspection is a later-consumer path, not model context injection.); SDK retained context delivery and possible generated application-memory consumers: uninspected (No accepted record determines whether these alternatives retain material and later pull or push it. A hypothetical generated reader is not a wired consumer.)	uninspected coverage (There is no supported positive push signal, but unresolved alternatives prevent axis-wide inapplicability or absence.); Inspected capability listing, source execution and human inspection routes: inapplicable (These routes fulfill requests, so no push selector exists to classify here. The requested name is an address, not identifier-based push.); SDK retained-context and generated workspace-memory push alternatives: uninspected (Trigger, selector input, selected retained part and even push existence are uninspected. No coarse or inferred signal can be assigned from a tool catalog alone.)	partial coverage (A claimed positive coexists with unresolved alternatives. No observed or causally supported learning, improved capacity or system-wide negative follows.); Requested failure-feedback to model repair to retained replacement to later Deno execution: known — yes [claimed] (The repair policy names the trace-fed update; wired persistence alone does not upgrade its claimed compliance.); General registration inputs and generated-program workspace updates beyond the repair policy: not-determinable (No actual inputs or program instances determine which further writes are trace-fed durable behavior-shaping updates. Lack of a supplied episode does not warrant a local no.); SDK retained summaries/checkpoints and their later consumers: uninspected (A qualifying continuation summary cannot be ruled out by inspecting only Tendril's constructor; dependency production and read-back are uninspected.)	partial coverage (Forces stay local to consumers and evidence layers. Index bookkeeping displayed to humans does not become a runtime validation or enforcement policy.); Capability description consumed by foreground model for tool fit: known — knowledge [wired] (The classified force is advisory reference at the model consumer, not factual acceptance.); Retained triggers/suppression interpreted as invocation policy by foreground model: known — instruction [wired], routing [wired] (Instruction and operation selection coexist at this consumer; no recorded relative-priority ranking is inferred merely from listing.); Retained TypeScript consumed as a program by Deno: known — instruction [wired] (Deno consumes the current source, not description text; actual program-specific checks and destinations remain outside this unit.); Failure feedback at the requested corrective-update consumer: known — learning [claimed] (Coverage is of the declared repair policy; compliance, actual feedback use and improved capacity remain unobserved.); Actual program-specific authority, generated application memory and SDK retained context consumers: uninspected (Arbitrary code could supply additional checks or memory consumers; opaque SDK alternatives are not classified from the inspected display.)	AAS-2026-10-15-tendril-f560fb299a37-01
Thalo	2	Retained/edited knowledge instances, entity schemas, synthesis source queries and prompts, generated Markdown and actualization checkpoints; reconstructed workspace/index access structures; shipped local maintenance, CLI/API request routes and action/Pi publication/read-back. Browser snapshots remain included with unresolved later-memory status; external provider parameters and retained state remain opaque included dependencies. Excludes static shipped system/skill instructions, transient command JSON/diagnostics, ordinary current-run presentation state, development workflows and unfrozen external implementations. External caller and provider alternatives remain unresolved rather than silently excluded.	partial coverage (File and Git retention and reconstructed access structures are supported. Browser later-consumption and opaque provider state prevent complete memory-store coverage.); Durable entry, schema, synthesis-query, retained-prompt, generated-prose and checkpoint text: known — files [wired] (These operative parts reside in .thalo files or Markdown files/fences; serialization is not a separate substrate.); Git-retained versions of synthesis files and checkpoint changes: known — repo [wired] (The shipped action commits and publishes file changes. Repository history supplements filesystem storage; actual publication is unobserved.); Reconstructed workspace documents, schema registry and access indexes: known — in-memory [wired] (Process Maps and source copies implement access and checking; files, not these transient maps, supply cross-job persistence.); Browser editable snapshots as possible later-consumed memory: not-determinable (Browser process state is inspected, but cross-page retention and a later consumer invocation of accumulated snapshots are not established. Ordinary current-session state is not automatically memory.); Opaque provider parameter dependency at synthesis: partial — model-weights [wired] (The selected model has external parameters; provider retention and mutable endpoint state are inaccessible. This does not classify a provider memory service.)	partial coverage (Automatic shipped admission is wired; manual decisions are afforded locally. Generic external callers and opaque provider writes prevent a complete agency classification.); Pi admission of generated Markdown and appended checkpoint: known — automatic [wired] (Software installs nonempty model output and a marker without an individual human content decision; starting CI/manual dispatch does not make this manual.); Action branch commit/publication after configured command completion: known — automatic [wired] (Software stages and publishes all working-tree changes after subprocess return; PR review is downstream, not pre-publication approval.); Initialization, formatting and merge-driver admission: known — automatic [wired] (Software computes and writes starter/formatted/merged text, including conflict-bearing merge output. Caller initiation is not per-content human approval.); Explicit human text supply, edit, revert or later PR-integration decision: partial — manual [afforded] (Human content decisions are an afforded separate mechanism; actual authors and exercised PR approval are not supplied.); Generic external editor/agent saves and arbitrary replacement command admission: not-determinable (Generic caller identity does not establish a human decision over individual content. Replacement command internals are external; their agency cannot inherit the shipped Pi classification.); Browser snapshot changes as later-consumed memory writes: not-determinable (setContent/reset are state updates, but durable/later-memory admission and per-content control remain unresolved.); Provider-retained state and parameter updates: uninspected (Provider admission/update mechanisms are outside the inspected source.)	partial coverage (Pull at explicit query/preparation and push at host-to-model context assembly coexist in one chain. Unresolved alternatives prevent a complete direction inventory.); Standalone query and requested actualization preparation: known — pull [wired] (Previously retained files reach caller-requested CLI/API query or actualization results; automatic code fulfillment is still pull.); Host supply of selected raw entries, retained prompt and whole prior synthesis file to a fresh Pi model: known — push [wired] (The host assembles context before the model call without a model request for the retained parts. This push is distinct from the upstream requested actualize result.); Checkpoint and schema consumption during requested selection/checking: known — pull [wired] (Callers request actualization or checking; code reads retained controls to fulfill those operations. Delivery is not counted as a second unsolicited route.); Additional external API/editor consumers and replacement commands: not-determinable (Access affordances exist but no additional later model reader is identified; direction cannot inherit the shipped Pi route.); Browser snapshot across consumer invocations: not-determinable (Current-session deterministic runner access does not establish a later accumulated-memory route.); Provider-retained state read-back: uninspected (Provider/Pi internal consumption is uninspected.)	partial coverage (Coarse full-file/prompt supply and identifier/freshness source selection are local host-push warrants. Requested returns are not extra pushes; opaque alternatives remain unresolved.); Pending Pi invocation: whole current target file automatically supplied to model: known — coarse [wired] (Trigger is pending execution; upstream file choice fixes the target and host includes all trimmed available content. There is no targeted selection within prose: coarse whole-file supply, not identifier matching merely because a path is present.); Pending Pi invocation: source entries selected upstream by stored entity/tag/link query and cutoff: known — identifier [wired], coarse [wired] (Trigger is pending synthesis. Exact entity/tag/link matches select current retained instances for host delivery; cutoff/change tests further filter that set. Identity predicates and coarse freshness eligibility remain separate findings, with no inferred relevance score.); Pending Pi invocation: retained prompt extracted from definition: known — coarse [wired] (Host includes collected prompt-header content as guidance. Header availability is coarse structural selection, not inferred lexical relevance to a consumer question.); Explicit requested query/preparation and requested schema/checkpoint reads: inapplicable (This unit is pull fulfillment; push-signal classification does not apply to the requested return.); Additional external and provider push selection: not-determinable (External consumers and provider internals have no established selector/payload inventory; pull-only inapplicability cannot cover these alternatives.); Browser snapshot later-memory selection: not-determinable (A current runner snapshot is inspected, but an accumulated-memory push route and its selector are not established.)	not-determinable coverage (No accepted record establishes a controlled yes or a bounded no. Durable recurrent generation is not itself evidence of trace-fed learning; unknown provenance and opaque updates prevent a system negative.); Pi automatic KB-to-Markdown-to-later-model route: not-determinable (Automatic durable rewrite and later context consumption are wired, but original inputs are caller-dependent KB entries, not authenticated session/tool/event/trajectory traces. This prevents yes; no bounded absence warrant supports no.); Checkpoint append, workspace/schema compilation and local maintenance: not-determinable (Known automatic transformations are symbolic control/content maintenance. Records neither identify a qualifying trace-fed behavior-shaping chain nor warrant complete absence across caller inputs; operational checkpoint metadata alone is not a trace source.); External authoring, commands and API consumers: not-determinable (Arbitrary caller provenance and transformations remain unresolved; an external agent label does not establish automatic trace learning.); Browser accumulated-state update and later consumer: not-determinable (Cross-invocation retained result and qualifying trace-fed update are not established.); Provider parameter/state updates: uninspected (Provider update behavior and internal trace use are inaccessible.)	partial coverage (Knowledge, instruction, routing, validation and enforcement have distinct local consumers. No record establishes retained relevance-ranking inputs or trace-fed learning authority; unresolved parts are not absent.); Raw source sections and prior generated Markdown at the later Pi request: known — knowledge [wired] (These parts supply reference/context to generation. Delivery establishes knowledge force, not factual endorsement or observed reliance.); Retained synthesis prompt at the Pi model consumer: known — instruction [wired] (Stored prompt supplies repeatable directions; natural-language guidance is not an enforcement gate.); Source queries and entry metadata at actualization/query membership consumers: known — routing [wired], enforcement [wired] (Predicates select the input scope and exclude nonmatching entries from delivery, giving routing and enforced membership rather than relevance ranking.); Checkpoint at future actualize/tracker consumer: known — routing [wired], enforcement [wired] (Retained cutoff/base state selects the next source window and excludes entries outside its predicate; it does not validate faithful incorporation.); Retained resolved schema directives at the checker and CLI exit consumer: known — validation [wired], enforcement [wired] (Constraints determine structural conformance. CLI errors enforce failed check completion, but cannot veto independent file writers.); Schema descriptions and other unconsumed metadata at further consumers: not-determinable (Readable documentation and metadata without an identified consequential consumer establish no additional force.); Browser snapshot at a later memory consumer: not-determinable (Current-session checks/preparation are inspected; later invocation and durable behavioral force remain unestablished.); Provider-retained state beyond the identified parameter dependency: uninspected (Opaque host/provider processing prevents classification of additional retained controls or update consumers.)	AAS-2026-10-14-thalo-86fcc3efe983-01
TheKnowledge	2	Included: accumulated raw/wiki content and metadata; derived lexical/dense access stores and cache; corrected examples, policy candidates/live updates and generated domain skills; demand/evaluation/conflict histories, intents and recovery provenance; research plans, notebook registry/source maps, opaque persistent remote corpus, downloaded artifacts and wrappers; requested development-session checkpoint and opaque host alternatives; remaining included but incompletely inspected acquisition, legacy, maintenance, editing, domain lifecycle, synchronization and export routes. Excluded: static shipped instructions/schema/code and inference model weights not changed through use, ordinary transient current-call frames, external provider internals as implementation evidence, deployment-specific state and unshipped consumers. Remote retained state and host alternatives stay in the inventory despite opacity.	partial coverage (Local stores and remote service objects are supported; opaque provider storage and incompletely traced included producers prevent a complete whole-system substrate inventory. Static model weights and shipped instructions are excluded, not classified as accumulated memory.); Local raw and wiki bodies, examples, demand, provenance, evaluation history, policies, generated skills, artifact downloads and operational intents: known — files [wired] (The records identify retained markdown, YAML, JSON, JSONL, CSV and binary files. Encoding does not change their filesystem substrate.); Requested development checkpoint and its requested commit: known — files [claimed], repo [claimed] (Instructions request a markdown checkpoint and a Git commit; this is a declared retention path, not a completed write.); Committed deposit history and accepted queued policy provenance: known — repo [wired] (Gate Git commits retain explicit paths and intent/policy provenance for later redelivery and audit decisions.); Derived lexical and vector index stores: known — sqlite [wired], vector [wired] (FTS and namespace/key/vector rows persist in SQLite. Vector names classify the operative vector store as well as its database substrate.); Reloadable dense index process cache: known — in-memory [wired] (Retrieval additionally caches derived index state by database mtime; this is retained selector state reused by later retrieval invocations.); Persistent remote notebook corpus: partial — service-object [wired] (Notebook IDs and later query establish provider-held retained service state. Its internal storage and additional substrates remain opaque.); Other included acquisition, backfill, promotion, editing, watcher and export state: uninspected (Converter sidecars, legacy bank production, additional authoring and upload/sync, and all vector cleanup paths are not completely traced; no complete substrate inventory follows.)	partial coverage (Automatic admission is supported on inspected routes. No authenticated per-content human decision establishes manual admission; unknown caller control, legacy initialization and opaque host/provider paths remain unresolved rather than negative.); Generated page admission, direct digest admission, query filing, skill emission and bootstrap: known — automatic [wired] (Code/model admission lacks compulsory per-content human approval. Starting these workflows does not make subsequent admission manual.); Worker admission of deposited content and deterministic recovery/demand output: known — automatic [wired] (Worker software independently renders/checks/adopts retained output; caller payload authorship is separate from this admission.); Derived index refresh, evaluation history and bibliographic metadata updates: known — automatic [wired] (Inspected software transformations and persistence decide admission without a per-result human approver.); Correction pinning into durable examples: known — automatic [wired] (Software validates and pins the feedback example; user-correction naming does not authenticate the proposer as human.); Correction labels, caller deposits/local edits, force overrides and queued policy replacement proposals: not-determinable (Generic caller identity and server principal do not establish explicit human content control. No manual value is inferred from CLI access or initiation.); Requested checkpoint writing and replacement: partial — automatic [claimed] (Hooks request agent selection, writing and committing without compulsory human approval; actual checkpoint and host alternatives remain opaque.); Legacy initialization, candidate promotion, additional domain lifecycle and upload/sync mechanisms: uninspected (Human versus software admission is unresolved on the remaining included paths; declared human review is not evidence of a completed manual promotion.)	partial coverage (Pull returns and internal automatic context push coexist. Static hook instructions themselves are excluded accumulated memory; their requested checkpoint delivery and generated skill affordance remain separate from opaque host/provider alternatives.); Explicit retrieve/answer, neighborhood context and persistent notebook query returns: known — pull [wired] (These supply retained material in fulfillment of a consumer request. Ranking or graph walking within a requested return is not another unsolicited push.); Explicit evaluate/status, preflight/demand and restore requests: known — pull [wired] (Requested operations reload retained evaluation trends, gaps and saved page content for their named consumer.); Existing-page context supplied to the plan model during ingest: known — push [wired] (Ingest code automatically selects retained pages on title/domain context and delivers them without the plan model requesting particular retained pages.); Policy and example context supplied to later filter model calls: known — push [wired] (Filter construction automatically delivers current criteria and selected retained decision examples on each invoked filtering path.); Retained context selected for evaluator and candidate distillation model calls: known — push [wired] (Orchestration automatically supplies selected retained wiki/source context and sampled examples/reference labels to these model consumers.); Requested checkpoint reanchor and generated skill discovery: partial — push [claimed], push [afforded] (Checkpoint reread is requested by host-event instruction rather than mechanically performed; generated skills afford later host-triggered delivery. Exact host selectors and actual delivery remain unresolved.); Remote selected corpus, downloaded payload rereads and remaining included callers: not-determinable (Remote selection/delivery, automatic model payload reload and untraced additional readers cannot be classified from local wrappers or known pull routes.)	partial coverage (Coarse, identifier and lexical push selectors have scoped support. Dense retrieval embeddings are a pull selector here, so inferred-embedding is not added as a push signal. Opaque host/provider and uninspected alternatives prevent complete assessment.); Ingest page push to plan model: known — inferred-lexical [wired], coarse [wired] (Trigger: authored ingest. Input: retained source title and chosen domain. Selected part: existing wiki pages. Ranked title-query selection and unranked glob fallback are distinct selectors.); Example and policy push to filter model: known — identifier [wired], coarse [wired] (Trigger: filter invocation. Input: effective domain and policy strategy/count. Selected parts: policy and example labels/scores/rationales/IDs, not stored excerpt bodies.); Example push to distillation model: known — identifier [wired], coarse [wired] (Trigger: distillation. Input: domain bank and decision category. Selected parts: sampled titles and capped rationales.); Wiki/source context push to judge and calibration classifier: known — identifier [wired], coarse [wired] (Trigger: evaluation/calibration. Inputs: domain and retained reference entries/source IDs. Selected parts: domain wiki bodies, enumerated raw source bodies or capped source text.); Checkpoint requested on host SessionStart: partial — coarse [claimed] (Trigger: SessionStart. Input: checkpoint existence/mtime. Selected part: requested whole checkpoint. This is coarse availability/time gating, not targeted identity selection within the file.); Generated skill host loading and opaque remote corpus selection: not-determinable (Host discovery and provider selection inputs are not disclosed sufficiently to classify their automatic selectors. Declared skill triggers alone are not wired identity matches.); Requested retrieve/neighborhood return operations: inapplicable (These local operations fulfill explicit retained-material requests; push selection signals do not classify their lexical, dense or identifier pull selectors.); Additional included context/maintenance selectors: uninspected (Unknown consumers/selectors on remaining included routes cannot be made inapplicable by the inspected requested-return paths.)	partial coverage (Correction reuse establishes a positive wired route; candidate guidance and session continuation have weaker local warrants. There are no supplied bounded absence records supporting a system-wide or local no; unresolved alternatives remain explicit. This axis is independent of demonstrated learning or self-improvement.); Correction decision feedback to pinned example to future filter guidance: known — yes [wired] (Code automatically admits durable decision/rationale examples; selected examples later shape filter model guidance. Human identity is unresolved, but automatic example admission and later consumer are wired. No improved capacity is claimed.); Example-fed durable policy candidates and intended review: partial — yes [afforded] (Automatic trace-fed candidate writing is wired; retained candidate is afforded for later review, but live installation and subsequent deployed use are not wired by this route.); Session working memory to durable checkpoint to resumed agent: partial — yes [claimed] (A generated continuation summary qualifies even without new claims. The source establishes requested automatic agent production and later reread, not completed production or exact host compaction inputs.); Demand history, evaluation results, recovery histories and conflict diagnostic snapshots: not-determinable (These have automatic history writes and named later consumers, but records do not establish a further durable behavior-shaping artifact on every route. Empty demand body, advisory score display and restoration are not by themselves universal trace-learning witnesses.); Static source-based digest/skill routes, mixed acquired source content and remaining legacy/host/provider update alternatives: not-determinable (Digest/skill records do not demonstrate trace-fed inputs; static acquired content alone is not a trace. Original mixed acquisition/legacy provenance and opaque alternatives prevent bounded route-wide no findings.)	partial coverage (Authority is consumer-specific: interpretive knowledge/instruction, priority inputs, update inputs, routing and admission/veto controls coexist. Provider payloads and additional included consumer paths prevent complete coverage.); Raw/wiki prose delivered to plan models, answering models and requested neighborhood readers: known — knowledge [wired] (Selected retained content provides evidence/reference/context, not mandatory execution policy.); Retained section text, title/link/trust/draft metadata and dense vectors at retrieval rankers: known — ranking [wired] (Retained text/features and access metadata influence relative priority even though ranking code is fixed.); Live editorial criteria and selected corrected examples at the filter model: known — instruction [wired] (Criteria and calibration examples direct future scoring; software thresholds separately decide source admission.); Example feedback at policy-candidate distillation: known — learning [wired] (Retained decision explanations feed an automatic durable guidance-artifact update. This learning force does not demonstrate improved capacity or live candidate installation.); Generated domain skill at a later host agent: known — instruction [afforded], knowledge [afforded] (Persisted skill supplies usage directions and selected domain reference material if the host discovers and loads it.); Requested checkpoint at resumed foreground agent: known — instruction [claimed], knowledge [claimed] (Instructions afford later checkpoint reread and interpretation, with tree verification and escalation requested rather than enforced.); Retained policy thresholds and queue retry/fencing/identity state at admission and worker consumers: known — validation [wired], enforcement [wired], routing [wired] (Thresholds and identity/version/lease/history checks provide criteria and can veto admission or redelivery; typed payload state chooses worker operation.); Query plans, notebook registry and source maps at research resumption/query rendering: known — routing [wired] (Retained plans and IDs choose resumed search/corpus access; source maps furnish source references for local rendering.); Gap history and trigger IDs at preflight/clustering: known — knowledge [wired], routing [wired], validation [wired], enforcement [wired] (Raw gaps supply planning diagnostics; recurrence selects queued work; retained trigger identities suppress duplicate submission.); Recovery provenance at reachability checks and restoration: known — validation [wired], enforcement [wired], routing [wired], knowledge [wired] (Saved paths determine cleanup eligibility; stored bodies supply reconstruction content; act identity locates withdrawal targets.); Evaluation history at status and retained goldens at evaluator gates: partial — knowledge [wired], validation [wired], enforcement [wired] (Status supplies advisory history; goldens supply validation criteria and policy gates can veto. Initial golden provenance and full result consumers remain uninspected.); Opaque remote payload consumers and incompletely inspected alternative maintenance consumers: not-determinable (Known wrapper and rank effects cannot identify every remote consumed part or authority on alternative included routes.)	AAS-2026-10-14-theknowledge-319a64927ee3-01
Tolaria	2	Included: accumulated vault Markdown and mutable root/legacy guidance, workspace transcript files/cache/frontend state, retained workflow flags and vault registry entries, with their admission, maintenance and later consumers. Included but unresolved: desktop index/cache and watcher/Git reconciliation, editor admission, web/server collaborative checkpoints, remote persistence and remaining maintenance/evaluator paths. Excluded: shipped static prompts/docs as such, transient request/read/snippet products, external CLI checkpoints and extension internals, and provider/model parameters.	partial coverage (Positive file, key-value and in-memory witnesses coexist with unresolved registry storage and included application alternatives. External model weights and external CLI checkpoints are outside Tolaria memory; transient tool results and shipped static prompts are excluded.); Vault Markdown bodies, frontmatter and mutable root/legacy guidance: known — files [wired] (Inspected retained files; transient parsed reads and request strings are not separate retained substrates.); Workspace transcript native file: known — files [wired] (JSON encoding does not imply an RDBMS substrate.); Workspace localStorage cache: known — kv [wired] (Browser key-addressed retained cache is distinct from the native JSON file.); Workspace frontend session map: known — in-memory [wired] (In-process retained state feeds later sends.); Organized workflow state and requested persistence: partial — in-memory [wired] (Frontend boolean state is inspected; lower persistence/reload dependencies remain uninspected.); Vault registry persisted entries: not-determinable (Registry persistence is recorded, but its substrate is not classified by the accepted record.); Included desktop index/cache, watcher/Git reconciliation, editor admission, and web/server collaborative and remote memory paths: uninspected (The accepted memory inventory leaves these paths untraced. Their storage facts and alternative operative parts are missing; inspected desktop witnesses cannot classify them.)	partial coverage (Automatic admissions coexist with the explicit manual workflow-flag decision. Human authorship or workflow initiation alone does not classify other admissions as manual.); MCP create/update/append admission: known — automatic [wired] (Local software admits caller-proposed changes without per-content human approval.); Direct-provider create-note admission: known — automatic [wired] (Starting the chat does not approve each returned note.); Managed guidance refresh admission: known — automatic [wired] (Content predicate preserves custom guidance but software selects replacement.); Transcript save/load normalization and prefix clone: known — automatic [wired] (UI initiation is separate from subsequent transcript admission.); Vault attach/clone registry and content admission: known — automatic [wired] (Caller proposes root or URL; local code admits registry/tree changes.); Organized flag supplied by operator toggle: known — manual [wired] (The user explicitly decides the retained boolean value; automatic persistence is physical I/O, not admission agency.); Native note save/delete and initial existing note/custom-guidance admission: not-determinable (Generic callers and untraced editor confirmation do not establish explicit human content control.); Included desktop index/cache, watcher/Git reconciliation, editor admission, and web/server collaborative and remote memory paths: uninspected (The accepted memory inventory leaves these paths untraced. Their write agency facts and alternative operative parts are missing; inspected desktop witnesses cannot classify them.)	partial coverage (Transcript push and requested note/MCP pull coexist. Static prompt delivery, immediate tool returns and registry scope use alone are not additional memory-content read-back routes.); Automatic transcript selection on later send: known — push [wired] (User requests a new answer, not the history payload; code independently supplies retained history.); Explicit note-reference request and hydration: known — pull [wired] (Requested note selection and its deterministic delivery are one pull chain, not an extra push.); MCP requested get/search/context, including recent metadata and AGENTS: known — pull [wired] (Requested context remains pull even when code sorts recent files.); Active-note snapshot and complete panel dispatch: uninspected (Builder inspection is described in the memory account but complete dispatch and automatic consumption are not established by a canonical route record.); Included desktop index/cache, watcher/Git reconciliation, editor admission, and web/server collaborative and remote memory paths: uninspected (The accepted memory inventory leaves these paths untraced. Their read-back facts and alternative operative parts are missing; inspected desktop witnesses cannot classify them.)	partial coverage (Identifier matching and coarse filtering are separate operations on the transcript push. Requested filenames, search terms and context recency do not become push signals. Untraced alternatives prevent complete signal coverage.); Transcript push session selection: known — identifier [wired] (Trigger: later send. Input: current session identity. Selected retained part: that session message array.); Transcript push eligibility and budget selection: known — coarse [wired] (Trigger: later send. Inputs: completed/nonmarker flags, recency and estimated character budget. Selected part: newest fitting user/response suffix.); Explicit references and MCP requested note material: inapplicable (These completely assessed requested operations are pull; push-selection signal does not classify their path/query/mtime selectors.); Included desktop index/cache, watcher/Git reconciliation, editor admission, and web/server collaborative and remote memory paths: uninspected (The accepted memory inventory leaves these paths untraced. Their push-selection signal facts and alternative operative parts are missing; inspected desktop witnesses cannot classify them.)	not-determinable coverage (No controlled yes or no is established by accepted records. Raw logging/replay is insufficient for yes; untraced maintenance and collaborative checkpoint alternatives prevent a system-wide no.); Raw transcript retention/replay and context trimming: not-determinable (The inspected chain retains raw turns, normalizes flags and evicts prompt text, not a recorded durable trace-derived summary. No bounded absence record supplies a controlled no finding.); Generic note writes and managed guidance refresh: not-determinable (Writes can carry arbitrary experience, but records establish no automatic trace-fed derivation into behavior-shaping artifacts. Template refresh is code-selected defaults, not trace learning.); Included desktop index/cache, watcher/Git reconciliation, editor admission, and web/server collaborative and remote memory paths: uninspected (The accepted memory inventory leaves these paths untraced. Their qualifying automatic trace-fed writes facts and alternative operative parts are missing; inspected desktop witnesses cannot classify them.)	partial coverage (Knowledge, instruction, routing, enforcement, validation and ranking are local consumer effects. No learning authority follows merely from storing traces; external model behavior and semantic approval are not inferred.); Retained note bodies and user/assistant history at later model context consumers: known — knowledge [wired] (Delivered evidence/context has interpretive force, not proven correctness or observed reliance.); Mutable root and legacy AGENTS guidance at MCP/agent instruction consumers: known — instruction [wired] (Instruction-bearing retained files are separate from the shipped static prompt.); Session IDs, status and streaming flags at history/recovery consumers: known — routing [wired], enforcement [wired] (These fixed-rule controls select a session and exclude ineligible material.); Note mtime at optional replacement check: known — validation [wired], enforcement [wired] (Sampled equality supplies an admission criterion and a veto, not a semantic verdict.); Note mtimes at recent-note orientation selector: known — ranking [wired] (Relative priority is fixed-code recency, not semantic quality.); Vault registry/root state at later tool-scope consumer: known — routing [wired], enforcement [wired] (Retained root entries widen later read/write scope.); Organized flag at UI inbox workflow consumer: partial — routing [wired] (State transition and next-note selection are inspected; complete persistent reload is not.); Other frontmatter/type/view and retained action/reasoning consumers: not-determinable (Complete type/view semantics and later consumption of stored reasoning/actions are not established.); Included desktop index/cache, watcher/Git reconciliation, editor admission, and web/server collaborative and remote memory paths: uninspected (The accepted memory inventory leaves these paths untraced. Their authority facts and alternative operative parts are missing; inspected desktop witnesses cannot classify them.)	AAS-2026-10-14-tolaria-42ddaf2ee8cd-01
Tracecraft	2	Includes accumulated project-bucket values/history, mail/cursor, claim/status, handoff, artifacts, mirrored transcript parts/meta, registration, local mirror checkpoint and optional durable evaluator trial JSONL. Includes opaque caller/file/host-checkpoint alternatives and unresolved handoff/evaluator later consumption. Excludes static shipped instructions/package source, credentials/configuration, external host storage/model loops and provider internals, and ephemeral evaluator fixtures/current-trial reads.	known coverage (The recorded retained-object inventory covers project bucket objects, the local ingestion checkpoint and optional durable evaluator output. Opaque file content does not make its bucket substrate opaque; provider implementation is excluded.); Project-bucket values/history, mail/cursor, claim/status, handoff, artifacts, transcript parts/meta and registration: known — service-object [wired] (All named operative parts are remote bucket objects on either supported backend.); Local mirror cursor checkpoint: known — files [wired] (The next mirror reads a local JSON file; host SQLite is an acquisition source, not Tracecraft retained storage.); Optional durable contention evaluator trial JSONL: known — files [wired] (The supplied record establishes file emission, independently of the uninspected later report reader.)	partial coverage (Automatic trace ingestion and software-derived access/result writes are established. Generic CLI caller identity cannot resolve per-content human control for caller submissions or replacements.); Caller submission/replacement of values, messages, claim/completion/handoff, files and registration: not-determinable (Human versus model/software content decision is unresolved. Software physical I/O and command initiation alone do not establish automatic or manual admission of the submitted content.); Per-flush transcript and access-state admission after mirror initiation: known — automatic [wired] (Mirror software selects deltas, optionally redacts and writes without per-item human approval.); Inbox cursor advancement after requested display: known — automatic [wired] (The software derives and commits reader position, without a separate operator decision on retained cursor content.); Optional Git-derived handoff path list: known — automatic [wired] (Enabling the flag does not approve each returned path; fixed code admits the measurement.); Compact-produced merged payload and replacement manifest: known — automatic [wired] (Requested maintenance triggers fixed concatenation and metadata replacement without per-result content approval.); Optional durable evaluator trial rows: known — automatic [wired] (Developer initiation selects conditions; software admits computed row fields.)	partial coverage (Established content/inspection routes fulfill caller requests. Local mirror code also reloads persisted access state during subsequent flushes. Handoff delivery and durable evaluator report reads are unresolved; they cannot be excluded to claim a complete pull-only boundary.); Requested get/history/inbox/status/wait/download/session-list/show/agents reads: known — pull [wired] (Automatic filtering inside the request is not independent push.); Mirror internal request for persisted cursor and prior metadata: known — pull [wired] (Subsequent mirror explicitly loads access state as its requested input, rather than unsolicited context delivery.); Handoff next-agent read-back: not-determinable (A storage GET capability and producer promise do not establish a later consumer route or its direction.); Durable evaluator trial JSONL later report read-back: uninspected (Emitter is established; downstream report-generation internals were not inspected. Same-trial fixture reads are not later consumption of durable output.)	not-determinable coverage (No positive push-selector value is established. Assessed requested routes are locally inapplicable, but unresolved handoff/evaluator routes prevent system-wide inapplicability.); Established requested dynamic reads, including mirror access-state loads: inapplicable (These are pull operations. Explicit identifiers, cursor filtering, sort and tail budgets on requests do not establish push signals.); Handoff delivery selector: not-determinable (next_agent is stored attribution, not an evidenced identity match selecting delivered content; trigger and consumer are missing.); Durable evaluator-output delivery selector: uninspected (Later report consumer/selection is uninspected, so push signal cannot be classified.)	not-determinable coverage (The accepted records establish raw/transformed trace retention and unresolved caller/host/evaluator routes, not a qualifying automatic trace-fed behavior-shaping update. They declare no bounded absence records, so a system-wide no is not authored.); Tracecraft mirror, scrub and compact to requested transcript inspection: not-determinable (Automatic writes are wired, but these records describe trace preservation/layout rather than a derived continuation/behavior-shaping artifact with a later guidance consumer. Raw-log storage alone is insufficient.); Caller coordination content and uploaded artifacts: not-determinable (Unknown caller decision/provenance and opaque file payload prevent determining whether any submitted part is a qualifying automatic trace-derived artifact.); Already-derived external-host checkpoint input: not-determinable (Checkpoint filename establishes ingestion only; original automatic derivation and later host guidance consumption are unresolved.); Access checkpoints, metadata, registration and claim/status: not-determinable (Fixed bookkeeping/routing is established, but no record establishes the qualifying trace-to-behavior-shaping-artifact learning chain. This is not evidence of improved capacity.); Durable evaluator results and later report/adoption: uninspected (Automatic trial recording is established; later report/adoption is uninspected and no recorded automatic policy update closes the chain.)	partial coverage (Fixed coordination effects, inspection displays and session-metadata ranking are wired; external-agent reference use is afforded. Opaque files, handoff delivery and later evaluator report consumption remain unresolved. Static shipped instructions are excluded.); Returned shared values/history and inbox text at external human/agent readers: partial — knowledge [afforded] (Shipped instructions afford later reference/context use, but unrestricted value/message content and actual uptake prevent complete downstream authority classification.); Claim owner at completion admission: known — validation [wired], enforcement [wired] (The same retained owner criterion both checks admissibility and vetoes a mismatch, with the recorded force bypass.); Claim/status at dependency waiter: known — routing [wired], enforcement [wired] (Fixed state comparisons control return, wait or error, not task truth.); Inbox cursor at new-mail selection and unread counting: known — routing [wired], enforcement [wired] (Full inbox is the separate cursor-ignoring route.); Local mirror cursor at subsequent ingestion: known — routing [wired] (Recorded cursor alone determines the next source window; source identity is not revalidated.); Session metadata and registration as inspection information at CLI views: known — knowledge [wired] (Display consumers supply descriptions, totals and age, not authenticated liveness. Session-list ordering and limit selection are classified separately below.); Retained session metadata at session-list ordering and limit selection: known — ranking [wired] (The recorded fixed-code sorting consumer establishes ranking independently of metadata display. This unit does not include registration display or claim observed agent uptake.); Handoff notes, next action, recipient and optional Git paths: not-determinable (Content-reader wiring is unresolved. Stored directive/recipient text alone establishes neither instruction nor routing authority at a later consumer.); Opaque artifact bytes at the downstream parser/agent: not-determinable (Download is wired and an agent consumer is afforded, but file content and its operative effect are not established.); Transcript tail at inspection reader: known — knowledge [wired] (The established consumer is requested inspection, not host restoration or automatic guidance.); Durable evaluator output at report/adoption consumer: uninspected (The later report reader was not inspected, preventing classification of force carried by durable trial rows. Immediate trial checks do not establish later authority.)	AAS-2026-10-14-tracecraft-9c4af961e8e2-01
Virtual Context	2	Canonical live/imported conversation content and provenance; segment summaries/full text, facts, structured envelopes, tag rollups, search vectors and actor cards; fact ledgers, checkpoint/working-set state, curation memo, proxy archives and retained engagement history; their admitted write, maintenance and later-consumer routes, including unresolved provider/backend, MCP, administrative repair and evaluator alternatives. Excludes static shipped core instructions, unmodified model weights, current-call-only checks/drafts and diagnostic artifacts without a traced later memory consumer.	partial coverage (Inventory covers canonical and derived content, access representations, archives, checkpoints, memo and engagement state. Configured store/provider alternatives and other maintenance/evaluator surfaces prevent complete storage coverage. Static core files, model service weights and diagnostic-only artifacts are excluded from accumulated memory.); Canonical text/provenance, fact content/links and verdict/refusal ledgers in the inspected relational storage family: known — rdbms [wired] (Relational rows and ledger tables establish this substrate locally; configured backend equivalence is not established.); Stored segment views, tag rollups, structured envelopes and actor cards through the configured store: not-determinable (Persistence is wired, but the records do not resolve a single controlled physical substrate for every configured store alternative.); Retained fact, chunk and tag-summary search representations: known — vector [wired] (The embedding storage/search backend holds dense retrieval state; backend and stale-index alternatives remain unresolved.); Proxy tool/chain rows and local media restore files in the inspected path: known — sqlite [wired], files [wired] (The inspected SQLite archive path and local media files are classified; this does not classify opaque media encoding.); In-process checkpoint state and optional Redis cache: known — in-memory [wired], kv [wired] (These two named alternatives are wired; durable store/provider-owned saving is a separate unresolved unit.); Durable checkpoint store, provider-owned checkpoint/curation memo and engagement history: not-determinable (Store/provider implementations and engagement history substrate are not fully classified; a known Redis witness cannot classify them.); Other MCP resources, migrations/backfills, evaluator caches and dated repair replacement surfaces: uninspected (Included alternatives remain incompletely inspected in the accepted accounts; their retained-object substrates cannot be enumerated.)	partial coverage (Known automatic transformations and the explicit human-admin path coexist. Generic caller identity, initial provenance control and other replacement/save mechanisms prevent complete admission-control classification.); Completed-turn tagging, summary/fact/vector writes and optional actor-card admission/replacement: known — automatic [wired] (Code/model admission has no per-derived-content human approval; enabling or manually starting compaction does not change this.); Normalized import admission, checkpoints/curation memo and proxy archive captures: known — automatic [wired] (Software performs content/state admission after caller input; provider-owned persistence internals remain separate.); Software fact comparisons/refusal decisions and engagement ledger claims: known — automatic [wired] (Posting intent is not approval of every retained draft; code owns claims and comparison disposition.); Explicit human-admin canonical edit and fact review path: known — manual [wired] (The records distinguish a human admin supplying exact edit/verdict from an arbitrary API actor. This finding is confined to that explicit human path.); Generic API edits/verdicts and initial host-supplied source/provenance admission control: not-determinable (An actor label or generic caller does not establish human decision versus software admission.); Provider-owned saving, additional administrative backfills/repair and evaluator alternatives: uninspected (The records leave provider save mechanisms and alternative replacement bodies incompletely inspected.)	partial coverage (Automatic supply to a later answer invocation and requested recall/restore/state reads coexist. Inventory retains unresolved consumer chains; host method calls and requested result delivery are not indiscriminately counted as pushes.); Automatic inbound selection into headless/proxy model context and requester-card delivery: known — push [wired] (Inbound user work triggers selection, but the answer model need not request each retained part. Requested API return is not a second automatic supply operation.); Model/host paging, recall and proxy archive restoration requests: known — pull [wired] (These are explicit requests for retained content; downstream formatting/return does not independently add another push.); Checkpoint restoration and fact-refusal/curation decision read-back by software consumers: known — pull [wired] (Code explicitly loads retained state/decisions for restoration or admission. Their effects are distinct from foreground automatic context supply.); Engagement history lookup at later repetition/day consumers: known — pull [wired] (History is requested internally during candidate/draft admission and posting; caller-owned scheduling is not an unsolicited memory notification.); Other MCP resources/searches, complete hint/envelope routes, provider alternatives and repair/evaluator reuse: uninspected (Known push/pull witnesses do not establish all direction units for incompletely traced alternative consumers.)	partial coverage (Signal findings describe automatic selection into model context only, with trigger/input/selected part retained locally. Inapplicable requested operations do not remove uncertainty about other automatic consumers.); Inbound summary/fact selection: query/context embedding and model selection: partial — inferred-embedding [wired], inferred-judgment [wired] (Trigger is inbound work; selector inputs are current query/recent context and eligible candidate facts. Exact retriever and curator alternatives remain incompletely inspected.); Inbound tag matching and availability/protection/budget selection: known — identifier [wired], coarse [wired] (Inbound work triggers tag/state filtering and measured-budget admission. Tag identity matches select history/topic parts; coarse gates separately bound inclusion.); Optional requester-card push: known — identifier [wired], coarse [wired] (Assembly trigger uses tenant/requester actor and audience to read that actor card; validity/confidence/token cap further select whole entries.); Explicit paging/restore/recall selection operations only: inapplicable (These units fulfill requests, so their supplied ref or requested topic is not an identifier-based push signal.); Untraced automatic selection alternatives and full hint/evaluator/MCP consumer chains: uninspected (Their selectors and selected retained parts are not completely established; known pull branches cannot make the entire signal axis inapplicable.)	partial coverage (Positive wired trace-fed durable memory and engagement routes coexist with unresolved alternatives. This is the comparison axis meaning, not measured improved capacity or exercised self-improvement; no unsupported system-wide no is asserted.); Completed-conversation compaction summaries/facts/tag derivatives to later answer context: known — yes [wired] (Automatic durable continuation memory qualifies even when lossy or source-text fallback; no new claim, parameter update or measured improvement is required.); Optional actor evidence to admitted card entries to requester context: known — yes [wired] (Actor-authored traces feed automatic durable card refresh and optional later guidance; the disabled default limits activation, not wiring.); Posting activity to retained question/day history to later engagement admission: known — yes [wired] (A claim may persist even after ambiguous send; later history constrains selection without demonstrating improvement.); Raw tool/chain/media capture and provider checkpoint alternatives as qualifying learned artifacts: not-determinable (Raw archive/reload alone is not a qualifying derived update; records do not bound all alternative transformations or opaque saving. No local or system-wide no is warranted.); Full rollup/curator alternatives, additional evaluators/admin repair and general diagnostics feedback: uninspected (Missing transformation and feedback-consumer facts prevent exhaustive qualifying-route coverage; diagnostic retention alone does not establish learning.)	partial coverage (Knowledge, instruction, learning, ranking, routing, validation and enforcement each have a local retained part and actual consumer. Unconsumed metadata and static policies are not assigned authority; unresolved consumers prevent complete coverage.); Canonical transcripts, summary/full text and facts at answer-model consumers: known — knowledge [wired] (Selected content supplies evidence/context, not a truth certificate or guaranteed model compliance.); Retained conversation experience at summary/fact/card update consumers: known — learning [wired] (The update path differs from downstream evidence and instructions; wired learning authority establishes neither weight training nor measured improvement.); Actor-card communication preferences and person context at optional answer consumer: known — instruction [wired], knowledge [wired] (The wrapper requests following response preferences and supplies relevance context; ships disabled and compliance is unobserved.); Source IDs/epochs, lifecycle state and refusal/verdict ledger fields at admission/disclosure consumers: known — validation [wired], enforcement [wired] (Criterion and veto are separate effects; structural proof does not entail statement truth.); Checkpoint watermarks, tag/working-set state and matching curation memo at later selectors: known — routing [wired] (Retained control fields select windows, eligible inputs and fact subsets rather than directing a model in prose.); Stored dense vectors at fact ranker and tag-summary text at embedding/scoring: known — ranking [wired] (Ranking is a software selection effect, not a semantic warrant; not all retriever strategies are inspected.); Engagement day claims and prior-question history at repetition/posting admission: known — validation [wired], enforcement [wired] (Earlier retained activity can veto later selection and day availability even after ambiguous sending.); Restored tool/chain text at provider continuation: known — knowledge [wired] (Sanitized selected raw context reaches later continuation, not unrestricted store content.); Opaque media effect, full hint/structured-envelope consumer routes and other MCP/repair/evaluator consumers: uninspected (Metadata or a readable display does not resolve opaque consumption; untraced consumers prevent an exhaustive authority union.)	AAS-2026-10-14-virtual-context-bbd06cc3ac95-01
VLM-wiki	2	Accumulated raw media/text and frames; analysis prose/status and physical metadata; import manifest; personal/project wiki articles, archives, index and logs; historical/work reviews and runtime briefs; helper-target report revisions and their admission, maintenance and established later consumers. Static shipped AGENTS/prompts/config/templates and external host/provider weights are not accumulated memory. Unknown initial authorship and missing output consumers remain included uncertainties; external host internals, private input/config state and unaudited media semantics are not silently classified.	known coverage (The accepted object inventory covers the shipped retained boundary. Git tracks some examples but is not their operative memory storage API; static configuration, prompts and external provider weights are excluded.); Raw copies, frames, analysis prose and JSON; manifest; personal/project articles, index, logs, review Markdown, runtime briefs and helper-target reports: known — files [wired] (All included accumulated objects are local files. Implemented file writers establish the substrate; instructed wiki writing and tracked examples use the same substrate, not a separate repository-backed storage engine.)	partial coverage (Known automatic mechanisms do not resolve initial control or hypothetical later human edits; no unsupported manual finding is added.); Importer admissions, decoder/metadata writes, direct response persistence, project/index/review/brief/log production and marker-controlled repair: known — automatic [wired] (Software admits individual outputs without a per-content human decision. Confirmation, configuration and workflow launch are not content approval.); Requested host compilation, related-page editing, archive and navigation/log updates: known — automatic [claimed] (The model chooses and admits content under instructions; an archive request initiates work rather than approving each assertion.); Original raw submissions and initialization of tracked reports/articles/index/history: not-determinable (The records do not identify per-content human supplying/editing/approval decisions. File editability alone cannot establish manual admission.)	known coverage (Read-back inventory distinguishes requested reads from unsolicited downstream delivery. Reports/briefs/logs without established readers add no inferred direction; their limitation is explicit in the cited object inventory.); Later importer manifest lookup; frame/image metadata requests; historical-review collection: known — pull [wired] (The invoked consumer requests retained files/fields needed for its operation. Internal delivery fulfilling that request is not another push.); Question-led host index/article reads and documented human browsing: known — pull [claimed] (The user question causes retrieval; model relevance choice within that requested retrieval remains pull.); Producer-initiated work-review preview delivery to configured WeCom recipients: known — push [wired] (After cycle invocation, code launches sender and supplies selected retained review content without a recipient query. Scheduler recurrence and successful receipt are not established.); Standalone historical-review sender documented in README: known — push [afforded] (A documented caller supplies the file to the same outbound sender; the daily preview marker mismatch changes selected content, not direction.)	known coverage (All recorded push branches use coarse content availability/section/budget rules. Host judgment belongs to requested pull selection and does not justify inferred-judgment push; manifest identity lookup is also on a pull route.); Work-cycle push: cycle completion triggers preview availability/heading and fixed line/character-budget selection of the produced review: known — coarse [wired] (The selector uses a static preview section and budgets, not question relevance, embedding similarity or matched recipient/content identities. The upstream review path is a requested file handoff, not identifier-targeted selection.); Documented historical-review push: caller invokes sender with daily file; fallback chooses first twenty nonempty cleaned lines: known — coarse [afforded] (The English preview heading fails the Chinese marker test; fallback and 4000-character cap determine recipient payload. Selector implementation is wired, but this additional consumer route is afforded.)	partial coverage (A positive qualifying work route coexists with unresolved origins/consumers. No bounded absence record licenses a global or local no, and unchanged rules do not negate trace learning through durable advice.); External project activity/Git metadata automatically converted to retained project pages and work advice, consumed by later history invocation or sender: known — yes [wired] (Automatic trace-fed durable behavior-shaping material has a later consumer. Fixed Python rules and human-facing advice do not disqualify the route; no parameter update, recursive review adaptation or improvement is asserted.); Historical article/diary transformation, host media/archive routes and initial sample content: not-determinable (Inputs mix personal text, derived work history and unknown origins. The records do not establish qualifying trace provenance and later-consumed guidance for every included alternative; brief retention alone supplies no later consumer.)	partial coverage (Actual retained parts and consumers determine force. Static AGENTS/config authority is excluded. Missing consumers are retained as uncertainty, not silently classified as instruction or enforcement.); Raw media/text at interpretation; project/article fields at historical report builders; work preview at human recipients: known — knowledge [wired] (Retained evidence/context shapes reports or supplies human advice; recommendations do not become standing host instructions merely by using imperative prose.); Personal/project articles at index-led host answering: known — knowledge [claimed] (The host is told to prefer article evidence; compliance and answer dependence remain unobserved.); Index links and summaries at host article selection: known — routing [claimed] (Retained navigation selects article inputs, without a demonstrated ranking algorithm or complete target consistency.); Manifest membership at later import admission: known — validation [wired], enforcement [wired] (A retained fingerprint is an exclusion criterion and executable skip veto; descriptive manifest fields have no equivalent force.); Retained article dates/tags and diary repetition at historical priority selection: known — ranking [wired] (Retained fields determine latest/top-place and repeated-line priority under fixed rules; no learned ranking policy is required.); Retained project activity at automatic durable report production for subsequent human/history consumers: known — learning [wired] (Activity records feed durable behavior-shaping advice and pages, with a later sender/history consumer. This update-consumer force is separate from downstream knowledge and proves no improved capacity.); Analysis reports, metadata/status outputs, briefs, logs and helper-target labels without an established later consumer: not-determinable (Retention or a trust warning alone cannot establish downstream authority; missing reader/activation facts prevent assigning a force to these alternatives.)	AAS-2026-10-14-vlm-wiki-311faef777b4-01
Voiden	2	Agent-facing MCP subsystem memory changed or accumulated through use: recorded response blocks; cross-call runtime/environment bindings and standalone durable snapshots; CLI execution history; publisher decisions/cache and unresolved provider status persistence; refreshed installed guide bodies, summaries and their access index. Includes opaque provider/script/initial-content alternatives and external guide-loading limits. Excludes unchanged shipped instructions/source code, executable plugin bytes and capability-installation metadata unrelated to these memory objects, outgoing transport session/token plumbing, ordinary current-run-only state, external host conversation/compaction and remote stores.	partial coverage (Files and process memory are established locally. The included provider write-back/storage branch prevents complete substrate coverage; no repository substrate is inferred from file extensions or a source checkout.); Recorded response blocks in project .void documents: known — files [wired] (The retained block substrate is directly established; project Git tracking is not established.); Runtime captures and selected environment across MCP calls; standalone saved variable snapshot: known — in-memory [wired], files [wired] (The two driver-specific retained locations are covered; current-run-only bindings are excluded.); CLI accumulated execution history: known — files [wired] (The history file has concrete later report consumers.); Publisher decisions and core verification-entry cache container: known — in-memory [wired] (This resolves core containers only, not provider-internal stores.); Provider status write-back and additional provider persistence: not-determinable (Project-file write-back is requested, but provider serialization/storage and any additional retention are unavailable. An interface is not a file-storage witness.); Use-refreshed installed guide collection and its paths/index: known — files [wired] (Includes the changed installed collection, not unchanged bundled source instructions.)	partial coverage (Automatic software admission is established for capture, driver outputs, guide refresh and status forwarding. Caller response control, initialization and provider-owned admission prevent a complete agency inventory; no manual value is inferred from workflow initiation.); MCP caller-proposed response admission: not-determinable (No per-result human confirmation is required, but generic caller identity does not establish whether a human explicitly supplied/approved this particular retained content. Serializer I/O is not the agency decision.); Runtime capture, CLI map save and history append: known — automatic [wired] (Authored capture rows or starting session mode do not make the subsequent values manual.); Environment-map replacement after profile choice: known — automatic [wired] (Global profile choice is not approval of each retained value; initial profile-file admission is unresolved separately.); Initial external variable/history edits, profile authorship and plugin script admission: not-determinable (Actual initial state and external content decisions are uncaptured; scripting receives a writable map but its admission policy is unavailable.); Guide refresh selection and generated installed collection: known — automatic [wired] (Operator setup/enablement is a trigger, not per-content manual admission.); Guide source-content admission: not-determinable (Source authors supply prose but actual human/model control of downloaded guide content is not established.); CLI provider-status forwarding: known — automatic [wired] (Known dispatch is classified independently of successful provider retention.); Publisher decision/cache admission and provider physical status writes: not-determinable (Provider creates/checks/adopts these retained parts; interface and caller identity do not resolve admission control.)	partial coverage (Both pull and push have supported local routes. Provider and host selection alternatives remain included and unresolved; a known request route cannot establish a globally pull-only boundary.); Requested file inspection and automatic later response-block callback context: known — pull [afforded], push [wired] (The document execution request and internal historical-context supply are different operations; immediate new execution output is not another memory push.); Later runtime-value substitution and durable snapshot reload: known — push [wired], pull [wired] (The automatic load and selector are distinct from requested variable display.); Requested history report/status/export: known — pull [wired] (Formatting and requested delivery are not independent pushes.); Core cached decision reuse and provider planning input: known — push [wired] (This establishes supply to provider calls, not that the provider reads every cached entry.); Provider targeted cache consumption, dynamic tool search and persistent status read-back: not-determinable (Actual selection/request/delivery semantics and status consumers are unavailable. The dynamic flag is not a resolved direction route.); Refreshed full-guide loading by foreground agent: known — pull [afforded] (This is a documented request route; actual guide loading is unobserved.); External host automatic inclusion of installed base/guide collection: not-determinable (Installation and file persistence do not establish host context assembly, so possible automatic inclusion remains unresolved.); Selected-environment carry-over and plugin script consumers: partial — push [wired] (Script-specific later consumers remain unavailable; the known fixed carry-over does not resolve them.)	partial coverage (Coarse supply and exact binding identity selection are established. Inapplicability is local to assessed requested reads; opaque push alternatives prevent an axis-wide inapplicability or complete selector inventory.); Automatic response-block supply to callbacks: known — coarse [wired] (A requested file path/placement UID is not an identity-matched targeted push selector.); Runtime binding substitution: known — identifier [wired] (This is an actual identity match selecting delivered values, not merely a catalog identifier.); Whole durable snapshot reload and selected-environment supply: known — coarse [wired] (Whole-map carry-over is separate from identifier targeting inside later substitution.); Core publisher planning/registration supply: known — coarse [wired] (Core supply is known; provider entry matching and dynamic-mode targeting are not.); History report requested return: inapplicable (This assessed route is pull-only: callers request all retained entries and their report format; no independent automatic selector supplies history.); Foreground full-guide requested read: inapplicable (The documented opening of a guide is a requested pull, so that operation has no push signal. This does not cover external host injection.); Provider dynamic targeting/cache selection, status read-back, host base loading and script consumers: not-determinable (Missing actual selectors, trigger inputs and selected payloads prevent additional signal values or inapplicability for these alternatives.)	partial coverage (A wired yes belongs to the durable capture/substitution route, not a system-level improvement claim. Raw reporting and deterministic guide packaging are locally inapplicable; unresolved write alternatives prevent complete trace-learning coverage. No unsupported no is asserted.); Automatic execution capture with standalone durable binding persistence and later substitution: known — yes [wired] (This wired trace-fed behavior-shaping chain needs neither novel knowledge nor an observed capacity improvement. The durable standalone route is the positive witness.); Recorded response admission and later plugin use: not-determinable (Caller provenance/admission control and unavailable plugin behavior prevent determining an automatic trace-fed durable behavior-shaping chain; serialization alone does not establish it.); Raw accumulated execution history and its report consumer: inapplicable (This inspected route retains raw outcomes for reporting, not an automatically produced durable behavior-shaping derivative. It cannot yield a qualifying learning route from history retention alone.); Refreshed guide collection and index summaries: inapplicable (The inspected composition route imports/statically reshapes guides rather than consuming activity traces. External guide-authoring processes are not resolved by this unit.); Initial bindings, environment acquisition, script writes and source-guide updates: not-determinable (Uncaptured initial provenance, resolver details and opaque script/source updates prevent qualifying or excluding further trace-fed writes.); Provider decision/cache/status transformation and later consumption: not-determinable (Provider verification might yield qualifying retained updates, but original inputs, actual write/persistence and later consumers are unavailable.)	partial coverage (Authority is assigned at each actual or afforded later consumer. Provider criteria and payload effects remain unresolved. Trace learning below does not imply learning authority for every retained output.); Response content supplied to later callback document context and afforded file readers: partial — knowledge [wired], knowledge [afforded] (Delivery as context is established; unavailable plugin consumers prevent stronger instruction, enforcement or factual acceptance claims.); Retained runtime bindings supplied to subsequent request substitution: known — knowledge [wired] (This is programmatic input authority, not factual endorsement or improved capacity.); Selected-environment state and unresolved script-written bindings: partial — routing [wired] (Resolved values feed later execution, but resolver internals and script meanings prevent enumerating all effects. Session IDs are access metadata, not a learned policy.); History consumed by later report/status/export commands: known — knowledge [wired] (No automatic future-request or model-prompt consumer is established for this history path.); Publisher retained served-state fields at core refresh comparison: known — routing [wired] (The core consumer is known; this does not establish the provider eligibility check.); Provider cache, substantive registration policy and persistent status read-back: not-determinable (Cache contents may inform checking/selection and status may inform future consumers, but actual provider use is unavailable. No validation/enforcement is inferred from verified labels.); Provider warning description at external agent tool-list consumer: partial — knowledge [afforded] (Formatting, admission to the description and actual host interpretation are not established.); Refreshed guide prose at external foreground agent: known — instruction [afforded] (Directions are advisory standing guidance, not code enforcement; obedience is unobserved.); Guide index and paths at on-demand guide loading: known — routing [afforded] (This is a documented foreground-agent loading route, not an automatic targeted push.)	AAS-2026-10-14-voiden-98188b4a2f55-01
Voyager	2	Active and archived retained skill code/descriptions, skill and QA access indexes, cached answers, chest maps, task histories, event archive/replay accounting, and encountered-block memory; includes checkpoint/library reload alternatives and their unresolved historical provenance. Working-frame replacement is assessed only where it carries accumulated memory to later invocations. Excludes shipped static prompts/primitives, external model weights as producer components, ordinary current-step payloads, and transient executor/plugin queues without an established retained-memory consumer.	known coverage (The recorded object inventory covers active/archived skills, both indexes, QA, chests, histories, event files/replay summaries and within-bot block memory. Opaque trial payloads have known file/access locations, not known contents. External models are producer components, not a Voyager memory store changed through use.); Active skill code and descriptions, including loaded checkpoints and versioned archives: known — files [wired], in-memory [wired] (Active entries have Python maps and JSON backing; code/description archives are separate files, not a repository write route.); Skill description documents, name metadata and dense access vectors: known — vector [wired], files [wired] (Chroma persistence is a vector access structure with local disk backing; binary trial files were not decoded.); QA answer map and question index: known — files [wired], in-memory [wired], vector [wired] (Answers have map/JSON storage; the separate question index has Chroma persistence. No rdbms/sqlite implementation is inferred from an unfrozen dependency.); Chest maps and completed/failed task histories: known — files [wired], in-memory [wired] (These checkpoint-backed objects have both operative in-memory and file parts.); Raw event archive and reconstructed recorder accounting: known — files [wired], in-memory [wired] (Persisted raw events and derived in-memory summaries are distinct parts.); Encountered-block set within one bot lifetime: known — in-memory [wired] (The recorded Set spans observations, but no disk reload is established.)	partial coverage (Live software/model admission is automatic unless a human makes the content decision. Manual critic approval is a distinct evidenced gate. Checkpoint selection by a generic caller does not establish per-entry manual admission.); Learning skill admission with automatic critic; direct manager admission logic: known — automatic [wired] (The driver success branch and manager deposit exclusion decide admission without further content approval. Generic direct-call origin itself is unresolved below.); Learning skill admission under manual critic approval: known — manual [wired] (A human explicitly confirms success of this candidate attempt before the learning driver retains its program; subsequent description generation remains automatic.); Skill-description generation and embedding/index writes after admission: known — automatic [wired] (Approval of a preceding task is not separate approval of the subsequently generated description.); QA misses, chest observations, history cleanup, event recording/replay and encountered-block updates: known — automatic [wired] (Models/code perform each store transformation. Human critic success may supply an outcome but does not manually edit each history entry or chest/event record.); Initial shipped/checkpoint contents, external replacements and caller-supplied direct-manager inputs: not-determinable (Caller identity and library-directory selection do not establish who supplied, edited or approved each entry. Historical/manual versus automatic admission remains not determinable.)	known coverage (The recorded consumer inventory includes requested retrieval/reload and independent automatic supply. Distinct operations within each chain are not counted twice; opaque historical content does not obscure these inspected delivery routes.); Driver-requested semantic skill retrieval and checkpoint/library reload: known — pull [wired] (The driver/constructor asks for retained material. Returning requested code is not an extra unsolicited push.); QA exact/nearest-question cache lookup and checkpoint reload: known — pull [wired] (A task-context or QA request obtains retained answers; the question index is an access part of that requested return.); Automatic all-active-code supply to each environment execution: known — push [wired] (The manager assembles every active definition without a per-skill consumer request; prompt retrieval is a separate pull route.); Automatic selected QA pairs, rendered chest/history context and encountered-block context in later model calls: known — push [wired] (Renderers supply memory as part of driver-owned model context; the models do not ask for these individual retained items.); Explicit resume of chest/history state and chronological recorder event replay: known — pull [wired] (Resume requests saved state; replay scans timestamp files, optionally to a cutoff. This is not implicit conversational-history reload.); Automatic use of retained progress and replayed iteration in later driver decisions: known — push [wired] (Once available, retained control state is supplied to sequencing/continuation logic without a per-item request from that consumer.)	known coverage (All recorded push operations use coarse availability, configured gates or whole-object supply. Embedding and name/key matching occur in requested pull chains and therefore do not establish inferred-embedding or identifier push findings.); Every-execution library assembly: trigger is an action step, selector is active map membership, selected part is all active skill code: known — coarse [wired] (No relevance threshold or per-name request selects definitions for this channel.); Curriculum QA display: trigger is eligible context rendering, inputs are progress/probability, eligible returned QA pairs and cap five: known — coarse [wired] (The push keeps eligible advice after fixed Unknown/language-model exclusions and a count budget. Nearest-question selection is separately a requested lookup, not an independent push.); Chest/history/block context: trigger is the relevant later model render, inputs are retained object availability and renderer/task/disclosure rules: known — coarse [wired] (Selected parts are rendered chest entries, task history and encountered names. Proximity is a chest acquisition rule, not identity-targeted push from the saved map.); Progress/iteration delivery to driver: trigger is later sequencing/continuation, inputs are retained counts and fixed thresholds: known — coarse [wired] (The selected retained parts are control counts; fixed branch use does not imply lexical or embedding targeting.)	partial coverage (Three recorded automatic trace-to-durable-behavior chains establish yes without observed gain. QA mixed input, archival replay and within-task/bot state do not establish a complete positive or bounded negative for every included write.); Successful task program to retained skill, then later action model/executor: known — yes [wired] (Task activity and success condition an automatic durable capability write with wired later consumers; direct admissions and historical libraries do not inherit this provenance.); Chest events to persisted map, then later planning/judgment/task selection: known — yes [wired] (The automatic trace-fed update persists and later affects model context and possible task override.); Task outcome to cleaned persisted history, then curriculum/disclosure/inference: known — yes [wired] (Critic/human outcomes feed an automatic history update; no observed improvement is needed for this classification.); QA answers whose questions can depend on observations: not-determinable (The recorded account distinguishes model-prior answers from trace extraction, but does not establish how much environmental experience reaches every qualifying generation path. Neither universal yes nor bounded no is assigned.); Raw event persistence and replay-derived accounting: not-determinable (Later control use is wired, but raw-log retention alone is insufficient and replay summaries are recorded as in-memory accounting, not a separately durable learned artifact. Qualification under the durable-output test remains unresolved.); Within-bot encounter set and within-task frame replacement: not-determinable (Cross-invocation context exists, but a durable trace-learned artifact is not established for these transient parts; a bounded absence record is not supplied.); Initial checkpoint/library production and unsampled historical writes: uninspected (The loaded artifacts are available, but no original production trace or complete admission history is supplied.)	partial coverage (Actual inspected consumers establish the listed forces, not compliance or benefit. Additional effects of unsampled executable bodies remain unresolved; static prompt instructions and provider weights are outside this memory boundary.); Retrieved retained code at the action model: known — knowledge [wired] (The model receives examples as context; these are not an access-control policy.); All active retained code definitions at Node eval: known — instruction [wired] (Symbolic executable directions enter the programs field. Actual function invocation is conditional; definition supply is not enforced primitive-only behavior.); Description text/vectors/name metadata and QA question index at similarity selection: known — ranking [wired] (Retained access representations affect relative retrieval priority; no learned ranking policy is required.); QA answers at curriculum/action context consumers: known — knowledge [wired] (Candidate advice is not an expected-result key or validated oracle.); Chest map at model renderers and hardcoded inventory-management selection: known — knowledge [wired], routing [wired] (The map has different effects at its two consumers.); Task history text at curriculum planning: known — knowledge [wired] (Judgment-derived labels are contextual evidence, not independent competence proofs.); Task-history count at disclosure, difficulty and inference sequencing: known — routing [wired] (The count selects inputs/next work; this is not a truth-validation criterion for stored skills.); Retained events/replayed iteration at learning-loop limit: known — routing [wired], enforcement [wired] (Only the iteration consumer has an established downstream driver consequence; other recorder metrics have no established model consumer.); Encountered-block set at curriculum rendering: known — knowledge [wired] (The set yields previously encountered other-block context rather than a spatial map.); Task experience at success-conditioned skill admission: known — learning [wired] (The automatic update turns task activity into durable executable behavior-shaping material; this authority is distinct from its later instruction/knowledge effects.); Unsampled retained skill bodies and their further executable effects: uninspected (The delivery/definition channel is inspected, but other checks, controls or knowledge assertions inside those bodies are not. No complete-value union is justified.)	AAS-2026-10-14-voyager-f16a44a0cfe3-01
WeKnora	2	Accumulated personal notes, note vectors/renderings, topic/affinity/tombstone controls; source files, document chunks/revisions and retrieval indexes; editable wiki bodies/revisions and access/withdrawal state; raw stored conversation/AgentSteps, generated or raw-fallback checkpoints and optional Q&A indexes; admitted skill files, instructions, snapshot/pointer and diagnostic state; later-queryable evaluation metrics. Included unresolved alternatives retain connector/manual/FAQ admission, graph/multimodal/enrichment derivatives, wiki map/taxonomy/issues, attachments/workspace payloads, suspended approval snapshots, MCP caches and host installation. Excludes shipped static prompts/code and ordinary unretained turn state, external model weights, provider internals and external host scheduling. Scope covers writes, maintenance, selectors and actual later consumers, not all runtime storage or permissions.	partial coverage (Known file, relational, vector, service-object and in-process parts coexist with backend-unspecified repositories and included uninspected alternatives. External model weights and shipped static prompts are excluded from retained memory.); Personal-note content and status/provenance, topic recurrence, familiarity and tombstones in GORM memory tables: known — rdbms [wired] (Table/repository evidence covers these relational objects, not every application backend.); Personal-note embedding rows: known — rdbms [wired] (The stored vector row is classified independently of the external producing model.); Document and optional conversation-passage retrieval indexes: known — vector [wired] (Vector is the operative access substrate, not a claim about backend implementation equivalence.); Uploaded source files and seeded installed skill files: known — files [wired] (Classifies the exposed file objects; remote object-store internals are not inferred.); Published remote skill snapshot: known — service-object [wired] (The exposed retained object is established; this classification does not describe its internal filesystem or provider database.); Evaluation metrics task map: known — in-memory [wired] (Memory includes later-invocation lookup; it need not survive process restart.); Stored message/AgentSteps/checkpoint rows, document chunks/revisions, wiki pages/revisions and Q&A Knowledge records: exact storage backend: not-determinable (Repositories and DB persistence are recorded, but these records do not pin every operative backend to a controlled substrate; personal-table evidence is not generalized.); Included unresolved alternatives: URL/manual/FAQ and connector admission; summaries/profiles, generated questions, OCR/captions, image vectors, graph derivatives and auto-tags; wiki map/taxonomy/issues; attachments/workspace artifacts, suspended approval snapshots and MCP caches; host installation alternative.: uninspected (Exact provider/cache/graph/multimodal storage and alternate backend coverage are not established in accepted records. Graph-related names alone do not establish a graph substrate.)	partial coverage (Automatic admission is wired on multiple routes. Manual-capable interfaces are retained as control uncertainty rather than converted into a supported manual finding.); Ordinary user-transcript extraction, recurrence promotion and reference-hit updates: known — automatic [wired] (Only flagged inferences are held pending; the positive does not classify manager/confirmation caller control.); Timed or forced memory maintenance, vector/backfill and cached rendering: known — automatic [wired] (Pressing ConsolidateNow does not make resulting content admission manual.); Document parsing/index enablement, wiki batch rewrite/structural AutoFix, Q&A indexing and eligible compaction checkpoint saves: known — automatic [wired] (Known automatic transformations do not settle initial upload control or every direct edit path.); Remote installer snapshot admission and verifier-directed repair: known — automatic [wired] (This covers inspected remote admission, not the host alternative.); Explicit/directive and manager create/edit/confirm/reject; chunk edit/revert; direct wiki user/agent write and supplied-source admission: not-determinable (API/service callers can be human or computation. Origin/manual/EditSource labels and authentication alone do not establish an explicit human content decision, so neither manual nor automatic is assigned to these caller decisions.); Connector/manual/FAQ details, graph/multimodal enrichment, approval snapshots/cache updates and host skill installation: uninspected (Distinct admission mechanisms remain incompletely recorded; automatic physical I/O cannot fill their control gaps.)	partial coverage (Both explicit consumer retrieval and application/maintenance supply are wired. Delivery inside a requested return is counted only as pull; unresolved payload and external-consumer routes remain included.); On-demand personal-memory, document, wiki and conversation-message searches/read requests: known — pull [wired] (Code selection and delivery fulfilling the request are not an additional push.); Evaluation-result lookup by task caller: known — pull [wired] (No automatic metric-driven successor is inferred.); Turn preparation personal recall and caller history/checkpoint loading: known — push [wired] (The user turn triggers the selector; the turn is not an explicit request for each retained item.); Query-understanding background and per-candidate familiarity conditioning: known — push [wired] (Candidate score adjustment is a later code consumer, not only model-prompt read-back.); Later extraction/review and page rewrite consuming retained notes, tombstones/topics and current page/reference state: known — push [wired] (These updater consumers are separate from requested foreground read tools.); Published skill pointer at subsequent workspace resolution: known — push [wired] (Specific later model selection of a skill is a different operation; full documentation selection was not traced.); Complete skill-document selection, external HTTP/MCP readers, attachments, approval/resume/cache payloads and graph/multimodal consumers: uninspected (Available routes or retained metadata do not settle every requested-versus-automatic delivery. Both established directions remain positive despite this incomplete inventory.)	partial coverage (Push routes establish coarse, identifier, lexical, embedding and model-judgment selection locally. Pull-only units are inapplicable locally, while unresolved push alternatives prevent whole-axis completeness or inapplicability.); Automatic personal recall at turn preparation: resident and situational selection: known — coarse [wired], inferred-lexical [wired], inferred-embedding [wired] (No chat-judgment selector is inferred: recall can invoke an embedder but does not generate recalled notes.); Query-understanding and affinity score push: known — coarse [wired], identifier [wired] (This identifier value is warranted by a match supplying retained counts, not merely by IDs attached to tool requests.); Automatic history/checkpoint replay to the next model call: known — coarse [wired] (Checkpoint IDs control exclusion/window routing; the recorded selector is recency/coverage, not an established targeted identity-match push finding.); Memory review candidate selection and merge-model decision: known — coarse [wired], inferred-lexical [wired], inferred-embedding [wired], inferred-judgment [wired] (Cluster admission and model-selected reduction are distinct selection operations in the maintenance chain.); Tombstone matching on later extraction/write admission: known — identifier [wired], coarse [wired] (The exact-fingerprint and bounded extracted-message checks differ from recent-list delivery; neither gives universal suppression.); Published image selection at later workspace startup: known — identifier [wired] (The selected snapshot is the delivered part, unlike a mere catalog identifier.); Full extraction existing-note/topic selector, wiki reducer page/source/link selector, wiki lint probes and skill-doc delivery: not-determinable (Some trigger/input/output facts are recorded, but complete selector alternatives and selected payloads are not classified. This prevents complete push-signal coverage.); Consumer-requested note/document/wiki/message and evaluation-result returns: inapplicable (These completely named requested-return operations are pull; push-signal classification does not apply to them. Their presence cannot make unresolved push routes inapplicable.); Graph/multimodal, attachments/workspaces, approval snapshots, caches and host-install alternatives: uninspected (Triggers and actual delivered parts of these included alternatives remain incompletely inspected.)	partial coverage (Wired yes findings cover personalization, familiarity feedback and continuation checkpoints. Other retention and capability paths stay unresolved; absence of improvement evidence is not trace-learning absence.); Automatic user-message extraction and recurrence-interest promotion to later recall/search/rewriter: known — yes [wired] (The trace-fed chain qualifies even without new knowledge or benefit measurements; inferred/pending outputs require later activation.); Completed-answer reference counts to later familiarity rerank and rewriter: known — yes [wired] (Counts encode selection experience, not satisfaction or verified relevance improvement.); Automatic model summary and raw-archive continuation checkpoints: known — yes [wired] (Both generated and raw-fallback continuation forms qualify as durable behavior-shaping artifacts. Transient previews alone are not this finding.); Q&A passage indexing to message-search results: not-determinable (Automatic trace retention and later search are wired, but accepted records stop at message-search output and do not establish a durable behavior-shaping consumer beyond that endpoint. No positive or bounded negative is assigned.); Verifier-directed setup repair and final skill image: not-determinable (Diagnostic findings affect repair and files later affect capability, but accepted records do not establish every persisted trace-to-durable-artifact write; transcript retention alone cannot classify this route.); Document/connector/wiki/multimodal/graph input origins; attachments, approval snapshots and remaining maintenance writes: uninspected (Imported content can include activity records, but exact original content and qualifying trace-fed later-consumer chains are not recorded for every included path. No system-wide no is warranted.)	partial coverage (Each value is attached to an actual retained part and consumer: knowledge, ranking, learning, validation, enforcement, routing and instruction coexist. Opaque consumers prevent complete coverage; delivery never establishes obedience or benefit.); Personal statements, document/chunk and wiki bodies, Q&A search results, replayed history and checkpoint text at their later readers: known — knowledge [wired] (Notes are explicitly wrapped as background data; preference content alone is not standing instruction authority.); Personal text/aliases and dense representations at recall/search selectors; document text/index features at retrieval: known — ranking [wired] (Ranking is the force at selection, distinct from knowledge at subsequent interpretation.); Familiarity counts at RAG rerank and profile/interests/titles at query rewriter: known — ranking [wired], knowledge [wired] (Rewriter context and executable score effects are separate consumers, neither a quality endorsement.); Retained user lines, prior notes/topics and answer references at automatic durable personalization updaters: known — learning [wired] (Authority is at the update consumer; wired trace-fed behavior shaping is not measured improved capacity.); Retained conversation/AgentSteps at automatic checkpoint writer: known — learning [wired] (This update-consumer force is separate from the checkpoint text knowledge effect and boundary routing effect.); Tombstone fingerprints/source-message windows at future write admission: known — validation [wired], enforcement [wired] (Tombstone topics also reach the extractor as guidance, but that does not turn deleted content into retained knowledge.); Retained memory status and chunk/wiki enablement at later eligibility checks: known — enforcement [wired] (Local selection constraints are not a universal stale-cache or backend-erasure guarantee.); Stored wiki table identities at machine rewrite gate: known — validation [wired], enforcement [wired] (The reference is retained content; the fixed shipped guard code is not itself accumulated memory.); Checkpoint coverage identifiers and published image pointer at later loaders/resolvers: known — routing [wired] (These are selected retained control parts, not model-interpreted text.); Retained skill documentation at future model skill use: known — instruction [wired] (Executable capability is described in rationale without inventing a controlled authority value for every executable effect.); Complete skill executable effects, attachments, approval snapshots, MCP caches, graph/multimodal derivatives, wiki issue/taxonomy and alternate admission consumers: uninspected (Opaque payloads and incompletely traced consumers prevent exhaustive authority classification. No blanket authority follows from retention alone.)	AAS-2026-10-15-weknora-7adf8a01c381-01
WikiSkill	2	Retained product task evidence, optional traces, original feedback, journal-backed patterns and replay/adoption metadata; research raw/selected experience, Wiki pages/indexes, impact/proposal/checkpoint artifacts, SKILL text and named SKILL/PURPOSE mappings; inheritance and deployment copies/backups; their write, maintenance and later-consumer routes. Additional OfficeQA and Spreadsheet callable controllers remain included but unassessed. Excludes shipped static role/prompts, external model weights and host context management, and task corpora except retained experience used by these routes.	partial coverage (All classified retained objects use files; the version-control repository is source distribution, not an evidenced operational memory substrate. Additional callable research controllers remain uninspected, so this is not complete system enumeration.); Inspected product retained patterns, original feedback, skill files, evidence payloads and journal metadata: known — files [wired] (Encoding and hash links do not turn local files into a graph or database.); Research Wiki, impact history, raw/selected traces, named SKILL/PURPOSE mappings and checkpoints: known — files [wired] (Temporary staging is distinguished from persistent source artifacts.); Exported/installed deployment copy and restoration backup: known — files [wired] (Filesystem deployment is wired; subsequent host consumption is only afforded.); Additional OfficeQA and Spreadsheet callable research controllers (officeqa/loop.py and spreadsheet/loop.py), including promotion and candidate-file resume alternatives: uninspected (The accepted epistemic coverage leaves these controllers unassessed; the named engine/study records do not enumerate them. Their storage_substrate cannot be inferred from the inspected routes.)	partial coverage (Known automatic role and controller paths coexist with unresolved caller-controlled admission; human authorship is not authenticated. Additional callable research controllers remain uninspected, so this is not complete system enumeration.); Role-produced Wiki/pattern edits, candidate submission and software successor admission: known — automatic [wired] (Starting a run, configuring a scorer or supplying verdicts does not make subsequent content admission manual.); Derived context views, event/history writes and checkpoint persistence: known — automatic [wired] (Physical writes and content control are distinguished.); Initial skill/feedback/task-evidence supply, generic direct submissions, custom adapter initialization and deployment replacement: not-determinable (Generic callers and origin labels do not establish a human decision. A caller may be human or agent; no manual finding follows merely from an explicit replacement flag.); Additional OfficeQA and Spreadsheet callable research controllers (officeqa/loop.py and spreadsheet/loop.py), including promotion and candidate-file resume alternatives: uninspected (The accepted epistemic coverage leaves these controllers unassessed; the named engine/study records do not enumerate them. Their write_agency cannot be inferred from the inspected routes.)	partial coverage (Push selection and downstream on-demand pull occur in the same chains; requested work and requested files are not conflated. Additional callable research controllers remain uninspected, so this is not complete system enumeration.); Automatic phase/round-specific selection for product and native role/task invocations: known — push [wired] (The work/status request is not a request identifying every selected retained part.); Automatic research/study learning and rollout context assembly: known — push [wired] (Library adapter wiring does not establish default CLI activation.); Explicit staged-file requests by research learning models: known — pull [wired] (A requested file return is not an additional push.); External host consumption after retained-skill export, installation or restoration: uninspected (Export/install/restore wire filesystem effects, not an independent later-consumer invocation. Future host consumption is afforded, but its request/selection/delivery operations are uninspected; neither pull nor push is established for that external route.); Additional OfficeQA and Spreadsheet callable research controllers (officeqa/loop.py and spreadsheet/loop.py), including promotion and candidate-file resume alternatives: uninspected (The accepted epistemic coverage leaves these controllers unassessed; the named engine/study records do not enumerate them. Their read_back_direction cannot be inferred from the inspected routes.)	partial coverage (Known push signals name actual selected parts; opaque host discovery and unassessed controllers prevent complete selector coverage. Additional callable research controllers remain uninspected, so this is not complete system enumeration.); Product/native automatic learning-context and task-skill push: known — identifier [wired], coarse [wired] (Trigger is work issuance; selectors use controller state, delivering evidence/history to authors or skill to executor.); Default research and isolated-study automatic evidence/context push: known — coarse [wired], identifier [wired] (Sampling by outcome and budgets is not embedding or lexical relevance retrieval. Optional exact-file pulls add no push signal.); External host discovery of installed retained skill: uninspected (Installation affords later host consumption but does not inspect its trigger or selector. No inferred-judgment or identifier push follows from a filename alone.); Additional OfficeQA and Spreadsheet callable research controllers (officeqa/loop.py and spreadsheet/loop.py), including promotion and candidate-file resume alternatives: uninspected (The accepted epistemic coverage leaves these controllers unassessed; the named engine/study records do not enumerate them. Their read_back_signal cannot be inferred from the inspected routes.)	partial coverage (Positive trace-fed procedural updates do not require weight training, novel knowledge, observed compliance or demonstrated improvement. Additional callable research controllers remain uninspected, so this is not complete system enumeration.); Default research trace-to-Wiki-to-skill and paper/study trace-fed durable instruction paths: known — yes [wired] (In the one-round study, Maintainer Wiki reaches Proposer and candidate reaches validation; no second deployed final-skill round is implied.); Product optional-trace-to-pattern/skill route: partial — yes [wired] (A qualifying trace-fed update is wired when optional traces contain execution experience; outputs/scores/feedback alone do not establish universal trace origin.); Opaque caller-supplied experience and custom adapter alternatives: not-determinable (Payload origin and alternative callback behavior are unresolved; no system-wide no or bounded absence is established.); Additional OfficeQA and Spreadsheet callable research controllers (officeqa/loop.py and spreadsheet/loop.py), including promotion and candidate-file resume alternatives: uninspected (The accepted epistemic coverage leaves these controllers unassessed; the named engine/study records do not enumerate them. Their trace_learning cannot be inferred from the inspected routes.)	partial coverage (Actual consumers distinguish authoring knowledge, update learning, downstream instruction and retained control forces; no ranking positive is established by staged-file availability. Additional callable research controllers remain uninspected, so this is not complete system enumeration.); Patterns, research Wiki, original feedback, purposes and history at learning authors: known — knowledge [wired], instruction [wired] (Knowledge and instruction apply to the interpreted portions, not every metadata field.); Retained execution evidence and Wiki at automatic procedural-update consumers: known — learning [wired] (Learning authority names the update consumer; it does not assert observed improved capacity.); Retained SKILL bodies at later task executors: known — instruction [wired] (PURPOSE does not accompany task injection; delivered instructions grant no new permissions.); Journal adoption/request metadata and research resume records at controllers: known — routing [wired], validation [wired], enforcement [wired] (These are retained control parts, not authority attributed to natural-language lessons.); Original explicit-intent source IDs at feedback finish: known — validation [wired], enforcement [wired] (Caller-provided fulfillment judgments remain unverified.); Opaque product outputs and optional traces at later authors: not-determinable (The evidence route is known, but payload-specific authority cannot be classified from a result row alone.); Additional OfficeQA and Spreadsheet callable research controllers (officeqa/loop.py and spreadsheet/loop.py), including promotion and candidate-file resume alternatives: uninspected (The accepted epistemic coverage leaves these controllers unassessed; the named engine/study records do not enumerate them. Their behavioral_authority cannot be inferred from the inspected routes.)	AAS-2026-10-15-wikiskill-baafad05d35e-01
wuphf	2	Use-accumulated private/shared notes, learning records, raw artifacts and derived facts/articles, embedding cache and retrieval metadata, compiled skill/registry state, saved executable tools, routine transcripts/artifacts and retained task outcomes/verification. Includes unresolved remote corpus/index/provider state, opaque native-session entries, skill/scanner/admission alternatives, app-state consumers and separate durable checkpoint/resume alternatives. Current-state snapshot helpers are assessed only as an unresolved retention alternative. Excludes shipped static prompts, transient invocation frames, uncaptured workspace contents and external-provider implementation claims; inference weights alone are not use-accumulated memory.	partial coverage (Positive file, repository and live-index witnesses do not resolve remote pages, provider state, checkpoints or app/registry internals; the included memory boundary is not a complete storage census.); Private note map and persisted broker-state encoding: known — in-memory [wired], files [wired] (Both live and persisted parts are identified; JSON is encoding, not an additional substrate.); GBrain shared page and remote retrieval index: not-determinable (Adapter identifies remote pages and queries, but does not establish the remote storage implementation. A service endpoint alone does not establish a service-object substrate.); Learning log, source artifacts, fact logs, entity articles and compiled skill files: known — files [wired], repo [wired] (File and repository retention coexist; article narrative has its own object record.); Live derived fact/index state: known — in-memory [wired] (Classifies the inspected live part, not every backing index implementation.); Cached dense text embeddings: known — files [wired] (Cache encoding is documented independently of provider inference; no dedicated vector database is established.); Saved tool code and metadata: known — files [wired] (Version counters do not establish retained prior code bodies.); Routine transcript trees: known — files [wired] (Only the wrapper-visible representation is characterized; native dependency internals remain opaque.); Separate durable session checkpoint and app-state/version internals: not-determinable (Snapshot structs are a current-state projection, not established retained checkpoint storage; app-store internals and retention details were not traced.); Provider-owned embedding state and additional registry/index persistence details: not-determinable (Provider initialization and registry backing details are not fully characterized. Static inference weights are not claimed as use-accumulated memory.)	partial coverage (Known software admissions and the specifically evidenced human edit coexist. Untraced initial/replacement mechanisms prevent a complete automatic/manual inventory.); MCP scoped-note and shared-promotion software admission: known — automatic [wired] (Caller initiation and caller-authored prose do not make this admission manual.); Explicit typed-learning validator and completed-task capture: known — automatic [wired] (Asserted user-stated source is not authenticated human approval.); Extraction, synthesis, compression/archive and skill-registration software writes: known — automatic [wired] (Known subsequent transformations remain automatic despite unresolved initial content control.); Saved-tool authoring/upsert, including routine fallback: known — automatic [wired] (Manual build endpoint and scheduled routine have different provenance conditions but both software admission.); Routine transcript/outcome persistence and vector-cache writes: known — automatic [wired] (Recording a human-supplied prompt is not a separate manual admission of generated outcomes.); Human-edited existing brief body: known — manual [wired] (This is the separately evidenced human edit, not a human veto over newly generated learned content.); Contradiction-resolution caller, rejection-tombstone authoring, remote initialization and app/checkpoint replacements: not-determinable (Synthetic human identity, generic caller identity, storage API and read selection do not establish who controls each content admission.)	partial coverage (Both pull and push are wired on named routes; unresolved alternatives prevent a complete direction inventory.); Requested note/learning/wiki searches and article-body reads: known — pull [wired] (A requested article return is not another push merely because code delivers it.); Automatic notification and task packet delivery: known — push [wired] (Backend Query inside an unsolicited packet remains part of push to the office consumer.); Saved-tool manual request and routine reuse: known — pull [wired], push [wired] (Two distinct selectors/operations, not duplicate counting of one requested return.); Synthesis and compilation maintenance reads: known — pull [wired] (These later update consumers are distinct from automatic delivery to the office model.); Routine transcript/artifact requested frontend retrieval: known — pull [wired] (No automatic conversational replay into tool authoring is established.); Full skill/native-session loading, remote alternatives, app-state reuse and durable resume: not-determinable (Known request and push routes cannot establish all unresolved consumers or durable recovery wiring.)	partial coverage (Known push selectors use lexical, dense, identity and coarse signals. Remote and untraced branches prevent complete classification; no inferred-judgment selector is established by the accepted records.); Notification-triggered private note push: known — inferred-lexical [wired], coarse [wired] (Remote backend selection is a separate opaque branch, not upgraded by this local witness.); Notification-triggered GBrain result push: not-determinable (Notification-derived query and hit limits are known, but remote ranking/selection mechanism is opaque; no lexical/embedding value is inferred from Query.); Task-triggered learning/wiki/article-pointer push: partial — inferred-lexical [wired], inferred-embedding [wired], coarse [wired] (Known local lexical and configured dense mechanisms do not resolve every remote wiki-index branch.); Dependency-triggered upstream outcome push: known — identifier [wired], coarse [wired] (Identifiers select delivered retained parts here, rather than merely naming a requested file.); Scheduled routine automatic saved-tool selection: known — identifier [wired], inferred-lexical [wired] (Due routine is trigger; current prompt is selector input; selected part is stored executable code.); Untraced skill/app/checkpoint push alternatives: not-determinable (Their direction and any automatic selector are unresolved; a known pull interface cannot make them inapplicable.); Explicit pull-only note/learning/transcript and maintenance-request operations: inapplicable (Push selection signal is inapplicable to these named requested operations, not to the mixed boundary.)	partial coverage (Task-derived capture and smoke-feedback-directed durable repair supply wired positives. Other trace origin/write/consumer chains remain unresolved; no system-wide negative or observed learning outcome is asserted.); Completed-task state to durable insight/entity articles to later task context: known — yes [wired] (Qualifies without a model update, novel claim or demonstrated improvement; the verification gate covers only one branch.); Artifact extraction to facts and future snippets: not-determinable (Automatic write and later knowledge consumer are wired, but original artifact content can be static or trace-bearing. Trace origin is not established for every included artifact.); Fact-fed brief rewrites and task-derived playbook/skill alternatives: not-determinable (Fact provenance and full draft/skill consumption chain are unresolved; procedure registration alone does not complete every trace-fed later-consumer route.); Smoke-error-directed repair to persisted code to later worker: known — yes [wired] (Conditional repair/persistence wiring qualifies even within one authoring episode. No successful business repair or repeated learning over routine history is established.); Routine transcript feedback alternatives: not-determinable (Raw history is retained for readers, but not supplied as author context in the inspected routine. Other native entry consumers remain opaque; this does not establish system-wide absence.); Remote initialization, native compaction, app state and durable resume alternatives: not-determinable (Opaque or untraced writes cannot be converted into bounded absence. Raw log retention and current-state projection alone do not establish qualifying trace learning.)	partial coverage (Authority is local to the consumed retained part. Known ranking, routing, validation, enforcement, learning and knowledge do not upgrade the afforded procedural instruction route or resolve opaque consumers.); Selected note/learning/fact/article/outcome text at office model or requesting reader: known — knowledge [wired] (Task apply/cite/read framing is shipped delivery guidance, not evidence that every stored insight is a standing instruction.); Retained text, confidence, recency and vectors at retrieval/matching selectors: known — ranking [wired] (Retained ranking inputs have ranking force even though scoring code is fixed.); Dependency IDs, skill source identity and synthesis fact-count checkpoint at next-input selectors: known — routing [wired] (These retained fields select scope or operation, not truth.); Stored executable code at worker: known — routing [wired] (Code has executable control, not merely advisory knowledge. Current-run host gates are not classified as retained memory controls.); Retained task results at automatic distillation: known — learning [wired] (Learning force at the update consumer is distinct from knowledge force at the later office model; neither establishes benefit.); Smoke error feedback at tool repair model: known — learning [wired] (The error-directed durable update is distinct from the stored code downstream routing authority; simulated feedback does not establish business correctness.); Persisted Pass and prior TaskID at learning admission: known — validation [wired], enforcement [wired] (The verification result licenses only its checked domain, not all narrative claims.); Rejection tombstones at compiled-skill registration: known — validation [wired], enforcement [wired] (Rejection authorship/undo is not thereby established.); Archive flag at article-file search and contradiction acknowledgment at lint: known — enforcement [wired], validation [wired] (These are bounded consumers; full expiry/deletion propagation through every query is unresolved.); Compiled wrapper directions at intended later agent consumer: partial — instruction [afforded] (Direction text and intended consumer are established, but assignment/loading/invocation are not fully traced.); Routine transcript and outcome artifact at requested frontend presentation: known — knowledge [wired] (Workflow annotation on the runtime route establishes a shipped request/presentation path, not human receipt.); Remote index effects, app-state consumers, checkpoint restoration and full skill/native-session consumption: not-determinable (These opaque or untraced alternatives prevent complete enumeration of authority paths.)	AAS-2026-10-15-wuphf-b246722f99c0-01
xMemory	2	Retained originals, episode title/content, semantic statements, provenance/access fields, embeddings, theme summaries/topology, semantic-kNN, diagnostic graph exports and original-message indexes; their admission, maintenance, update-consumer and shipped QA read-back routes. Includes unresolved generic input admission and injected storage/generator alternatives. Excludes static shipped prompts, inference model weights not updated by these routes, current-query NLL/prefix-KV state, and offline QA/metric artifacts with no memory-update consumer. Hidden provider retention remains uninspected, not presumed absent.	partial coverage (The recorded default inventory supports files, RAM, vector and graph storage. Opaque alternatives prevent complete substrate coverage.); Filesystem episode/original/semantic objects, metadata and theme/kNN/vector companion artifacts: known — files [wired] (Covers the recorded local-file objects, not arbitrary injected repositories.); RAM repository alternatives, loaded memory dictionaries/caches and original-message lexical indexes: known — in-memory [wired] (Current-query scoring state is excluded; loaded accumulated memory and callable lexical indexes are included.); Chroma episode/semantic retrieval index: known — vector [wired] (Optional theme-Chroma helpers do not establish mandatory theme synchronization.); HierarchicalMemoryGraph diagnostic export and reload: known — graph [wired], files [wired], in-memory [wired] (Graph representation and its physical persistence are both supported.); Injected repository/index implementations and hidden provider-retained state: uninspected (Injection and remote endpoint boundaries do not reveal alternative memory substrates or provider retention.)	partial coverage (Automatic generation/admission is established locally. No positive manual memory admission is established by caller identity or human workflow initiation.); Dataset formatting and automatic episode/original, semantic, vector/index, theme and graph admission: known — automatic [wired] (An operator starting construction or selecting maintenance timing does not make subsequent content admission manual.); Generic incoming content supply, explicit withdrawal and arbitrary injected implementation admission: not-determinable (Generic caller identity leaves human control unresolved. Selecting a startup model is not an explicit human edit or approval of accumulated memory.)	partial coverage (Requested retrieval is pull; automatic prepared-input supply to update models is push. Unknown helper/injected consumers prevent complete direction coverage.); Facade/baseline/lightweight search and shipped adaptive question answering: known — pull [wired] (Automatic query ranking is not a separate unsolicited push.); Episode-event semantic continuation, selected prior statements into prediction and statements into theme summarization: known — push [wired] (A driver may initiate maintenance; automatic selection and model delivery within that update remain separate operations from the upstream operator request.); Diagnostic graph reload and operator-invoked visualization: known — pull [wired] (No automatic graph-to-answer push is inferred.); Standalone original-message helper callers and injected/hidden-provider read-back: not-determinable (Callable search capability alone supplies no additional later consumer; opaque alternatives have no recorded complete delivery chain.)	partial coverage (Recorded update pushes use embedding, identity and coarse availability/order selection. The whole axis is not inapplicable because pull-based QA coexists with update pushes and unresolved alternatives.); Prior semantic statement selection for automatic prediction calls: known — inferred-embedding [wired], coarse [wired] (The memory annotation on the runtime route supplies both alternatives and their default count limit.); Episode-created task reload into semantic generation: known — identifier [wired] (Cold start supplies episode content; comparison can supply original messages.); Maintenance selection of statement text for theme summaries: known — identifier [wired], coarse [wired] (Availability and actual ID matching are separate selection inputs, not proof of complete membership coverage.); Requested baseline/adaptive/diagnostic pull operations: inapplicable (This axis classifies push selectors; lexical, embedding and judgment ranking within these requested pull chains is not an additional push signal.); Unknown helper/injected/provider delivery alternatives: not-determinable (Unknown direction and later consumers prevent either a positive push-signal classification or an inapplicability conclusion.)	partial coverage (There is a wired positive trace-fed memory route, while alternative update coverage remains unresolved. No system-wide negative follows from inference-only model calls or offline scores lacking feedback.); Automatic episode summaries, semantic extraction/comparison and downstream theme updates: known — yes [wired] (This is the profile's durable trace-fed write meaning. It establishes neither improved capacity nor parameter training, source faithfulness or successful operation.); Arbitrary injected generators/storage and hidden provider update paths: uninspected (No complete write/retention/later-consumer chain is recorded for opaque alternatives; the known positive does not classify them.)	partial coverage (Knowledge, update learning, ranking, routing, validation and context/insertion enforcement have distinct consumers. No retained standing-instruction path is established; opaque alternatives remain unresolved.); Originals, episode prose and semantic statements at final answering; theme prose at adaptive scoring: known — knowledge [wired] (Delivery is wired; compliance, beneficial influence and factual correctness remain unobserved.); Retained conversations/episodes/statements consumed to produce durable episode, semantic and theme material: known — learning [wired], knowledge [wired] (Update authority is separate from downstream knowledge authority; no parameter update is required.); Retained prose, vectors, lexical structures and weighted adjacency at retrieval/representative selection: known — ranking [wired] (The unspecified final callers of standalone original-message helpers do not establish another final-model path.); Stored owner/source IDs and theme membership at scope and candidate selection: known — routing [wired], enforcement [wired] (Unconsumed confidence and revision counters receive no authority merely because they persist.); Prior same-type semantic content/embeddings at duplicate checking and insertion: known — validation [wired], enforcement [wired] (Validation and enforced insertion veto are distinct effects of the same retained comparison material.); Diagnostic graph interpretation and unspecified helper/injected consumers: not-determinable (A diagnostic reader exists, but the records do not establish a further behavior-shaping model/update consumer for every graph part or arbitrary helper caller.)	AAS-2026-10-15-xmemory-b9a2b9f58eda-01
Zikkaron	2	Use-accumulated memory text, raw tool/sensory traces, archives, enrichment, profiles/beliefs, reminders, checkpoints, narratives, hierarchy/graph/vector access structures, rules and selection/provenance state; includes RAM buffers and use-dependent dense caches, library-only alternatives and shipped evaluator ingestion/consumption. Excludes fixed pretrained model artifacts, static shipped host instructions/hook configuration, transient current-call outputs and benchmark reports without an established automatic later update consumer; unresolved input origins and library integration remain included uncertainties.	known coverage (The accepted object inventory covers SQLite retention and its graph/vector structures plus use-dependent RAM state. External pretrained models and static installation files are excluded from this memory boundary, not silently classified as learned memory.); SQLite-retained text, traces, archives, enrichment, profiles, beliefs, checkpoints, reminders, narratives, rules and access state: known — sqlite [wired] (Object and route records cover the retained database inventory, including callable/library-populated parts; unknown writer deployment does not make their documented substrate unknown.); Relational graph, causal edges and hierarchy access structures within SQLite: known — graph [wired] (Graph is the operative structured substrate, not a claim of a separate graph server.); Dense embedding/vector index and retained centroid/HDC representations: known — vector [wired], sqlite [wired] (The enriched row/vector-index discrepancy concerns contents and consumers, not the existence or location of these stores.); Sensory buffer, mutable HDC codebook, Hopfield patterns and cognitive-map caches: known — in-memory [wired] (Includes RAM state accumulated or rebuilt through use, not all transient variables or externally loaded fixed model weights.)	partial coverage (Automatic derived admissions are established locally. Unresolved caller control and initial population prevent complete coverage and prevent asserting manual agency.); Post-gate curation and query-triggered revision: known — automatic [wired] (Calling remember or recall is not approval of every subsequent mutation.); Capture, consolidation, sleep, compression, seeding, conditional extraction/enrichment, hierarchy and event-checkpoint admissions: known — automatic [wired] (Includes automatic placeholder/micro-checkpoint mechanisms separately from caller-supplied snapshots.); Benchmark conversational-input reformulation and retained-row admission: known — automatic [wired] (Operator selection starts the benchmark but does not approve each resulting memory.); Library CRDT and belief-derivation admissions: known — automatic [afforded] (Afforded admission is not upgraded by the wired profile or ordinary-memory alternatives.); Caller-supplied memory, anchor/snapshot/rule/trigger content, ratings and initial beliefs: not-determinable (Generic MCP/library caller identity does not establish an explicit human content decision. Manual agency is not inferred from tool names or a workflow being operator-started.)	partial coverage (Both requested and unsolicited routes are wired. Library caller and resource-route coverage limits remain explicit instead of dropping those alternatives.); Requested recall, project context, hierarchy/drill-down, story, explicit replay and benchmark retrieval-to-generation: known — pull [wired] (Requested result delivery is not counted as another push merely because selection is automated.); Host prompt/session events and post-compaction hook supply: known — push [wired] (Host event execution and stdout consumption remain unobserved external contracts; the delivery path is wired.); Prospective reminder matching and related-memory reinjection in remember responses: known — push [wired] (The current remember request does not itself request the independently selected retained additions.); Archive decompression, library-only consumers and incompletely recorded resource routes: not-determinable (An archive method alone is not a later host route; ordinary callers and alternative resource details are not fully established. These gaps cannot be turned into pull-only or no-read-back findings.)	partial coverage (Known push selectors support lexical, embedding, identifier and coarse values. Prospective grammar and canonical reinjection selector detail remain unresolved; service-only model rerankers do not establish inferred-judgment push.); UserPromptSubmit hook: prompt/cwd to selected current memory rows: known — inferred-lexical [wired], inferred-embedding [wired], identifier [wired], coarse [wired] (Each finding names a selection step, not a generic identifier on an output row. Global fallback does not erase directory targeting.); SessionStart/compact event: checkpoint, directory memories, anchors and recent actions/memories: known — coarse [wired], identifier [wired], inferred-embedding [wired] (Separates global/coarse sections, directory matching and optional numerical prediction within the unsolicited restore chain.); Remember-time prospective matching: partial — coarse [wired] (The trigger uses current content/directory/tags/time to select reminder text, but full matching grammar is not established. Lexical, identifier or other targeted-signal coverage cannot be completed from trigger names.); Related-memory reinjection after remember: not-determinable (The canonical route establishes optional reinjected related context, but does not fully record its selector. Member prose alone cannot establish an additional controlled signal finding.); Explicit recall, project-context, hierarchy, story and requested restore delivery: inapplicable (These operations fulfill requests. Their selectors are not push signals; automatic hook alternatives are classified separately.)	partial coverage (Positive automatic trace-to-durable-artifact-to-later-consumer chains, including benchmark conversational-input reformulation and later generation context, establish yes at wiring strength. Judgment feedback remains separate and unresolved. No absence record warrants a no union; mixed-origin and external-write routes keep this axis partial.); Captured tool activity to summary memory and sensory episodes to graph/derived text: known — yes [wired] (Raw retention alone is not the positive: the recorded transformations and later consumers close the architectural chain.); Recall queries and access transitions to revised text or navigation/selection state: known — yes [wired] (The records establish durable behavior-shaping updates, not successful learning or actual exercise.); Usefulness feedback to retained confidence/stability: known — yes [wired] (Caller evaluation provenance is unresolved, but automatic derivation/admission and the later consumer are wired.); Compaction event to retained checkpoint epoch/placeholder to later restore: known — yes [wired] (This narrow event-fed chain does not acquire a complete conversational trace or demonstrate benefit.); Remember-derived micro-checkpoints and mixed-source extracted/profile/narrative products: partial — yes [wired] (Original caller content is not necessarily a trace. The mixed-origin branches cannot all be asserted trace-fed from their common storage alone.); Initial beliefs, library sync, external caller admissions and uninspected maintenance alternatives: not-determinable (The records do not settle whether every included write has qualifying trace input and a later consumer. A wired positive elsewhere cannot establish a system-wide negative for these routes.); Benchmark conversational turns reformulated into retained rows and later generation context: known — yes [wired] (The workflow account and annotation specify this preparation/read-back chain on the canonical route. Qualification does not depend on judgment-to-production feedback, observed execution or demonstrated benefit.); Saved benchmark judgments at the measured-report boundary: not-determinable (Judgments can be saved, but no automatic production-policy consumer or later behavior-shaping transformation is established. Uninspected later report use prevents a bounded no finding.)	partial coverage (The actual consumers support knowledge, ranking, routing, validation, enforcement and learning. No static shipped instruction authority is imported into memory; particular reminders and untraced consumers keep coverage partial.); Current memory, trace summaries/actions, profiles, beliefs, narratives, hierarchy summaries and replay prose at host or benchmark model: known — knowledge [wired] (Informational delivery is wired; actual model reliance and benefit are unobserved.); Retained prospective reminder text delivered on matching remember calls: partial — knowledge [wired] (The records allow reminder text to contain instructions, but do not establish a standing-guidance consumption contract. Instruction authority of particular reminder contents remains unresolved.); Memory text/vectors, heat/confidence/access state, graph/hierarchy and numerical patterns at selectors: known — ranking [wired] (The enrichment record preserves row/index divergence; ranking authority is consumer-specific, not uniform across indexes.); Retained retrieval rules at RulesEngine: known — enforcement [wired], ranking [wired] (These are retrieval controls, not host tool permissions or a system-wide authorization boundary.); Checkpoint active/epoch fields, graph navigation and trigger conditions at fixed-code consumers: known — routing [wired], enforcement [wired], validation [wired] (Separates operation/scope selection, eligibility criteria and veto effects from informational text delivery.); Retained traces and experience at automatic durable update consumers: known — learning [wired] (Learning authority at update consumers is separate from downstream knowledge/ranking authority, and does not establish improved capacity.); Recorded benchmark turns at automatic memory reformulation: known — learning [wired] (The workflow specifies the update and later consumer. This authority is separate from downstream knowledge and from judgment/report feedback.); Archive/enrichment rationale and untraced control consumers: not-determinable (No ordinary archive/rationale reader is established; full graph/domain and derivative consumers are not resolved. Stored metadata alone establishes no additional authority.)	AAS-2026-10-15-zikkaron-0102ec63c6e5-01

doc-grounded (15)

System	Profile revision	Compared boundary	Storage	Write agency	Read-back	Push selection	Trace learning	Authority	Evidence
AgeMem	2	Included: accumulated LTM content, retained dense encodings and access metadata; active conversational material after accumulation, Summary or Filter and its later consumers; collected training experience, reward-derived supervision and changed policy weights; documented explicit and driver-owned admission/read-back alternatives, including unresolved state serialization and checkpoint/provider persistence. Excluded: static shipped tool/summarizer/evaluator prompts and schemas, ordinary unchanged current-call state, imported benchmark reference material as a separate memory store, publication-only metrics, external framework internals and other systems. The horizon established for textual memory is later invocations and stages of an episode; training parameters reach subsequent training decisions and reported evaluation, not an inspected deployment.	partial coverage (The inventory includes LTM text, embeddings and metadata, active context and summaries, training experience and updated policy parameters. Only the parameter substrate has a controlled value; abstract stores and buffers do not identify files, RAM, databases or a vector backend. Deployment checkpoint and provider-state persistence remain unresolved.); LTM content, embedding tuples and access metadata: not-determinable (An abstract external store is specified. Dense encodings do not establish a vector storage substrate; backend and reload implementation are uninspected.); Active conversational context, surviving messages and replacement summaries: not-determinable (A message list and replacement prose are retained between invocations. Their abstract notation does not establish an in-memory backend or durable transcript storage.); Training experience buffer and numerical supervision: not-determinable (An abstract buffer is reinitialized in training. Serialized logs, replay retention and the physical substrate are unspecified.); Policy parameters changed by GRPO and used in later invocations: known — model-weights [claimed] (The operative retained result is updated weights; checkpoint files and serving infrastructure are not separately identified.)	partial coverage (The documented store mutations, context transformations, trajectory collection and gradient updates use computational admission. Researcher setup and preference input do not make these admissions manual. Initial backbone admission, checkpoint replacement and complete context/metadata initialization control remain unknown.); Policy Add/Update/Delete request admitted to the abstract store: known — automatic [claimed] (For this documented mechanism the model proposes and the described tool installs the change, without a per-content operator decision.); Summary replacement and Filter omission, including task-derived criterion alternative: known — automatic [claimed] (Both criterion alternatives and the summary installation mechanism are computational. Missing criterion derivation or preservation checks do not turn the stated admission manual.); Experience collection and GRPO parameter-update admission: known — automatic [claimed] (Researchers define the objective and launch training; the driver collects experience and gradient ascent applies each update.); Initial context/metadata supply and backbone or deployment checkpoint admission: not-determinable (The records name input selection and a backbone but do not identify all content-control principals or a separate serving replacement decision. Read selection cannot establish manual or automatic checkpoint admission.)	partial coverage (Both model-requested retrieval and independently triggered Stage 1 delivery exist in the specified design. Later context and training consumers are included separately. Exact full-store state serialization and deployment checkpoint routes remain unresolved rather than being excluded as pull-only.); Policy-requested Retrieve_memory selection and return: known — pull [claimed] (The selector fulfills the policy request. Its return is not a second push merely because code inserts results into context.); Stage 1 pre-action retrieval: known — push [claimed] (Algorithm 3 retrieves before policy choice, independently of an explicit request for retained material.); Replacement summary and surviving context supplied to subsequent policy invocation: known — push [claimed] (Requesting a transformation does not itself request later read-back; the driver later carries retained transformed context forward.); Buffered experience to optimizer and changed parameters to subsequent training decisions: known — push [claimed] (The training driver schedules consumption of the collected trajectory group and current policy parameters; no target-policy retrieval request controls these paths.); Additional model-visible store serialization and deployment checkpoint read-back: not-determinable (Abstract state tuples do not settle full-store input serialization, and no deployment loader is inspected. Known delivery paths do not resolve these alternatives.)	partial coverage (Signals classify automatic delivery operations, not query syntax on the pull route. Stage retrieval and Filter use embedding-based selection; context-span carryover and grouped training supply are coarse. Unknown state serialization and serving selection prevent complete coverage.); Explicit policy-requested retrieval return: inapplicable (This completely named operation fulfills a read request; signal characterizes push selection rather than this pull response.); Stage 1 driver retrieval selecting stored entries: known — inferred-embedding [claimed] (Trigger: sampled Stage 1 message before the policy action. Input: message encoding and existing memory embeddings. Selected part: cosine top-k entries; k limits count, not tokens.); Filter-selected surviving messages delivered on later policy call: known — inferred-embedding [claimed] (Trigger: context-maintenance action followed by the next invocation. Input: retained messages and supplied or task-derived criterion. Selected part: messages surviving the similarity threshold. The automatic criterion derivation remains unspecified.); Summary span and ordinary retained-context carryover to subsequent policy calls: known — coarse [claimed] (Trigger: later invocation or Stage 3 transition. Input: current retained context and policy-selected all/N span for compaction. Selected part: replacement prose and remaining messages, without an evidenced identity match or targeted query.); Training-group experience and current updated weights supplied by the driver: known — coarse [claimed] (Trigger: group optimization and subsequent training invocation. Input: current episode group and current updated policy. Selected part: collected experience and changed parameters, not retrieved prose entries.); Whole-store state visibility and deployed checkpoint selection: not-determinable (No serialization or serving loader establishes a selector, trigger-input contract or selected payload for these included alternatives.)	partial coverage (Positive qualifying writes are conversational store extraction, continuation summarization and trajectory-fed parameter optimization, all design claims. Original context provenance and checkpoint/provider alternatives remain incompletely assessed; no system-wide no is warranted.); Automatic conversational extraction to LTM followed by later retrieval: partial — yes [claimed] (The model writes information from presented conversation and illustrative user preferences, retaining it for later policy context. Static supporting-fact inputs and arbitrary generated strings need not all be trace-fed; their provenance and exact derivation remain unresolved.); Automatic conversation summary to later policy context: known — yes [claimed] (Selected prior conversational activity produces retained replacement prose that affects the input available at the next invocation; within-episode retention suffices for this route.); Automatic trajectory/reward processing into policy weights: known — yes [claimed] (Collected multi-step experience produces parameter updates that govern later training decisions; actual execution, causal benefit and production loading remain unestablished.); Initial backbone provenance and alternative serving/checkpoint updates: not-determinable (The supplied model identity and update route do not establish the initial training inputs or separate deployment replacement mechanism. Provider opacity warrants neither yes nor no for these alternatives.)	partial coverage (The documented consumers establish knowledge, ranking, routing and learning paths. Exact rendering, opaque policy processing and deployment admission controls prevent complete authority enumeration. Static tool prompts and schemas are excluded rather than counted as acquired instruction or validation memory.); Retrieved LTM prose and retained conversational prose at the task policy: known — knowledge [claimed] (This is contextual interpretation, not evidenced instruction-priority elevation, enforced compliance or content acceptance.); LTM embeddings at cosine retrieval selector: known — ranking [claimed] (The retained feature representation influences relative retrieval priority; similarity establishes no truth warrant.); Retained message text at Filter similarity selector: known — ranking [claimed] (The retained text is scored for removal eligibility; criterion polarity and encoder opacity limit reliability, not the stated score influence.); Access metadata at search restriction and retrieved ID at Update/Delete targeting: partial — routing [claimed] (Filtering and target addressing are specified. Authentication, criterion enforcement and actual returned field layout are uninspected; source labels establish no trust elevation.); Experience and reward-derived supervision at optimizer: known — learning [claimed] (The update consumer differs from the downstream policy consumer; scalar feedback is not a memory-entry admission gate.); Updated policy weights at subsequent action selection: partial — routing [claimed] (The specified output distribution selects operations and language; opaque internal content and exact checkpoint loading prevent a complete enumeration of downstream forces.)	AAS-2026-10-12-2601-01885v1-32274e5bddc1-01
aide2	2	Use-derived summaries, error signatures and reviewer results; retained score/strategy access metadata; previous proposal/grade and evaluation-artifact history; accepted harness code and its embedded instructions. Include unresolved debug/reviewer delivery and AIDE47/AIDEhuman bounded-history alternatives. Raw code/output is included only as accumulated later-consumed history or input to derived-memory updates. Exclude plain immediate execution state, static task inputs, fixed provider weights/internals, external benchmark implementation and wider production product.	not-determinable coverage (No controlled substrate is established for the included inventory. Fixed provider weights and static task inputs are excluded; encoding and delivery do not resolve storage.); AIDE85 summaries and error signatures: not-determinable (Prompt/context state is documented, but physical retention substrate is not. Prompt is a consumption channel, not evidence of files, in-memory storage or a prompt registry.); Reviewer feedback, scores and accumulated strategy metadata: not-determinable (Candidate-associated state and grading artifacts do not identify a controlled storage substrate.); Accepted harness, embedded instructions and proposal/grade archive: not-determinable (A codebase does not establish Git/repo storage or a filesystem implementation; archive and reload are unknown.); Other role and agent context routes: debug/reviewer payloads, AIDE47/AIDEhuman bounded-history alternatives, and outer evaluation artifacts: not-determinable (Full payloads, production rules and later delivery contracts are unavailable. AIDE85 witnesses do not classify these included alternatives.)	partial coverage (Automatic mechanisms are documented locally; unknown summary creation, initialization and alternative artifacts prevent complete agency coverage.); Summary inclusion in later prompts: known — automatic [claimed] (Known delivery admission does not resolve summary creation control.); Summary creation admission: not-determinable (Producer and production control are not stated; automatic inclusion alone does not classify every creation mechanism.); Signature extraction/deduplication and gated admission: known — automatic [claimed] (Human experiment setup is not per-signature approval.); Reviewer score/feedback extraction and strategy state updates: known — automatic [claimed] (Model authorship and physical I/O are distinct from admission.); Harness successor and embedded-instruction admission: known — automatic [claimed] (Experiment initiation does not make each successor manual.); Initial content and opaque artifact admission mechanisms: not-determinable (Initial task/harness content control and alternative artifact-production mechanisms are not specified; no manual finding follows from designer participation.)	partial coverage (Documented automatic supply supports push. Unspecified exploration/reload does not establish pull, absence of pull, or complete direction coverage.); AIDE85 summary and signature supply: known — push [claimed] (Root/recent selection and conditional error injection are independent supply operations.); AIDE0 accumulated code/output at later generation: known — push [claimed] (This read concerns prior retained trace input to later calls, not arbitrary current state.); Outer history and incumbent reuse: partial — push [claimed] (Artifact exploration includes unspecified requests and delivery; no complete push-only boundary is established.); Reviewer feedback delivery, artifact exploration and other agent prompts: not-determinable (Full payloads, production rules and later delivery contracts are unavailable. AIDE85 witnesses do not classify these included alternatives.)	partial coverage (Coarse supply is supported locally. No embedding, lexical query, inferred judgment or identifier matching is established for the recorded pushes; unresolved routes remain included.); Root/recent summary push selector: known — coarse [claimed] (Trigger is prompt assembly; selector input is root/recent membership; selected retained part is the compact history account.); Failure-signature push selector: known — coarse [claimed] (Trigger is prompt assembly at bug rate at least 15%; input is recent failed outputs; selected part is at most three final lines. Text dedup is not lexical query retrieval.); Full-history automatic supply: known — coarse [claimed] (Trigger is drafting/improvement; prior-history availability selects material without targeting a query.); Outer proposal/grade supply and incumbent choice: partial — coarse [claimed] (Grade comparison and availability are coarse selection; no documented identity-match delivery, exact budget or complete artifact selector.); Other payload selection and artifact exploration: not-determinable (Full payloads, production rules and later delivery contracts are unavailable. AIDE85 witnesses do not classify these included alternatives.)	partial coverage (Positive claimed automatic trace-fed writes coexist with unresolved mechanisms. No bounded absence record supports a local or system-wide no.); Automatic signature update to later generation context: known — yes [claimed] (Qualifies within a task run; does not require cross-task memory or observed improvement.); Summary creation and later automatic delivery: not-determinable (Earlier code/output history feeds a retained summary consumed by later draft/improve calls, but production control is unspecified. Automatic inclusion establishes delivery admission, not an automatic trace-fed write producing the summary; neither yes nor no is established.); Reviewer extraction and later score selection: partial — yes [claimed] (Score-selection path is documented; exact feedback delivery and persistence are unresolved.); Evaluation-guided harness rewrite and private admission: known — yes [claimed] (Durable behavior-shaping code, not model-weight training. Only claimed route evidence is available.); Other bounded-history alternatives and opaque artifact updates: not-determinable (Full payloads, production rules and later delivery contracts are unavailable. AIDE85 witnesses do not classify these included alternatives.)	partial coverage (Authority is consumer-specific. Update learning and downstream knowledge/instruction are separate. Unknown full payloads prevent an exhaustive force inventory.); Derived summaries and signatures consumed by later draft/improve: known — knowledge [claimed] (The downstream advisory consumer is distinct from the transformation consumer.); Raw execution history consumed by summary/signature and reviewer updates: known — learning [claimed] (Learning names the update consumer, not demonstrated improved task capacity.); Feedback text at subsequent generation: partial — knowledge [claimed] (Exact delivery after filtering/summary is unresolved; no complete feedback-consumption guarantee follows.); Scores and strategy labels at fixed inner and outer selectors: known — ranking [claimed], routing [claimed] (These are explicit selector inputs, not new executable policies themselves.); Previous proposals, grades and evaluation artifacts at outer rewriting: partial — knowledge [claimed], learning [claimed] (The documented update route is positive; exact artifact contents and rationale delivery remain unresolved.); Accepted executable search/context rules at later executions: partial — routing [claimed], ranking [claimed] (Exact code decomposition is unavailable; this does not classify every retained control.); Accepted robustness rules and reminder at their consumers: known — instruction [claimed], validation [claimed], enforcement [claimed] (The criterion is a length proxy, not semantic correctness or outer harness admission.); Other role and agent context routes: debug/reviewer payloads, AIDE47/AIDEhuman bounded-history alternatives, and outer evaluation artifacts: not-determinable (Full payloads, production rules and later delivery contracts are unavailable. AIDE85 witnesses do not classify these included alternatives.)	AAS-2026-10-14-2609-26457-6b3bcd4791cf-01
arsumbris	2	Changed repository knowledge and types, indexed opaque assets, their live graph access structure, device metadata reused by later merges/launches, declared session ledger and saved pane/viewer state, governed write metadata, and unresolved accumulated outputs of ingestion/computed or changed guidance. Includes their admission, maintenance and later-consumer routes. Excludes unchanged shipped instructions/profile policy, executable platform code as such, ordinary transient UI coordination and external harness/provider weights; absent sibling implementations limit all opaque alternatives rather than establishing absence.	partial coverage (Documented file/graph substrates and implemented device files are resolved locally; ledger backing, viewer state, pane storage and computed/ingested alternatives remain unresolved.); Changed repository knowledge, symbolic fields, type definitions and indexed assets: known — files [claimed], repo [claimed] (The durable substrate is Git-tracked files, including opaque asset payloads; their encoding does not change storage.); File-derived live graph and access index: known — graph [claimed], in-memory [claimed] (The documented graph is an in-memory access structure, not a separately established durable database.); Device access metadata: known — files [wired] (Producer persistence is inspected; downstream host/engine storage contracts add no other substrate.); Session ledger backing: not-determinable (Appends are declared; no physical storage or durability contract identifies a controlled substrate.); Shared pane configuration and viewer-local state: not-determinable (saveConfig promises persistence but exact storage is unspecified; viewStore lifetime and backing are opaque.); Computed context, source-note ingestion and governed stamp outputs: not-determinable (Their existence or output labels do not resolve all retained payloads or backing stores. Source notes may join repository files, but their complete persistence chain is uninspected.)	partial coverage (Software admission and explicit human content editing are distinguished from workflow initiation; ask, generic filesystem callers, initialization and opaque package writes remain unresolved.); Mediator admission and model-supplied overwrite on governed tool path: known — automatic [claimed] (The model proposes replacement text and software permits/denies it without an established per-content operator decision; ask is assessed separately.); Direct human file editing branch: known — manual [claimed] (This branch is limited to a human supplying/editing text, not every projection-generated change or agent launch.); Change-driven graph update: known — automatic [claimed] (Code indexes admitted file changes without a separate human decision for each indexed part.); Device registry discovery/merge admission: known — automatic [wired] (--write authorizes the run, not each discovered mapping; entry selection is deterministic.); Session observation append: known — automatic [claimed] (The observe stage appends pipeline events rather than awaiting content approval.); ask approval, arbitrary projection/filesystem writers, initial content, saved/viewer state and ingest/stamp/context transforms: not-determinable (Generic caller identity or a save API does not resolve per-content human control; the ask interface and opaque transformation admissions are absent from the records.)	partial coverage (File and registry reuse are established locally. Subscription delivery is distinct from requested reads; guidance, ledger, saved/viewer state and opaque ingestion retain unresolved later-consumer coverage.); Later explicit file/graph request: known — pull [claimed] (A repository/path request followed by returned/rendered content is pull, not identifier push.); Subsequent seeder configuration load/merge: known — pull [wired] (The seeder rereads retained configuration as an input to another invocation; the load requests that material.); Change subscription delivery: known — push [claimed] (After subscription, the declared service streams changes without a new content request for each change; subscription setup remains a request.); Ledger replay, saved pane/viewer reload, computed/changed guidance and source-note later consumption: not-determinable (A retained object name or static launch materialization does not establish a later accumulated-memory consumer. Exact guidance-selected parts and reload/replay routes are unspecified.)	not-determinable coverage (The subscription push is documented, but accepted records do not identify its actual selector input and delivered retained parts sufficiently to classify a signal. A pull witness cannot make unresolved push alternatives inapplicable.); Subscription selector and delivered changed parts: not-determinable (Subscriptions stream changes; filter granularity, identity matching and exact delivered retained content are not specified. Mounted-domain scope alone does not establish a coarse selection operation.); Computed/changed guidance and opaque later delivery alternatives: not-determinable (Graph-walk seeds are targeting inputs, not an inspected selector or identified payload. Triggered skill matching and later accumulated-content delivery remain unresolved.); Explicit file and configuration requests: inapplicable (Read-back signal classifies push selection; these local operations fulfill requests and introduce no push selector.)	not-determinable coverage (No accepted bounded absence or qualifying positive chain resolves the included opaque alternatives. Event recording alone is insufficient; do not turn an unestablished learning route into a system-wide no.); Session ledger to later behavior-shaping artifact: not-determinable (Raw events are recorded, but no automatic derived artifact and later consumer are established.); Computed injection, source-note transforms and changed guidance: not-determinable (Inputs, durable derived result and later invocation chain are unspecified. Session-start timing alone does not establish trace feeding.); File indexing and device configuration merge: inapplicable (These known transformations index files or merge configuration, rather than automatically derive behavior-shaping artifacts from activity traces.); Initial knowledge, host/viewer state and stamped output alternatives: not-determinable (No qualifying trace-fed update and later consumer are recorded; incomplete provenance prevents a bounded negative.)	partial coverage (Each positive names a consumer and force; opaque payload consumers and later context delivery prevent a complete union. Static profile/tool policy is not counted as accumulated memory.); Changed knowledge prose at later graph/UI/agent readers: known — knowledge [claimed] (Retrieval/rendering supplies content without warranting its truth or model compliance.); Retained type constraints at engine diagnostics: known — validation [claimed] (Validation assesses structural conformity; diagnostics do not veto save/read.); Type and runtime metadata at extension discovery/load: known — routing [claimed], validation [claimed], enforcement [claimed] (These effects concern subsequently used definitions, not the mere presence of static SDK code.); Device registry rows at subsequent seeder merge: known — routing [wired] (Preserved mappings affect later discovery/merge output; this consumer is implemented independently of the documented launcher.); Device locations/tool paths and saved composition at documented host/engine consumers: known — routing [claimed] (Launch/resolution and UI arrangement are documentary consumers, not inspected execution.); Ledger, stamp payloads, viewer store, source notes and computed or changed guidance: not-determinable (Ledger replay is not established; metadata interpretation and viewer lifetime are unknown. Guidance instruction intent is documented, but no changed-guidance-to-later-invocation chain establishes accumulated instruction memory. No learning or ranking authority follows from mere retention.)	AAS-2026-10-15-arsumbris-a9be6bac297f-01
claude-code-dynamic-workflows	2	Use-accumulated run/reusable workflow scripts and their embedded tasks/criteria, saved worker results including research claims, populated prompt-prefix caches, use-changed consent/guidelines/opt-in settings, and their admission, maintenance and later-consumer paths. Includes unresolved worker-context, usage-wait checkpoint and progress/tool-trace retention alternatives named in accepted records. Excludes shipped static bundled/plugin instructions as such, ordinary same-run variables and current-output filtering as such, general external project-file versioning, provider model parameters/training and general host compaction not established within the workflow capture. Plugin copies of task-generated or revised reusable scripts remain included; their admission details are unresolved.	partial coverage (The documented inventory includes run/reusable scripts, saved results, populated prefixes and use-changed settings. Files and cloud service objects are supported; provider cache infrastructure, administrative consent storage and alternative checkpoints remain unresolved.); Run and saved project/personal/plugin script artifacts: known — files [claimed] (Documented .js file locations classify the operative artifact, not every loader implementation.); Locally retained agent results: known — files [claimed] (The record names local session-directory storage; serialization is separate.); Cloud retained results: known — service-object [claimed] (The documented conversation-history service object is classified without claiming its physical backing store.); Populated prompt-prefix cache: not-determinable (Opaque provider/request-cache infrastructure does not identify a controlled physical or logical storage substrate; no vector, kv or model-weights value is inferred.); Use-changed consent, guideline and opt-in settings: partial — files [claimed], in-memory [claimed] (User settings files and session state are described; exact named-consent and managed administrative backing remain unknown.); Included worker context, usage-wait checkpoint and progress/tool-trace retention alternatives: not-determinable (The records do not establish the retained payload or later reload route for these alternatives; Storage cannot be completely classified. A partial tool-result display is not the consumed payload.)	partial coverage (Admission mechanisms are separated from authorship and I/O. Positive manual and automatic routes are supported, but plugin admission, managed configuration and alternative checkpoints are unresolved.); Runtime persistence of generated run script and completed worker outputs: known — automatic [claimed] (Starting or approving execution is not approval of each retained content item.); Operator save and direct human edits of reusable scripts: known — manual [claimed] (A human selects the run/location to save or supplies edited content.); Claude-authored requested script edits: partial — automatic [claimed] (Requesting edits alone does not document human approval of each resulting content revision; direct-file edit control is not fully specified.); Prefix cache population: known — automatic [claimed] (Settings controlling TTL do not supply individual cached content.); Explicit human consent and preference/configuration choices: known — manual [claimed] (This unit covers specified operator choices rather than every administrative pathway.); Plugin installation/admission and administrative settings backing: not-determinable (Separate installation and administrative admission details are outside the recorded contract; generic tooling identity does not resolve content control.); Included worker context, usage-wait checkpoint and progress/tool-trace retention alternatives: not-determinable (The records do not establish the retained payload or later reload route for these alternatives; Write agency cannot be completely classified. A partial tool-result display is not the consumed payload.)	partial coverage (Named loading, inspection and relaunch requests coexist with automatic selectors supplying retained parts. Full worker context and usage-wait reload routes remain unresolved.); Operator/Claude request for saved script invocation or run-file inspection: known — pull [claimed] (The requested artifact is returned or loaded; command naming alone does not establish identifier-based push.); Replay request and automatic completed-prefix substitution: known — pull [claimed] (Completed-prefix selection and substitution fulfill the requested recovery of retained run results. The request need not name each value; automatic compatibility selection does not establish a separate push operation.); Automatic matching-prefix reuse on later worker request: known — push [claimed] (The worker initiates its task request; the cache selector supplies retained prefix processing without a request for sibling memory.); Persistent control and guideline supply to later launch/authoring: known — push [claimed] (Settings are automatically read in the documented launch/startup/next-prompt channels.); Included worker context, usage-wait checkpoint and progress/tool-trace retention alternatives: not-determinable (The records do not establish the retained payload or later reload route for these alternatives; Read-back direction cannot be completely classified. A partial tool-result display is not the consumed payload.)	partial coverage (Actual push selectors are classified independently of upstream requests. No embedding or semantic retrieval signal is supported; opaque checkpoint and inherited-context selectors remain unresolved.); Requested completed-result recovery during replay: inapplicable (Replay is pull: prior order, prompt equality and failure state select reusable results within the requested recovery. Read-back signal classifies push selection, so it is inapplicable to this completely assessed requested-return unit, not to the whole partial axis.); Automatic prefix-cache reuse: known — identifier [claimed] (Trigger is a later matching worker first request or staggered fan-out release. Model, effort, agent type, tools, schema and working-directory identity select the cached prefix.); Named/project consent control delivery at launch: known — identifier [claimed] (Trigger is named launch evaluation. Workflow/project scope identities select applicable retained consent.); Global/user/session settings and next-prompt size advice: known — coarse [claimed] (Startup, launch or next-authoring prompt triggers availability of applicable global/session controls and guideline.); Included worker context, usage-wait checkpoint and progress/tool-trace retention alternatives: not-determinable (The records do not establish the retained payload or later reload route for these alternatives; Push selection signal cannot be completely classified. A partial tool-result display is not the consumed payload.)	not-determinable coverage (No accepted record establishes a qualifying automatic trace-fed durable behavior-shaping update and its later consumer, or a bounded absence across all included routes. Lack of demonstrated improvement is not the reason for withholding yes.); Automatic saved task-output production and replay: not-determinable (Durable outputs and a later script consumer are documented, but complete original inputs and a trace-fed behavior-shaping derivation are not established. Task products are not automatically traces or learned methods.); Generated/saved/revised workflow scripts: not-determinable (Durable orchestration can affect later behavior; the records do not establish automatic extraction or revision from activity traces. Human-selected saves do not alone prove a bounded no.); Automatic populated prompt cache: not-determinable (Reuse of tools/system prefixes is not recorded extraction of experience; opaque provider population details prevent a bounded absence claim.); Use-changed consent and settings: not-determinable (Explicit human choices are supported, not an automatic trace-fed update; complete alternative administrative updates are unresolved.); Included worker context, usage-wait checkpoint and progress/tool-trace retention alternatives: not-determinable (The records do not establish the retained payload or later reload route for these alternatives; Qualifying trace learning cannot be completely classified. A partial tool-result display is not the consumed payload.)	partial coverage (Authority is assigned at actual retained-part consumers. Opaque cache representation, worker context and administrative details prevent an exhaustive authority union; no learning or ranking force follows merely from retention.); Retained script control flow at workflow runtime: known — routing [claimed] (The script selects workers, branches and returned values.); Embedded retained task text at worker model: known — instruction [claimed] (Prompt delivery supplies task directions, not guaranteed compliance.); Retained script schema criteria at validator/runtime admission: known — validation [claimed], enforcement [claimed] (Only embedded reusable criteria are included; ephemeral schema state is excluded.); Saved result content at replayed script and downstream reader: partial — routing [claimed], knowledge [claimed] (The documented generic and example consumers establish routing/context; arbitrary saved/plugin consumer effects are not enumerated.); Replay compatibility state at result tracker: known — routing [claimed] (Prior start order, completion/failure and prompt matching determine which saved values remain reusable.); Cached delivered tools/system prefix at matching worker: partial — instruction [claimed] (System guidance is established; complete tool definitions and provider payload effects are not exposed.); Retained consent/disable/opt-in control at launch evaluator: known — enforcement [claimed], routing [claimed] (Controls govern workflow permission and launch surfaces, not factual acceptance.); Retained size guideline at foreground script author: known — instruction [claimed] (Advice remains advisory even when persisted.); Included worker context, usage-wait checkpoint and progress/tool-trace retention alternatives: not-determinable (The records do not establish the retained payload or later reload route for these alternatives; Behavioral authority cannot be completely classified. A partial tool-result display is not the consumed payload.)	AAS-2026-10-14-workflows-f9c2f80a3a94-01
context-constitution	2	Agent memory accumulated or revised through use: system and external memory blocks, their Git-tracked MemFS projection, prose and opaque skill assets, curated discovery cues and automatic access metadata, retained conversation history, recursive compaction summaries, foreground and delegated maintenance, and later prompt or requested retrieval consumers. Automatic metadata is included as an access structure although its own persistence is unspecified. Environment context is included only as an alternate discovery/input surface, not attributed to agent-owned memory. Excluded are the static published Constitution, unmodified environment instructions, model weights and internal training, and publication-hygiene state. This is the declared policy artifact's memory loop, not a classification of an inspected Letta implementation.	partial coverage (Inventory covers declared retained objects and access structures; opaque backend and summary persistence prevent complete substrate classification.); Authoritative memory blocks underlying system and external memory: partial — service-object [claimed] (Logical block storage is described; backend substrate and initial representation are unspecified.); Local MemFS projection, external files and skill folders: known — files [claimed], repo [claimed] (The local file/versioning boundary is classifiable without inferring synchronization implementation.); Compaction summary in the later inference buffer: partial — in-memory [claimed] (Buffer placement is specified; backing persistence and prior-summary storage are not.); Raw message database and automatic metadata backing state: not-determinable (A database and automatic indexing are named without relational, vector or other backend facts; metadata persistence is unresolved.)	partial coverage (Manual and automatic mechanisms are both declared; unresolved initial, direct-tool and delegated gates prevent a universal admission classification.); Agent-selected foreground memory and index edits: known — automatic [claimed] (File I/O and push are propagation stages, not evidence of human semantic approval.); Explicit human edits of projected memory files: known — manual [claimed] (This separately allowed operator-content decision does not make all other writes manual.); Automatic raw-message storage, metadata assembly and size-triggered compaction: known — automatic [claimed] (Host automation is explicit at design level; summary quality and implementation are separate unknowns.); Initial content, direct bespoke memory tools, delegated write admission and desired external fact checkpoint: not-determinable (Initial admission, alternate-tool control, delegate veto/merge ownership and fact-extraction producer are unspecified; agent proposal or workflow launch cannot resolve per-content approval.)	partial coverage (Both requested content and unsolicited prompt/summary supply are declared. Unresolved delegate and asset paths prevent complete direction coverage.); System memory and discovery metadata supplied in later prompts: known — push [claimed] (Unsolicited baseline supply is distinct from subsequent agent-requested expansion.); External files/skills and stored history requested by the foreground agent: known — pull [claimed] (Relevance reasoning precedes an explicit request in the disclosed workflow.); Compaction summary carried into later inference: known — push [claimed] (This within-conversation continuation is read-back because a later inference receives retained derived context.); Delegate retained inputs/returns, non-prose skill consumers and desired extracted-fact checkpoint: not-determinable (Exact delegated context construction and returns, asset consumption and a particular fact checkpoint's later delivery are unspecified. Shared memory alone does not resolve these operations.)	partial coverage (Coarse automatic supply is supported at design level; no targeted identity, lexical, embedding or judgment push selector is established for full retained content. Opaque alternatives remain included.); Baseline system memory and automatic discovery-metadata push: known — coarse [claimed] (The described push selects baseline classes by placement/availability. Agent relevance judgments subsequently request full files and do not classify this push as inferred-judgment.); Size-triggered summary push: partial — coarse [claimed] (Threshold/age selection is described; internal summarizer content selection, retained tail and exact thresholds are not, preventing a complete signal classification.); Agent-requested file/skill/history delivery: inapplicable (This unit is explicit pull, not automatic push. Paths identify requested objects but establish no identifier-based push.); Delegated/opaque asset supply and desired fact-checkpoint delivery: not-determinable (Unresolved delivery direction and selectors prevent signal findings; a known pull route cannot make these alternatives inapplicable.)	partial coverage (Positive design routes coexist with unresolved admissions and transformations. No system-wide negative or exercised learning conclusion is inferred.); Automatic conversation compaction to retained continuation summary: known — yes [claimed] (This meets the trace-fed write definition by doctrine; no novelty, observed operation or demonstrated improvement is implied.); Model-selected durable experience-based foreground edits: partial — yes [claimed] (The prescribed trace-fed branch qualifies; concrete inputs and actual writes are unspecified, and this does not classify every edit as trace-fed.); Reflection admission, other delegated maintenance, initial/alternate-tool writes, non-prose assets and desired extracted-fact checkpoint: not-determinable (Reflection reviews conversations to change memory, but final admission control and outputs are unspecified; other branches lack a complete automatic trace-to-artifact-to-consumer chain.); Raw history storage, automatic access metadata and explicit human content edits considered alone: inapplicable (Plain trace retention is not a derived behavior-shaping update; counts/tree metadata alone establish no such artifact, and explicit human edits are not automatic admission. These local exclusions do not negate qualifying routes.)	partial coverage (Inventory separates inference guidance, routing and update-input authority. Opaque consumers remain unresolved; the file push gate is not itself retained memory with enforcement authority.); System memory at later model inference: known — instruction [claimed], knowledge [claimed] (Both described content effects are model-interpreted, not enforced invariants; actual activation remains unobserved.); External prose and raw history at requested model consumption; summary at subsequent inference: known — knowledge [claimed] (Procedural advice here is not automatically standing system guidance or executable enforcement.); Curated and automatic discovery cues at foreground retrieval decisions: known — routing [claimed] (Cue-based input selection is described, but no relative scoring or retained veto criterion is established.); Retained conversational experience at compaction and Reflection update consumers: known — learning [claimed] (Learning authority denotes the update input's role, not demonstrated improved capacity or accepted knowledge.); Opaque skill-asset consumers and particular delegated memory outputs: not-determinable (No actual asset interpreter or delegate payload determines its force; capability names do not establish execution or enforcement.)	AAS-2026-10-15-context-constitution-7a352d7408f6-01
From Agent Memory to Portable Skills	2	Included: scoped retained conversation/decision text, tool-call structure and opaque payloads, ontology/reference identities, derived graph and prose skill products, grounding, review and version metadata, and their documented admission/maintenance/later-consumer routes. Excluded: upstream ingestion implementation, enclosing host execution/context policy, transient Redis job state and synthesis buffers, and model weights with no established through-use memory update. Unknown backing provenance and payload forms, patch release and consumer selection remain included limits, not inferred absence.	known coverage (Known substrate classification for the documented retained inventory, not verified deployment. No retained filesystem copy is established by the export name. Redis jobs and transient synthesis buffers are not additional accumulated-memory objects.); Scoped conversation text, tool records including opaque payloads, and ontology entities: known — graph [claimed] (All three backing-input classes are described as workspace Neo4j graph material.); Derived skill graph, grounding and version relations, and packaged skill including prose fallback: known — graph [claimed] (The graph and packaged artifact are persisted in Neo4j; SKILL.md names an export format, not an independently evidenced filesystem store.)	partial coverage (Automatic construction and repair coexist with manual initial publication; they do not settle upstream acquisition or alternate revision controls.); Software/model construction and gate admission to pending persistence: known — automatic [claimed] (Scope initiation is not a human decision approving each generated item.); Human initial publication approval or rejection: known — manual [claimed] (The reviewer explicitly decides on the retained product, independently of who started the job.); Salvageable single-step grounding patch admission: known — automatic [claimed] (Software admits surviving-grounding changes or refuses unsupported repair; later patch publication is separate.); Backing-content initial admission, manual replacements and repaired-patch publication: not-determinable (Initial acquisition control, direct edits/replacements and mandatory human patch approval are not specified.)	partial coverage (Supported requested routes establish pull, not system-wide absence of push. Automatic delivery in fulfillment of a request is not counted again as push.); Scope-requested retained-evidence retrieval and bounded synthesis delivery: known — pull [claimed] (The distiller fulfills a scope request; snippet filtering within that requested chain is not an additional unsolicited delivery.); Reviewer evidence access and MCP/REST published-package fetching: known — pull [claimed] (These documented request/load routes have later consumer roles, not observed deployments.); Host skill loading, governance scheduling, repair invocation and downstream version/flag delivery alternatives: not-determinable (No independent unsolicited selector is described; host loading policy and governance/repair initiation are unresolved. This prevents a complete pull-only conclusion.)	not-determinable coverage (Known pull routes cannot make the unresolved loading/maintenance alternatives inapplicable. No controlled push-selection signal is supported.); Potential unsolicited host loading and governance/version delivery: not-determinable (Trigger, selector input and selected retained parts for a push route are not established. Scope IDs and evidence links on requested retrieval do not establish identifier-based push.)	partial coverage (A claimed yes survives incomplete inventory coverage. This is not measured learning, parameter updating, autonomous publication or improvement of the distiller itself.); Automatic scoped trace distillation into durable procedure graph/package for later review and host guidance: known — yes [claimed] (The documented raw-to-derived-to-later-consumer chain meets the write-route definition, including the prose alternative; human publication veto does not erase automatic pending-artifact construction.); Repair-derived behavior changes beyond grounding, alternate writers and upstream processing: not-determinable (Grounding revision is documented, but original trace consumption and downstream behavior-shaping effects of every alternative update are not specified. No evidenced-absence record warrants a local no.)	partial coverage (Consumer-specific forces coexist without an implication from instructional delivery to enforcement or observed benefit. Metadata whose consequential consumer is unresolved remains unclassified.); Retained conversations and recorded execution outcomes at the distillation update consumer: known — knowledge [claimed], learning [claimed], ranking [claimed] (This unit is the described update use, distinct from downstream instruction use.); Graph identity and grounding at reference validation and repair consumers: known — validation [claimed], enforcement [claimed] (The classified force concerns retained identities and survivor relations, not truth of synthesized claims.); Retained expected status and recorded current ToolCall status at governance comparison: known — validation [claimed] (The comparison checks status consistency; its downstream blocking effect is not established.); Pending skill and linked backing evidence at human review: known — knowledge [claimed] (The reviewer consults steps and evidence before the publication decision.); Published procedure and references at later host-model consumers: known — instruction [claimed], knowledge [claimed] (Natural-language guidance does not grant tools or enforce compliance.); Review state, versions and drift flags at downstream delivery/selection consumers: not-determinable (Publication is the documented condition, but actual state enforcement, flag consultation, version resolution and withdrawal are unspecified. No routing or enforcement value is inferred for these opaque consumer mechanisms.)	AAS-2026-10-18-from-agent-memory-to-portable-skills-4a509c7732b1-01
incremental-self-improvement	2	Documented single-life memory: live policy distributions; stack oldP and evaluation/access metadata; work inputs, GetP snapshots, registers and arbitrary retained integers; external variable values and retained navigation state; their automatic write, retention, restoration and later-consumer routes. Documentary alternatives include prior-code reuse, direct teacher protection and circular history overwrite only at their claimed/proposed strength. Excludes shipped static primitive/evaluator definitions, transient comparison predicates, base-design sampled code with no later reuse, ordinary immediate action effects without a later retained-state consumer, referenced systems, proposed neural variants, and unsupported restart durability. Navigation representation and arbitrary history semantics remain included but unresolved.	partial coverage (All documented retained arrays and stack parts are classified. Navigation state substrate remains unresolved; files and restart persistence are not inferred from a paper capture.); Internal work cells: inputs, GetP snapshots, registers and remaining integers: known — in-memory [claimed] (The documentary inventory places all these parts in the same in-life addressable array; semantic uncertainty does not obscure its substrate.); Live policy table: known — in-memory [claimed] (The documented distributions are runtime numerical arrays.); Stack distribution payload oldP and access/evaluation metadata: known — in-memory [claimed] (Both parts occupy the finite in-life stack; payload encoding is not a separate storage technology.); External variable array: known — in-memory [claimed] (The task describes an integer array, not a service or disk store.); Proposed reuse of previous sampled program-cell contents: known — in-memory [claimed] (Included only as a documentary alternative, not as activated experimental memory.); Retained navigation-environment position between actions: not-determinable (The documented maze retains and changes task state, but these records do not specify its operative storage representation. No substrate value can be assigned to that part.)	partial coverage (Documented base mechanisms are automatic. Proposed teacher protection retains its claimed manual control locally, without upgrading it to wired or treating programmer task design as per-content admission.); Birth initialization and ordinary work/input writes: known — automatic [claimed] (The records describe fixed initialization and software/task writes without per-content human approval.); GetP snapshot admission: known — automatic [claimed] (A selected primitive writes a snapshot without a human content decision.); Base Pi update, stack saving and keep/restore: known — automatic [claimed] (Proposal, physical copying and subsequent retention are different operations, all software-controlled in this design.); External variable Write and periodic zero replacement: known — automatic [claimed] (Starting the experiment is not approval of each value.); Proposed sampled-code reuse admission: known — automatic [claimed] (The interpreter samples code for later reuse in the proposed variant.); Proposed direct teacher protection of selected policy changes: known — manual [claimed] (This is a separate declared override, not evidence of a human gate in the base loop.); Navigation retained-state write/admission details: not-determinable (Task actions and resets are described, but the records do not fully identify the retained payload and all admission paths; this prevents a complete write-agency inventory for that part.)	partial coverage (Pull and push coexist at different operations. Known requested routes do not resolve opaque semantic alternatives or make the entire selection axis inapplicable.); Explicit work reads, GetP and external Read: known — pull [claimed] (Selected primitives request particular retained values and receive them; delivery fulfilling that request is not another push.); Live policy supplied to automatic later sampling: known — push [claimed] (At the next instruction visit the interpreter consumes retained Pi without a content retrieval request from an executing primitive.); Surviving stack history supplied to later evaluation/restoration: known — push [claimed] (Completed cycles with checking enabled trigger automatic selection of retained baselines and recovery payloads.); External variables supplied to periodic scoring: known — push [claimed] (The driver automatically scores retained array values independently of explicit Read.); Proposed reused prior code at repeated program address: known — push [claimed] (Only the declared alternative reuses code; base sampled contents are ordinary current-run realization.); Additional later consumption of arbitrary encoded work history and navigation state: not-determinable (The records afford event representation and describe task state but do not enumerate all retained history encodings or their later consumer routes.)	partial coverage (Each positive signal names a concrete automatic selector and selected retained part. Address-based primitive requests remain pull; unknown alternatives prevent complete signal coverage.); Policy push at each instruction visit: known — identifier [claimed] (Trigger: next interpreter visit. Selector input: IP/program address. Selected part: its associated Pi and subsequent argument distributions.); Stack push on enabled credit assignment: known — coarse [claimed], identifier [claimed] (Trigger: completed instruction outside an unfinished modifier. Newest surviving entry is chosen positionally; first/address fields identify grouped history and restoration targets.); Variable-array push to periodic scorer: known — coarse [claimed] (Trigger: 1,000-step payoff event. Selector input: interval timing and full array membership. Selected part: all current variable values.); Proposed prior-instruction push: known — identifier [claimed] (Trigger: repeat visit in the run. Selector input: program-cell identity and visit history. Selected part: its previous instruction.); Explicit primitive work/GetP/Read requests: inapplicable (These are pull operations, so their requested addresses are not push signals.); Unresolved work-history and navigation consumer alternatives: not-determinable (No complete selector inventory for these included parts is recorded. Known pull routes cannot make their unknown alternatives inapplicable.)	partial coverage (Yes is established at claimed design strength for the payoff-driven retention chain, independently of observed improvement. Other included transformations remain unresolved, not negative.); Historical payoff/time to automatic policy keep/restore to later sampling: known — yes [claimed] (The automatic transformation consumes experience measurements, retains or restores durable behavior-shaping policy within one life, and supplies later interpreter invocations. No raw transcript or new proposition is required.); Other modifier proposal inputs and arbitrary work/event-memory updates: not-determinable (The records do not establish the particular source lineage of modifier operands or actual arbitrary event encodings and later use. A local positive evaluator route does not determine every other possible trace-fed write.); Navigation-state and proposed reuse/teacher/circular alternatives as separate qualifying writes: not-determinable (Their memory/control effects are described, but no separately complete automatic trace-to-durable-behavior chain is recorded for each. No unsupported local no or system-wide absence is assigned.)	partial coverage (Distinct update and downstream consumers are classified separately. Arbitrary work-cell roles and retained navigation representation prevent complete authority coverage; software constraints themselves are static and outside this memory inventory.); Live Pi at later instruction/argument sampler: known — ranking [claimed], routing [claimed] (This consumer selects code, not a truth-apt proposition or standing natural-language instruction.); Current Pi at GetP and snapshots at later computations: known — knowledge [claimed] (GetP provides current-policy reference data; subsequent arithmetic use is afforded, not observed.); Input cells at selected later primitive operands: known — knowledge [claimed] (Encoded observations supply data, with conditional source warrant; individual exercised effects remain uninspected.); Registers and addressed work values at indirect addressing and branching: partial — routing [claimed] (Documented operand/control roles are positive; arbitrary cells may have additional unresolved roles.); Stack historical t/R at rate comparison and policy retention: known — knowledge [claimed], learning [claimed], validation [claimed], enforcement [claimed] (Measurements and baselines feed the fixed evaluator; this is not acceptance of an explanatory claim.); Stack address/first metadata at grouping and restoration: known — routing [claimed] (Fields identify restoration destinations and self-modification groups.); Stack oldP at restoration and later sampler: known — routing [claimed] (The payload replaces live policy rather than supplying an explanatory rationale.); External variables at explicit Read and periodic scorer: known — knowledge [claimed], learning [claimed] (Read makes retained values available to computation; scoring supplies the experience-responsive retention chain.); Proposed prior-code reuse at interpreter: known — instruction [claimed] (The proposal gives earlier code an effect absent from base-design resampling.); Navigation-environment state at later observations and payoff: not-determinable (The records identify task consequences but not the consumed retained representation and all its actual consumer paths; encoded input authority cannot classify that unseen payload.)	AAS-2026-10-14-fki198-94-pdf-90fb0a2a9508-01
interview-doc-agent	2	User-admitted raw experience, JD and persistent template replacements; initialized and maintained dossiers, profile and wiki index; generated products and their human consumption; described cloud persistence and opaque later knowledge-source delivery. Excludes unchanged shipped skill/templates/examples, external model weights and ordinary current-conversation state. Includes unresolved profile initialization and cloud intermediary representations where they affect this memory loop.	partial coverage (The named local and cloud stores are classified; opaque cloud access structures prevent complete coverage. Repository distribution does not make personal memory a repo substrate.); Local raw materials, dossiers, profile, wiki index, retained template replacements, JD and generated products: known — files [claimed] (These records enumerate the local retained-object inventory; successful persistence is unobserved.); Described cloud knowledge pages/Base records and product documents: known — service-object [claimed] (Classifies the named persistence objects, not the host's internal access structures.); Cloud knowledge-source intermediary retained state/access structures: uninspected (Host retrieval and transformations are uninspected; the page display cannot resolve internal storage or establish embedding-free delivery.)	partial coverage (Manual and automatic admission both exist in doctrine; unresolved initial fields and cloud internals prevent a complete agency inventory.); Human admission/edit/replacement of raw records, dossiers, JD and preferred templates; manual cloud copying: known — manual [claimed] (Human copying is a separate admission route, not evidence of factual review.); Starter creation, direct dossier/index updates and product admission in local or writable cloud mode: known — automatic [claimed] (Automatic names documented admission control; no executable or successful write is established.); Initial populated profile fields: not-determinable (The allocation of real biographical fields between user and model is unspecified; known subsequent index updating does not resolve that admission.); Cloud service-side derived/access-state admission: uninspected (No inspected selector or transformation implementation establishes control over any intermediary retained state.)	partial coverage (A known local pull route cannot settle the cloud direction or justify a system-wide pull-only conclusion.); Local foreground generation/detail reads and requested human consumption of saved products: known — pull [claimed] (Covers selected dossiers/profile/templates/JD, raw-detail fallback and whole-library alternative in the documented local route.); Bound cloud knowledge-source selection and delivery to a later invocation: uninspected (The tutorial does not expose whether a consumer pulls pages or the host independently pushes selected content.)	uninspected coverage (There is no supported positive push signal. Unresolved cloud delivery prevents axis-level inapplicability despite the pull-only local unit.); Completely described local requested read-back: inapplicable (This branch is pull-only; requested path names and model relevance matching do not establish a push signal.); Cloud knowledge-source selection, if push occurs: uninspected (Trigger, selector input and delivered retained parts are uninspected. No identifier, lexical, embedding, judgment or coarse push signal can be assigned.)	not-determinable coverage (Local negatives do not establish system-wide absence. Records support neither a qualifying positive trace-fed route nor a complete negative across heterogeneous inputs and opaque cloud alternatives.); Manual source/JD/template admission, blank initialization and generated products without later agent reuse: known — no [claimed] (The negative is bounded to these mechanisms and the shipped doctrine, not all downstream host use.); Automatic dossier/index compilation from heterogeneous human experience records: not-determinable (A durable derived artifact and later consumer are described, but chats/retrospectives are not identified as qualifying session/tool/event/trajectory traces. Absence of agent-execution compaction or method revision does not rule out other trace-fed context production; qualifying input origin is unresolved.); Cloud intermediary retained transformations: uninspected (Opaque host transformations expose no trace-fed write inventory or later derived-artifact consumer.)	partial coverage (Knowledge, instruction, ranking and routing have distinct consumers/effects. No blanket enforcement, validation or learning force follows from retention or a host permission check. Cloud consumption is unresolved.); Raw facts at consolidation/detail lookup; dossiers/profile facts at generation; products at human use: known — knowledge [claimed] (Names both update and downstream consumers without treating knowledge delivery as compliance.); Retained template replacements at the document-generation model: known — instruction [claimed] (Instructional force is requested, not independently enforced.); JD requirements and profile persona/project content at relevance selection: known — ranking [claimed] (Relevance priorities do not authenticate accomplishments or supply an answer oracle.); Wiki and profile locator entries at dossier/template lookup: known — routing [claimed] (The broader epistemic index container includes profile locators; locator authority is not factual endorsement.); Actual cloud-selected retained payload at later host consumption: uninspected (Knowledge-source intent is recorded, but opaque selection/transformation prevents complete classification of the consumed payload and its effects.)	AAS-2026-10-14-interview-doc-agent-c7597c33f23d-01
IsaacCLupus_mnemosyn_spec	2	Included: use-accumulated raw/draft/published/committed pages; normalized legacy vault and optional Git history; conversations, profile facts, rejection feedback, metadata and draft governance; graph access views, packs and optional proof outputs; their admission, maintenance and later-consumer routes. External Link/LLM-WIKI retained state and history, optional embedding/reranking state, response cache and provider-retained state remain included unresolved alternatives at their recorded interfaces. Excluded: shipped static templates/configuration/AGENTS instructions, configured model weights without an established use-fed update, and volatile current-request proxy queues. Proposed unified OS, draft extensions/studies and wired legacy bridge remain distinct evidence layers.	partial coverage (The inventory covers proposed retained content/access state and legacy publication. External backend state, cache and optional embedding storage prevent a complete substrate classification.); Legacy normalized Markdown and optional version history: known — files [wired], repo [wired] (Files are wired outputs; Git history is conditional on publication configuration and command success, not observed.); Proposed raw, draft, published and committed page bodies; rejection text and page metadata: known — files [claimed], sqlite [claimed] (Filesystem page bodies and SQLite page fields are the recorded proposed storage split.); Conversation, profile, link, job, audit, contradiction and draft governance rows: known — sqlite [claimed] (DDL and proposed additions specify SQLite, without establishing populated runtime state.); Reconstructed knowledge-graph access view: known — graph [claimed], in-memory [claimed] (Graph topology and RAM residence are separate supported properties of the proposed view.); Exported articles, index, navigation, manifest and optional proof serialization: known — files [claimed] (Ordinary pack files and optional draft proof output remain design-level storage.); External Link/LLM-WIKI state, optional embedding storage, response cache and provider-retained state: not-determinable (Payloads, backend storage and cache/embedding implementations are unresolved; neither service-object nor vector nor model-weights follows from endpoint names or model aliases.)	partial coverage (Known automatic publication and proposed automatic/human gates do not resolve generic caller control or every initialization and backend admission alternative.); Legacy per-page publication/copy/overwrite admission: known — automatic [wired] (The caller starts a workflow; fixed software chooses and writes each page without per-page approval.); Proposed conversation extraction and daily maintenance admission: partial — automatic [claimed] (Automatic extraction/promotion is specified; separate profile-row veto/control is not.); Proposed explicit human commitment of promoted memories: known — manual [claimed] (This is the distinct final approval mechanism, not the preceding automatic extraction.); Proposed configurable article and explicit-memory approval: known — manual [claimed], automatic [claimed] (Human and configured automatic alternatives remain distinct findings at the same claimed evidence layer.); Proposed feedback-guided writing, graph reconstruction and pack generation: known — automatic [claimed] (Software/model transformations are automatic even when an operator initiates recompilation or export.); Draft claim-level human verdict admission: known — manual [claimed] (The future governance gate expressly assigns content verdicts to humans.); Initial retained content, conversation logging control, external ingest/compile and Link HTTP versus CLI fallback; cache/embedding writes: not-determinable (Generic caller content and unknown backend admission do not establish human decisions or automatic admission; HTTP/fallback scope differs and neither resolves this missing fact.)	partial coverage (Request-led backend/file routes coexist with proposed automatic context supply inside compilation/query chains. Opaque backend assembly and unsettled profile/cache consumers prevent complete direction coverage.); Legacy client search/ask fulfillment: known — pull [wired] (A caller requests retained material; wrapping a requested return as MCP text does not add push.); Proposed graph get_context request and automatic model-input assembly: known — pull [claimed], push [claimed] (Two different operations occur: page-seeded retrieval is requested; the assembler supplies selected context to the answering model.); Proposed retained criticism delivered to a later writer: known — push [claimed] (The writer is supplied feedback by orchestration, not shown issuing a retrieval request.); Proposed retained source/related-page rendering into compilation: partial — push [claimed] (Automatic rendering supplies retained material, but the selector providing related pages is not shown.); Documented exported-file consumer: known — pull [afforded] (Directory import is an affordance; actual external reads are not inspected.); Conversation extraction/history, direct profile delivery, backend-internal context, cache/embedding reuse and optional proof consumers: not-determinable (Proposed later consumers or storage are named without enough invocation/selection facts to classify every direction. This uncertainty does not negate the independently supported push/pull routes.)	partial coverage (Push selectors are classified locally. A requested file ID alone is not a push signal, and opaque selectors remain unresolved rather than being made inapplicable by a known pull route.); Proposed QueryEngine context insertion on a query with supplied primary page ID: known — identifier [claimed], coarse [claimed] (Trigger: query generation. Selector input: supplied page_id and graph edges. Selected parts: matching primary page and reachable outbound entries with edge context.); Proposed feedback insertion during affected-concept recompilation: not-determinable (Trigger: another compilation attempt. The record identifies affected concepts and prior feedback, but inconsistent lookup scope leaves the actual identity-match selector unresolved; an identifier finding would overstate it.); Proposed compiler source/related-summary truncation: partial — coarse [claimed] (Trigger: template rendering. Selected parts: supplied retained source text/related summaries/flagged pairs. Length truncation is known; the upstream related-page selector is unknown.); Legacy requested backend results and documented requested pack reads, considered only at their pull operation: inapplicable (These operations fulfill requests rather than independently selecting unsolicited context; no push signal applies to those specific operations.); Unsettled profile/history delivery and external backend/cache/optional embedding automatic selection: not-determinable (No actual push selector is recorded for these alternatives. Optional vector re-ranking and model aliases alone cannot establish inferred-embedding selection; neither lexical nor judgment selection is inferred from a tool name.)	partial coverage (The proposed conversation and audit-feedback update chains support local positives without observed improvement or parameter learning. Other writes lack bounded absence evidence, so no system-wide no is asserted.); Proposed automatic extraction from logged conversation to durable facts/proposals and later committed knowledge: known — yes [claimed] (Automatic derived writes and later committed-memory retrieval are specified; final human commitment is a separate admission step.); Direct user-profile branch of conversation extraction: not-determinable (Automatic profile production is proposed but direct later visibility is undecided, preventing a separately complete qualifying profile-to-consumer chain.); Proposed retained audit criticism to automatic article rewrite to later compilation/query: known — yes [claimed] (The trace-fed durable rewrite chain is design only; broken sketch plumbing prevents wired evidence, not the claimed route.); Legacy imported-page publication, generic remember, original raw sources, audit/job/governance rows, graph/pack outputs and opaque cache/embedding/provider updates: not-determinable (The records do not establish every original input, automatic trace-fed write and later behavior-shaping consumer. Ordinary copying or raw log retention alone is insufficient, and no bounded absence record supports no.)	partial coverage (Authority follows actual proposed or wired consumers, not labels. Profile delivery, opaque backends, cache/embeddings and unconsumed governance metadata leave coverage partial.); Published page bodies and graph relationship context at proposed query/compilation consumers: known — knowledge [claimed] (Knowledge supply is proposed; delivery does not establish model compliance.); Exported article bodies at documented file-aware agents: known — knowledge [afforded] (The external consumer role is afforded, not deployed wiring.); Persisted rejection criticism at ArticleWriter update consumer: known — instruction [claimed], learning [claimed] (Update force and prompt force coexist; neither establishes improved capacity.); Recorded conversation turns at fact-extraction consumer: known — learning [claimed] (This force belongs to the raw experience at the update consumer, not an assumed profile prompt injection.); Heat/reference state at promotion and cold-retention selector: known — ranking [claimed], routing [claimed] (Heat represents relevance, not factual acceptance.); Rejected-count and hand-edit/hash state at proposed scheduler: known — validation [claimed], enforcement [claimed], routing [claimed] (The scheduler tests retained state before queuing another compilation; deployed plumbing remains absent from these sketches.); Published-stage/page-ID/link state at proposed graph/export selectors: known — routing [claimed] (Fixed selection fields determine included material rather than factual truth.); Direct profile consumption, governance verdict effects, backend ranking/retrieval, cache and optional embedding consumers: not-determinable (No settled profile delivery or inspected backend force is established. Stored metadata or trust labels alone cannot establish instruction, validation or enforcement authority.)	AAS-2026-10-15-isaacclupus-mnemosyn-spec-1887e13dcc13-01
llm-context-base	2	Accumulated or use-revised profile/context/training observations, instructions/templates, inbox captures, optional source originals, working knowledge/projects/journal/private/blog/decision content, metadata/indexes, logs and local reload-control recipes; their admission, maintenance and later wiki/linked-code/reference-consumer routes. Initial shipped static guidance/examples are excluded as memory until use revises them; throwaway exports, host-native transcripts/summary/checkpoints/model weights, linked code bodies, and external plugin/sync/platform implementations are excluded. Opaque original and embedded artifact parts remain included with local uncertainty.	known coverage (The recorded inventory locates target-retained bodies, access state and optional reload control in files; tracked copies also use a repository. External sync implementations and host-native stores are excluded, not classified as service memory.); Use-derived profile/context/instructions/templates; inbox, sources, knowledge, decisions, indexes and durable logs: known — files [claimed], repo [claimed] (All named target objects have file storage; repository persistence is conditional on tracking and committing. Ignored journal/private bodies retain the file finding.); Optional journal persistence variants: known — files [afforded], repo [afforded] (The recipes preserve local files and optionally repository copies; underlying encryption/synchronization machinery is external.); Optional compaction sentinel and read-path log recipe: known — files [afforded] (The documented local paths are known; installation and /tmp lifetime are not observed.)	partial coverage (The named target admission inventory distinguishes explicit content decisions from automatic model/software admission. Starting capture, migration or a session is not blanket content approval. Unspecified outcome, saved-synthesis and publication admissions prevent complete coverage; external host summary writes are excluded.); Initial profile answers, human/editor context and capture/journal entry, explicit file destinations and filing decisions: known — manual [claimed], manual [afforded] (These mechanisms have explicit operator content supply/edit/disposition; they do not establish human review of every model-generated sentence.); Clear conversation-pattern writes, observation-log updates, extraction/drafting and item-level source preservation: known — automatic [claimed] (Assistant admission follows the request/setting without a universal per-content human decision; ambiguous capture and approved structures are a separate branch.); Ambiguous pattern capture, new directories and instruction restructuring after approval: known — manual [claimed] (The human explicitly decides these proposed changes, unlike clear immediate rule capture.); Lint/format/status/index/log updates not requiring per-item approval, and human-resolved cleanup: known — automatic [claimed], manual [claimed] (Conditional instructions distinguish automatic requested edits from user-controlled disposition; no universal approval invariant is inferred.); Individual matching-file admission by metadata converter: known — automatic [wired] (The caller selects invocation/root; fixed code selects and overwrites individual matches without per-file approval.); Optional sentinel and read-path hook writes: known — automatic [afforded] (Installation is an operator choice, but callback writes do not require per-record content approval.); Outcome reporting, saving query syntheses and specialized publication/status/URL admission: not-determinable (Outcome entry timing/control is unspecified; a save decision is distinct from actual content approval; external author/assistant/platform roles do not establish who admits each published or tracked retained part. Manual versus automatic admission cannot be completely assigned.)	partial coverage (The file loop includes requested retrieval and automatic startup/task/precedent supply. Optional linked-code consultation and source reprocessing lack fully specified request/selection chains; these unresolved included routes prevent complete direction coverage.); User knowledge/status query and returned selected-content synthesis: known — pull [claimed] (The return satisfies a request for retained material; code delivery is not counted as an additional push.); Session startup and task-triggered retained guidance/template/context loading: known — push [claimed] (A user initiating a task/session is not a separate request to retrieve all learned preferences and guidance.); Decision-context precedent and failure-constraint supply: known — push [claimed] (This branch proactively selects outcome-bearing past decisions without a separate retrieval request.); Old-inbox startup nudge and review-date cadence supply: known — push [claimed] (Age/date control state prompts triage or review without a dedicated request for those retained entries.); Optional post-compaction reinjection and durable-file reload: known — push [afforded] (The next-prompt callback supplies re-init without a separate request to retrieve retained wiki context.); Linked-code consultation and optional preserved-original reprocessing: not-determinable (A documented external reader/reprocessing role establishes read-back affordances, but the records do not fully separate explicit requests from automatic selected supply. No installed caller is supplied.)	partial coverage (Classified push routes use coarse control state, topic/tag lexical inputs and model judgment. Pull-only query operations need no push signal. The external consultation/reprocessing chains remain unresolved; file paths alone do not establish identifier-based push.); Startup and task-triggered accumulated context/guidance supply: known — coarse [claimed], inferred-judgment [claimed] (Trigger is session start or a routed task; phase/date/manual override and task category select current profile/context/rules or relevant module/template bodies.); Decision-context push of completed outcomes and constraints: known — inferred-lexical [claimed], inferred-judgment [claimed] (Trigger is a decision conversation; topic/domain, summaries/tags and filled Outcome content select retained decision parts within the 1–3-file budget.); Startup inbox-age and optimization-date nudges: known — coarse [claimed] (Startup/cadence checks read dates/age and select old captures or the overdue-review state; thresholds do not target semantic facts.); Optional post-compaction reload: known — coarse [afforded] (Trigger is next prompt after manual compaction; sentinel availability selects the reload instruction, then startup conditions select file bodies.); Requested knowledge/status answer: inapplicable (This unit is a requested pull return; its relevance selection is not push selection.); External linked-code/original-reprocessing routes: not-determinable (The records do not specify every automatic selector, trigger or selected payload; known local push routes do not resolve these alternatives.)	partial coverage (Conversation-fed durable behavioral writes are positively supported at the claimed layer. Other named routes are nonqualifying or have unresolved automatic-write/input chains; these limits prevent a system-wide negative or complete trace-route inventory.); Immediate patterns, training observations and conversation-derived personal context/instructions/templates: known — yes [claimed] (Live conversational turns and usage observations feed selected retained text that later sessions/tasks are instructed to consume. This axis requires neither observed improvement nor parametric learning.); Assistant conversation capture/extraction into retained working knowledge: known — yes [claimed] (This unit covers the automatic conversational branch, not independently imported clips or raw manual parking.); Explicit human entry/approval and purely mechanical metadata conversion: inapplicable (Manual content decisions are not automatic trace-fed writes; the converter consumes file metadata rather than activity traces to learn durable behavior. No local no is inflated into a system-wide absence.); Sentinel and read-path logging without a recorded automatic behavioral-update consumer: inapplicable (Raw control/path logging alone is not qualifying learning. The records disclose no automatic path-log-to-guidance pipeline; sentinel re-init reloads existing files rather than producing a conversation summary.); Decision-outcome updates, saved query syntheses, operation/report summaries and source-derived knowledge of unspecified origin: not-determinable (Outcome reporting has no mandatory automatic write schedule; synthesis save and original source content/admission are not fully identified. Operation summaries and logged issue counts have later status/provenance readers, but the original trace inputs to their writes are unspecified. These missing chains prevent classification of additional qualifying automatic trace-fed writes.)	partial coverage (Known effects are local to actual consumers: evidence, guidance, relevance selection, routing and procedural admission criteria. Unspecified artifact consumers and original reprocessing effects prevent complete authority coverage; host permissions and static criteria are not treated as learned memory.); Working content, personal facts, sources and decision reasoning/outcomes at answering or deliberating assistants: known — knowledge [claimed], knowledge [afforded] (Content supplies context/evidence; it does not become an independent factual oracle or guaranteed decision rule.); Use-derived conventions, response preferences and revised module/template guidance at later assistant tasks: known — instruction [claimed] (Standing textual directions are loaded in later sessions/tasks; compliance is unobserved.); Selected experience and observations at the conversation-fed behavioral update consumer: known — learning [claimed] (This learning effect is the update path, separate from downstream guidance and from improved capacity.); Summary/tags/status and decision-outcome content at relevance/prioritization selection: known — ranking [claimed] (Retained descriptions/features affect relative relevance; no learned scoring algorithm is implied.); Profile phase/date, project indexes, directory navigation, review date and operation history at startup/query/cadence consumers: known — routing [claimed] (These retained values choose scope, input windows, questions and review timing; plain descriptive log text has no blanket authority.); Optional compaction reload flag at hook consumer: known — routing [afforded] (The flag changes the next-prompt reload operation, not the host compaction payload.); Use-revised instructions/templates containing completion or format criteria at subsequent candidate-check consumers: known — validation [claimed] (Only the accumulated/revised criteria are memory here; initial shipped static checklist bodies remain outside memory scope. Procedural judgment is not executable enforcement.); Opaque original/embedded artifact semantics, specialized code/data/image consumption and reprocessing outputs: not-determinable (The records establish reference roles but not all downstream artifact consumers/effects. No enforcement or parametric learning authority is inferred.)	AAS-2026-10-14-llm-context-base-f63a7c0fb5f2-01
llm-project-wiki	2	Specified retained Sources snapshots, compiled Wiki content and metadata, index, logs, templates, health dashboard, schema, generated commands and repaired live project/maintenance rules, including pre-existing vault/input provenance and unresolved host delivery of these rules. Excludes static shipped setup text as current invocation input, provider parameters, host compaction/working state, consuming project source code except retained imports, and external plugin implementations. Classifies the prompt-defined memory design, not an inspected installation.	known coverage (Complete classification of the specified substrate inventory, not proof of actual persistence or deployment.); Specified vault and project support files: Sources, Wiki content/metadata, index, log, templates, health, schema, commands and repaired live rules: known — files [claimed], repo [claimed] (The accepted inventory covers the prompt-defined retained file surfaces, including generated and repaired support files; provider state is excluded, not classified.)	partial coverage (Known automatic admissions coexist with unresolved initial and alternate copy control. No supported explicit human content gate justifies manual.); Bootstrap page/support generation, task gap/stale bookkeeping, diff ingest and audit model repairs: known — automatic [claimed] (These specified admission mechanisms are computational, although actual host write permissions and execution are unavailable.); Invoked source sync copy/header admission: known — automatic [claimed] (Original source authorship is separate from copy admission.); Initial existing-vault/project-memory admission and G2 manual-copy alternative: not-determinable (Initial content control is unspecified. Manual copy names a physical recovery action without establishing an explicit human content decision; the foreground audit model may perform it.)	partial coverage (Known requested content reads coexist with unresolved host instruction delivery. Host context management and provider state are excluded, but included live generated rules are not silently dropped.); Question/edit/status page lookup and requested bootstrap/ingest/audit reads of retained references and metadata: known — pull [claimed] (Intent matching and diagnostic selection occur inside the requested read chain; no unsolicited page injection is specified.); Later delivery/loading of generated live project rules and command guidance: not-determinable (The records specify future instruction channels but leave host discovery/loading uninspected; whether delivery is automatic push or explicit requested pull cannot be determined.)	not-determinable coverage (No supported push signal is classifiable; an unresolved included delivery alternative prevents a pull-only system-wide inapplicable assessment.); Requested file/page/query read chains: inapplicable (These inspected design routes are pull; their path identifiers and intent matching are not push selectors.); Possible host delivery of generated live rules/commands: not-determinable (No automatic trigger, selector input or selected delivered part is recorded for host loading. Neither coarse nor targeted push signal is supported.)	partial coverage (The gap-fed route supports a positive at the design layer while opaque project-memory provenance prevents complete trace-learning coverage. No bounded absence record supports a local or system-wide no.); Task-insufficiency gap report to ingest/audit page revision to later lookup: known — yes [claimed] (This is a claimed automatic experience-fed update, not a claim of raw conversation collection, weight training, exercised learning or benefit.); Existing project-memory input and other compilation/repair inputs: not-determinable (Installation-specific memory content and provenance are unknown; the records do not establish whether additional trace-fed writes qualify.)	known coverage (Coverage concerns the specified consumer/effect paths, including live-rule instruction and wiki-first ranking, selection routing and diagnostics. Ranking is warranted by the ordered lookup policy in runtime.md#rt-clm-wiki-first as well as its rule channel and lookup consumer. No retained veto or enforcement consumer is established; all positive force remains claimed design.); Sources snapshots and Wiki domain content at bootstrap, answering, status and maintenance model consumers: known — knowledge [claimed] (This is the intended content force; source warrant and deployed compliance remain unknown.); Schema, commands, template guidance, archive warning and live repaired rules at subsequent model consumers: known — instruction [claimed] (Policy force is specified; it is not host enforcement.); Retained live wiki-first project rules at future question/edit model consumers: known — ranking [claimed] (This specified priority effect coexists with standing instruction and operation-selection effects; actual host loading and model compliance remain unobserved.); Index mappings, freshness/path metadata, health predicates and log deferrals at lookup and maintenance selectors: known — routing [claimed] (These selector effects concern scope and operation; the separate live-rule unit preserves wiki-first relative priority as ranking without requiring a numerical algorithm.); Retained schema/template criteria at authoring and audit conformity checks: known — validation [claimed] (Diagnostic/schema force is bounded to conformity; incomplete grep checks do not certify page truth.); Task-derived gap reports at ingest/audit updates to durable pages: known — learning [claimed] (Learning authority at the update consumer is distinct from knowledge authority at the downstream page reader and from improved capacity.)	AAS-2026-10-14-llm-project-wiki-c6c664168f3e-01
Prove2Me_workspace	2	Use-retained mathematical library and submissions; prose, comments, milestone/review/description histories, discovery metadata, collection/campaign state and read-back testimony; local credentials, environment/build caches, declaration mirrors, solutions and interrupted-upload recovery files. Source extraction facts are included only within the retained upload-state chain, not merely as current-run stdout. Static shipped skill/playbooks, ordinary current invocation context and external mathematical references before retention are excluded. Opaque host/provider retention and updates remain explicitly uninspected limits, not inferred stores.	partial coverage (All inventoried artifact-level storage surfaces have classifications; external host/provider retention remains uninspected. Neither graph-shaped dependencies nor JSONL encoding establishes a graph database or repo substrate.); Remote mathematical library, submissions, explanations, comments, histories, reviews, discovery metadata, collections, campaigns and read-back fields: known — service-object [claimed] (The report inventory covers the exposed remote retention surfaces. Persistence is a contract, not inspected service implementation.); Local mirrors, authored solutions, credentials, environment/build caches, generated upload files and recovery bookkeeping: known — files [claimed] (The specified local objects are files; the resumable state format and uploader implementation remain unspecified, without changing the stated substrate.); Opaque host/provider retained state, if any: uninspected (The external host/provider is not supplied. No retained checkpoint, update or later-memory consumer can be classified from the model component alone; model computation is not evidence of a model-weights memory store.)	partial coverage (Automatic and manual admission mechanisms are separately supported. Caller identity, authorship and physical HTTP/file writes do not resolve the remaining controls.); Agent proof/direct publication, explanation/comment composition, mirroring, parsed access metadata, history capture, extraction and prescribed uploader action recording without per-content human approval: known — automatic [claimed], automatic [wired] (Starting the workflow or supplying a project is not approval of each resulting object; the helper witness does not upgrade the unshipped uploader/backend.); Human proposal item confirmation/launch, moderator reports and campaign value admission: known — manual [claimed] (These human gates are local. They do not apply to direct contribution or upload-first publication.); Generic account/editor saves, ratings, votes, collection edits, initial live content and opaque host/provider writes: not-determinable (Account identity alone does not determine human versus computational admission. Initial live-state control and host/provider updates are not supplied.)	partial coverage (Requested retrieval is supported across the specified artifact paths. Uninspected host alternatives prevent a system-complete pull-only or no-push conclusion.); Solver scouting and later compiler imports; captain review reads; human/moderator testimony reads; collection/campaign browsing and template copying; setup reuse and interrupted-upload recovery: known — pull [claimed] (The report inventory explicitly identifies the later readers; ordinary extraction output within its producing invocation is excluded.); Opaque host/provider memory delivery and unresolved optional-history consumers: uninspected (No host memory injection route is inspected, and optional endpoint availability does not establish a later invocation. This prevents a complete direction inventory.)	uninspected coverage (There is no supported push-signal positive. Local pull-route inapplicability does not make the whole axis inapplicable while external delivery remains uninspected.); Inspected artifact later-consumer routes assessed as requested retrieval: inapplicable (Push-signal classification is inapplicable to these pull operations. IDs, filters and pagination on a request are not automatic push selectors.); Opaque host/provider memory injection alternatives: uninspected (No push trigger, selector input, selected retained part or delivery channel is supplied for host/provider state. Therefore no controlled push signal can be classified.)	not-determinable coverage (No accepted record establishes the complete automatic trace-to-durable-derivative-to-later-consumer chain. This is unresolved classification, not evidenced absence or a claim that authored agent guidance cannot qualify.); Authored experience comments/explanations, retained failed source/status, captain reasons and product-review correction: not-determinable (These records establish experience externalization and later advisory use, but do not establish a particular automatic write consuming recorded traces to produce a durable behavior-shaping derivative. Agent authorship is not an exclusion: a recorded trace-fed agent write could qualify without weight updates or demonstrated improvement.); Uploader recovery, extraction-derived content, copied templates and refreshed static procedure text: not-determinable (Action-ID recovery and source-code extraction are not by themselves qualifying trace-fed behavior updates. No bounded absence record covers all possible writes, so a local no is not asserted.); Opaque host/provider trace-fed updates: uninspected (The external host/provider is not supplied. No retained checkpoint, update or later-memory consumer can be classified from the model component alone; model computation is not evidence of a model-weights memory store.)	partial coverage (The listed knowledge, instruction, routing, ranking and validation paths have distinct consumers. Unspecified consumers and external implementation prevent a complete authority inventory.); Formal declarations, prior code, explanations, comments, reviews and read-back testimony delivered to later solvers, captains, humans or moderators: known — knowledge [claimed] (The consumers and advisory/reference effect are specified; delivery does not establish actual changed behavior or truth.); Milestone source specification and captain rejection reasons at later solver guidance: known — instruction [claimed] (This is standing mission guidance, not enforced compliance or revocation of existing imports.); Saved IDs, tags, collection membership, graph frontier and upload action IDs at selection/recovery consumers: known — routing [claimed] (Selection of an operation/input differs from relative ranking and from mathematical evidence.); Votes/trust, milestone order, campaign value/date/status and frontier closability at ordered discovery consumers: known — ranking [claimed] (The vote/trust and campaign selectors are documented. Rating-to-sort linkage is not included in this finding.); Retained campaign template/foundation and milestone specification at matching/attestation decisions: known — validation [claimed] (Validation is scoped to these decisions, not a source-truth certificate or a shipped backend check.); Optional rating reasons, historical description snapshots, credential/cache internals and opaque host/provider state at stronger or unspecified consumers: not-determinable (Optional reason and history storage does not establish a routine behavioral consumer; authentication enforcement and provider memory effects are uninspected. No learning or enforcement finding is inferred merely from retention.)	AAS-2026-10-14-prove2me-workspace-187ef30582c9-01
sig-releases	2	Accumulated private factual records and personal interpretations, approved team knowledge, reusable shared skills, and their admission, maintenance and later-consumer paths. Raw capture retention, organizational draft retention, access metadata, optional Slack/voice acquisition, conversation retention and provider-side retained context remain unresolved included alternatives. Excluded are static release documentation, provider model weights not shown to change through use, credentials, analytics linkage and telemetry, and public support issues: no route from those diagnostic objects to durable app behavior-shaping updates is established. This is the documentary memory boundary, not the whole runtime's storage.	partial coverage (Local files are a supported existence finding, not an inventory of all storage in the partial application loop.); Private factual records and personal-layer notes on the local computer: known — files [claimed] (File retention is classified at the documented layer; file-versus-section layout does not change this substrate.); Published team knowledge and reusable skills, including recipient copies: not-determinable (Git synchronization is named, but operative remote storage and skill placement are unspecified; transport alone does not establish repo or files.); Raw captures, drafts, access metadata, integration outputs, transcripts and opaque provider context: not-determinable (Independent retention and substrate are unresolved for these alternatives; provider internals and optional acquisition paths are unavailable.)	partial coverage (Only organizational publication has an explicit per-content human admission warrant; automatic local admission is not inferred from Sig's authorship.); Admission of selected organizational text to team publication: known — manual [claimed] (The documentary publication decision is manual; gate enforcement is not inspected.); Local private-note and draft admission, initial imports and capture retention: not-determinable (Sig proposes structured text and writes files, but per-content local approval is unspecified. Human initiation and physical software writing settle neither automatic nor manual admission.); Skill creation/replacement/publication, recipient admission, integration and provider writes: not-determinable (One-click skill sharing does not disclose a content approval decision; recipient, optional integration and provider controls are unknown.)	partial coverage (Human file pull is supported; opaque AI and skill routes prevent a pull-only or complete direction assessment.); User opening and reading accumulated local note files: known — pull [claimed] (The explicit user file-access path has a described requesting consumer.); Private history and published team knowledge reaching later AI consumers: not-determinable (The documents claim context use but do not distinguish requested retrieval, automatic injection and mere synchronized availability.); Skills and other opaque retained-input alternatives reaching later consumers: not-determinable (Invocation triggers, loading channels and independent retention are unresolved; one-click publication is not a read request.)	not-determinable coverage (The known human pull unit cannot make the whole signal axis inapplicable while model routes remain unresolved.); Human-requested opening of local note files: inapplicable (This unit is requested file access, not push; a filename on that request establishes no push selector.); Possible private/team context push, skill delivery and opaque integration/provider alternatives: not-determinable (No actual push operation has a disclosed trigger, selector input and selected retained part; immediate team availability does not establish coarse or identifier selection.)	not-determinable coverage (Neither yes nor bounded no is established. The context-payoff claim concerns benefit, not a qualifying automatic trace-fed write.); Meeting narrative to private notes, organizational material and later AI context: not-determinable (Derived retained context is claimed, but automatic private admission and original trace content are unresolved. Manual team publication alone does not warrant a negative for all precursor writes.); Skill updates, optional integration outputs, conversation summaries and provider context writes: not-determinable (No recorded automatic trace-fed update and durable later-consumed result can be classified; opaque alternatives remain included rather than absent.)	partial coverage (Knowledge is supported at described consumers only. Neither instruction, learning, ranking nor enforcement follows automatically from retention or publication approval.); Private work-history content at its intended later AI-context consumer: partial — knowledge [claimed] (Exact selected parts, especially personal-layer provider inclusion, and actual influence are unresolved.); Approved team account at teammate Sig's intended context consumer: partial — knowledge [claimed] (Context supply is claimed, while selected payload, invocation and activation are not established.); Private notes at the user's direct file-reading consumer: known — knowledge [claimed] (This bounded human-reference path does not establish any AI selector or consumer.); Skills, retained capture/draft alternatives, metadata, integration and provider context consumers: not-determinable (Workflow bodies and actual downstream effects are unspecified. A review/send policy is not itself an accumulated retained enforcement object.)	AAS-2026-10-14-sig-releases-50379145ac31-01
trajectory-informed-memory-generation	2	Documented cross-task memory: generated and consolidated strategy, recovery and optimization tips; original task, generated subtask and canonical descriptions; stored embeddings; structured tip metadata; extraction/admission, consolidation and both pre-task retrieval routes. Completed trajectories supply the trace-fed write, but durable raw-input retention and intake control remain unresolved. Extraction intermediates and transient query vectors have no established later invocation consumer. Static prompts, provider model weights, external API state and benchmark machinery are not accumulated memory within this profile; no parameter-update route is established. Physical persistence, initialization, maintenance implementation and actual activation remain unverified.	partial coverage (The retained product/access inventory is documented, but only vector storage has a controlled positive; other physical substrates and possible raw retention remain unresolved.); Stored tip and description embeddings: known — vector [claimed] (Covers stored content/purpose and description vectors, not transient query vectors or provider parameters.); Tip text, description text and structured metadata storage: not-determinable (Memory-system storage and structured attributes are described, but no controlled substrate is identified for these parts. Encoding does not establish files, a relational database or a key-value store.); Possible durably retained raw trajectories: not-determinable (Neither durable raw retention nor its substrate is specified; no raw-log storage value follows from its use as extraction input.)	partial coverage (Known computational successor admission coexists with unresolved initialization; no manual finding is established.); Model extraction and subsequent consolidation admission: known — automatic [claimed] (Covers both task/subtask extraction and semantic management, not merely physical I/O.); Initial/raw-input admission and corpus initialization: not-determinable (Intake controller and initial-content control are unspecified; neither manual nor automatic initialization follows from model authorship downstream.)	partial coverage (The documented cross-task routes are push; implementation/reload gaps prevent a complete inventory rather than establish pull-only or push-only deployed behavior.); Pre-task cosine and LLM-guided tip selection and guidelines insertion: known — push [claimed] (Both alternative selectors share the same delivery channel; internal query execution is not a separate consumer pull.); Other later-consumer routes, including raw-memory reload: not-determinable (Raw retention, reload and mid-run memory access are unspecified, not evidenced absent; no additional pull route can be classified.)	partial coverage (Embedding and judgment selection are documented locally. Missing access-route coverage and index wiring prevent complete system enumeration.); Cosine branch: new task to stored descriptions to associated tips: known — inferred-embedding [claimed] (Top-3 and uncapped variants remain configuration-specific; an item cap is not a token budget.); LLM-guided branch: task interpretation and hybrid query selecting tips: known — inferred-judgment [claimed], inferred-embedding [claimed] (The documented hybrid selector targets tips; source IDs alone do not establish identifier-targeted push.); Selection on unspecified additional reload/access routes: not-determinable (Additional read-back and any corresponding selector are unresolved; known initial push does not classify them.)	partial coverage (A positive design-level route is established; unknown additional writes do not erase it or justify a system-wide negative. Actual benefit and exercised self-improvement remain separate questions.); Completed trajectories through extraction and consolidation to later-task guidelines: known — yes [claimed] (Task/subtask and labeled/self-inferred outcomes share this declared durable-guidance chain; operation is unobserved.); Unspecified initialization and other maintenance writes: not-determinable (The capture does not enumerate intake control or executable maintenance, preventing assessment of every possible qualifying write.)	partial coverage (Declared consumers support advisory, update, ranking and scope effects. Raw retention and uninspected maintenance consumers prevent a complete accumulated-memory authority inventory; benchmark scoring is not a memory authority route.); Retained tips at the later ReAct agent's guidelines consumer: known — instruction [claimed], knowledge [claimed] (Both effects concern delivered text; actual compliance or content-dependent activation is not observed.); Accumulated tips consumed by semantic consolidation: known — learning [claimed] (Authority at the update consumer is separate from the successor's advisory authority and from demonstrated improvement.); Stored descriptions/vectors and priority/outcome fields at selection consumers: known — ranking [claimed] (Fixed scoring still gives retained representations ranking force; no learned ranking policy is inferred.); Stored application/task/category metadata at LLM-guided retrieval: known — routing [claimed] (This is access-scope selection, not an API capability grant or independent truth-validation gate.); Possible retained raw trajectories at extraction: not-determinable (Trace evidence feeds extraction, but raw retention across invocations is unspecified; its authority as an accumulated raw-memory object cannot be completed.)	AAS-2026-10-14-2603-10600v1-a956324c9a82-01
wikiskill	2	Accumulated raw training trajectories; wiki pattern pages, relevance index, evolution log and intervention/validation history; candidate/active SKILL.md and companion PURPOSE.md; their automatic writes, edits, withdrawal and later maintainer/proposer/task consumers, including wiki-enabled inference. Unresolved provider caches, validation-trace retention, durable restore checkpoints, initial admission and companion delivery remain named coverage limits. Excludes static shipped prompts, pretrained model weights without an evidenced update, task documents/current outcomes except where retained within experience, ordinary current rollout/history state and the live scalar best-score variable. Cross-run reload is not established; described persistence spans evolution rounds and final skills/wiki return.	partial coverage (The described seven-object workspace has files storage at claimed strength. Potential additional retention and restore substrates remain unresolved; this is not a whole-runtime storage inventory.); Documented Raw, Wiki and Skill filesystem objects, including the separately retained PURPOSE.md: known — files [claimed] (The seven named objects cover the documented three-layer workspace. The log filename ambiguity does not change its declared filesystem substrate; companion rollback mechanics do not change its declared storage.); Potential retained provider state, validation-trajectory archive and durable restore checkpoint: not-determinable (The accepted records leave provider caches, validation-trace retention and durable incumbent snapshot storage unresolved. Their existence, content and later-consumer routes are not established, preventing complete storage-substrate coverage; model backend weights alone are not evidence of accumulated memory.)	partial coverage (Known computational admission routes coexist with unresolved initialization, companion maintenance and potential additional retention. The automatic findings do not upgrade unknown alternative admission mechanisms.); Harness admission of generated raw training trajectories: known — automatic [claimed] (Starting an experiment is not a per-content admission decision.); Maintainer-selected pattern edits, replacement index and log append: known — automatic [claimed] (Model authorship, admission choice and physical application are separate roles, all computational in the documented route. Edit failure handling remains unknown.); Single-skill creation/patch and associated creation-time rationale admission: known — automatic [claimed] (This classifies the described candidate route, including creation-time PURPOSE.md supply; synchronized companion patch/restore coverage remains unspecified.); Harness append of intervention history after acceptance or rejection: known — automatic [claimed] (Rejected content can be admitted as history without being admitted as active instructions.); Initial workspace creation, alternate human seeding/replacement and undocumented companion-file updates: not-determinable (Initial content is declared empty but initializer identity/control and alternate content replacement are unspecified. Generic caller identity cannot establish manual agency; no manual finding is inferred.); Potential retained provider state, validation-trajectory archive and durable restore checkpoint: not-determinable (The accepted records leave provider caches, validation-trace retention and durable incumbent snapshot storage unresolved. Their existence, content and later-consumer routes are not established, preventing complete write-agency coverage; model backend weights alone are not evidence of accumulated memory.)	partial coverage (The default routes support both push and pull. Missing companion, alternative inference and additional-state routes prevent complete enumeration; default wiki denial is not a universal absence of direct wiki read-back.); Training traces and current wiki delivered to the maintainer at each maintenance call: known — push [claimed] (Stratified sampling and full-wiki supply are declared. Complete intervention-tracker inclusion is separated as unresolved.); Initial index and intervention-history delivery to each proposer invocation: known — push [claimed] (The initial all-task outcome summary also appears, but current outcomes alone are not an additional accumulated-memory object.); Proposer-requested pattern, trace and history file reads: known — pull [claimed] (Trace aliases and filenames resolve requested content; they do not establish unsolicited identifier selection.); Candidate/active procedural content automatically supplied to task inference: known — push [claimed] (Provisional validation and later accepted inference are both specified. Full injection is not task-relevance retrieval.); Separate PURPOSE.md delivery, complete maintainer tracker supply and wiki-enabled inference access: not-determinable (These included alternatives lack a settled consumer/delivery chain. Wiki access alone does not decide pull versus push; conflicting maintainer completeness claims cannot supply a joint complete route.); Potential retained provider state, validation-trajectory archive and durable restore checkpoint: not-determinable (The accepted records leave provider caches, validation-trace retention and durable incumbent snapshot storage unresolved. Their existence, content and later-consumer routes are not established, preventing complete read-back-direction coverage; model backend weights alone are not evidence of accumulated memory.)	partial coverage (Coarse push selection is supported for the default automatic supplies. Model relevance judgment guides requested pulls, so it is not an inferred-judgment push finding. Unknown alternative directions cannot be made signal-inapplicable.); Maintenance-call trace push: success/failure stratification and capped logs: known — coarse [claimed] (The trigger, selector input and selected retained traces are named. Tie-breaking and which characters survive the cap are unspecified; categories/budgets are not targeted identity matches.); Maintenance-call current-wiki push and proposer-start index/history push: known — coarse [claimed] (This classifies the documented wholesale/default payload supply, not the later model-requested detail reads. It does not claim complete maintainer tracker delivery.); Task-invocation full skill push: known — coarse [claimed] (Accepted or provisionally applied configuration supplies the entire set. File identities alone do not establish identifier push.); Wiki-enabled inference selector and uncertain rationale/maintainer-history delivery: not-determinable (Access is described without selection inputs, budget or delivery mechanics for the alternative inference route; rationale and necessarily complete tracker supply remain unresolved. Their push signals cannot be inferred from the known default paths.); Potential retained provider state, validation-trajectory archive and durable restore checkpoint: not-determinable (The accepted records leave provider caches, validation-trace retention and durable incumbent snapshot storage unresolved. Their existence, content and later-consumer routes are not established, preventing complete push-selection coverage; model backend weights alone are not evidence of accumulated memory.)	partial coverage (Several claimed automatic trace-fed artifact routes qualify. Unresolved companion and additional-state chains keep the axis partial; no local or system-wide no is inferred from missing execution evidence.); Automatic trace-fed wiki pattern, index and evolution-log production with later maintainer/proposer consumers: known — yes [claimed] (The complete claimed raw-to-derived-to-later-consumer path qualifies without weight updates or demonstrated improvement. Raw archiving alone is not the qualifying write.); Trace/wiki/feedback-fed skill creation or refinement followed by validation or later task consumption: known — yes [claimed] (Acceptance controls the later active instruction set but provisional validation consumption can precede it. Rationale delivery is not required to establish the SKILL.md path.); Automatic validation-feedback history append followed by later proposer consumption: known — yes [claimed] (This is a qualifying feedback-to-durable-guidance route, not proof that every entry changes behavior or improves capacity.); PURPOSE.md considered separately from consumed procedures, complete maintainer tracker feedback and additional provider/checkpoint writes: not-determinable (Rationale is retained but separate later consumption is unresolved; complete maintainer feedback delivery is disputed. Potential additional trace retention and durable writes have no established raw-to-derived-to-later-consumer chain. No bounded system-wide no follows.)	partial coverage (Supported knowledge, learning, ranking and instruction findings stay on actual retained-part consumers. The scalar current best score and static prompts are outside this accumulated-memory boundary, so their admission/enforcement roles are not attributed to retained history. Unresolved consumers and additional state keep coverage partial.); Raw traces consumed by maintainer and proposer to construct durable wiki and skills: known — knowledge [claimed], learning [claimed] (Evidence supply and update use are distinct effects of the same retained traces. Neither establishes truthful diagnosis or improved future capacity.); Patterns, index and evolution log at subsequent maintainer/proposer consumers: known — knowledge [claimed], learning [claimed] (This is the default mediated wiki pathway. Pattern advice is not automatically standing instruction to default task inference.); Index summaries influencing which pattern pages the proposer reads: known — ranking [claimed] (Ranking influence is model-interpreted relevance judgment, not an evidenced vector index or learned retrieval scorer. Requested read selection is still pull.); Intervention history supplied to later proposer: known — knowledge [claimed], learning [claimed] (The proposer route is specified. The history advises; no machine-enforced veto on repeating proposals is established.); Procedural SKILL.md at provisional validation inference and later active task inference: known — instruction [claimed] (Candidate exposure can precede acceptance. This is interpreted guidance, not enforced compliance or explanation endorsement.); Separate PURPOSE.md consumers; complete maintainer intervention-history supply; wiki-enabled inference force: not-determinable (Rationale delivery is not separately specified. Runtime and specialist records disagree about necessarily complete maintainer tracker delivery. Alternative inference wiki access is described but its consumed payload/channel is unresolved. These limits prevent a complete consumer/effect inventory.); Potential retained provider state, validation-trajectory archive and durable restore checkpoint: not-determinable (The accepted records leave provider caches, validation-trace retention and durable incumbent snapshot storage unresolved. Their existence, content and later-consumer routes are not established, preventing complete behavioral-authority coverage; model backend weights alone are not evidence of accumulated memory.)	AAS-2026-10-15-2608-27454-65116a9d0de7-01

Input identities

• Profile revision 2; kb/agentic-system-analyses/retained/wujiangxu--a-mem-sys/overview.md: 0ad1b8032e7646758b862e29f0d5b8fa721265c6514b77ae6852834eded14cdd; artifact kb/agentic-system-analyses/retained/wujiangxu--a-mem-sys/ARTIFACT.yaml: 16fad56f4a88e833d4c0f5ffa129af99bca8a0383acbd6799d2dc99b28cd2f28; source https://github.com/WujiangXu/A-mem-sys at f303dfc71e07bdc787f4bc135d4cea328ae30e99.
• Profile revision 2; kb/agentic-system-analyses/retained/imbad0202--academic-research-skills/overview.md: 60a3cf6cc5c6ff6ba75fc1c16ab82a3a793d406818fe88849cc0c195d07358f5; artifact kb/agentic-system-analyses/retained/imbad0202--academic-research-skills/ARTIFACT.yaml: bf7dab128a42096b3e3a5c88eb974ea888d5c20f1cd1149562a0188d5b67b3c3; source https://github.com/Imbad0202/academic-research-skills at 060cb3cbffc2ca1bdfb8f7c6d2ffa4883cf09b1a.
• Profile revision 2; kb/agentic-system-analyses/retained/2601-01885v1/overview.md: 3456493c0e9003e294fc4952d097ba7dae938292a5d99746474978784834e0b4; artifact kb/agentic-system-analyses/retained/2601-01885v1/ARTIFACT.yaml: ec5ef488e3864432af6583e3410daeba596badd0839a85b58c257f4f7322306e; source https://arxiv.org/html/2601.01885v1 at arxiv-2601.01885v1-html-2026-10-12.
• Profile revision 2; kb/agentic-system-analyses/retained/simular-ai--agent-s/overview.md: f7bc6079e7098b02bd98b7f43de6bc9c95ee905036e190060da206df9854a260; artifact kb/agentic-system-analyses/retained/simular-ai--agent-s/ARTIFACT.yaml: 5334d5e19c716cfc3639f70644bcf090d89eb701fbc33783f90bd928f4d98e1e; source https://github.com/simular-ai/Agent-S at 15776a8dc659658b1963816ea006f7af1978e595.
• Profile revision 2; kb/agentic-system-analyses/retained/muratcankoylan--agent-skills-for-context-engineering/overview.md: 77077f3cc331bc9663e43f219011907bf2f1e280c51b25f2bea8cbb493f39878; artifact kb/agentic-system-analyses/retained/muratcankoylan--agent-skills-for-context-engineering/ARTIFACT.yaml: ac0f099fabd58597f6927f44f8591cdb8d2495d6fd0b3ee451ec93411a57b3cb; source https://github.com/muratcankoylan/Agent-Skills-for-Context-Engineering at 9ff0c99108b36d3b886defd825d8ac1f47dba374.
• Profile revision 2; kb/agentic-system-analyses/retained/zorazrw--agent-workflow-memory/overview.md: ebd57cc5adea92f797c6bf9638aa084984772b1e0b4257a030bdf1826b3bd44c; artifact kb/agentic-system-analyses/retained/zorazrw--agent-workflow-memory/ARTIFACT.yaml: 048c2bd8bd64ef458ec712465dfd418d42123000bb9dee665dfe26c2c209cded; source https://github.com/zorazrw/agent-workflow-memory at 8c0ff8cd11d648c8fceb99e4e42f37e3b75381b1.
• Profile revision 2; kb/agentic-system-analyses/retained/bytedance-seed--agent-r/overview.md: ccfa7f7c467a91046b01cbb54205872adbbd81e53febd75092531d749b8e9ef7; artifact kb/agentic-system-analyses/retained/bytedance-seed--agent-r/ARTIFACT.yaml: 1f1168ff9af8f458b9e09b87bc9ece10474db7ef895248635a74622181231a57; source https://github.com/ByteDance-Seed/Agent-R at 82fcc1ca7873b460949ed49022146dd988c32e31.
• Profile revision 2; kb/agentic-system-analyses/retained/agent-on-the-fly--agentfly/overview.md: 7b9f958ecc3937ee550babdbdc004006bfa60981b5a0c10df86d2ed89d54767f; artifact kb/agentic-system-analyses/retained/agent-on-the-fly--agentfly/ARTIFACT.yaml: fd049fd1c1f36bd124f3eeecb22960129b138cc29eb69b2d75f9cf36d9f17d35; source https://github.com/Agent-on-the-Fly/AgentFly at 42fbbcac63dd58ed6856c0761357345a58e4f032.
• Profile revision 2; kb/agentic-system-analyses/retained/agent-creativity--agentic-local-brain/overview.md: a99223626deb2034f7377b7a132aeb67a9c6dcd2a327d5d20c200699d7a0afe9; artifact kb/agentic-system-analyses/retained/agent-creativity--agentic-local-brain/ARTIFACT.yaml: aed7791f2b5d2c53eeb314cf5c74b91a79b03c3472b76c6a0d8c205f2def6476; source https://github.com/agent-creativity/agentic-local-brain at 41ea0cb104686eb0d39f32c94c2b84d37cdff0e7.
• Profile revision 2; kb/agentic-system-analyses/retained/agno-agi--agno/overview.md: 6159044f672c5174d490e21530b3e0981dd19f45de20f0a10032267990648807; artifact kb/agentic-system-analyses/retained/agno-agi--agno/ARTIFACT.yaml: e2adeea6acc18fd42b3bcead06a03c505b91ca72ea74f95e0122dfcac9e35a2b; source https://github.com/agno-agi/agno at 00497422d4a386d93ab3b45c1a11faf42fcc932f.
• Profile revision 2; kb/agentic-system-analyses/retained/bojieli--ai-agent-book/overview.md: 4da820fcf02d07bb132074ed1eab8aefbec280b5cb156e2b843012c00c69d24d; artifact kb/agentic-system-analyses/retained/bojieli--ai-agent-book/ARTIFACT.yaml: d192ebf0d7ca4ca81a64171cfd9c4da5f36f2fc57a1043a8658751f8564d6209; source https://github.com/bojieli/ai-agent-book at dbc046eb896ac4e39aa19c7774c8bf49583b89a6.
• Profile revision 2; kb/agentic-system-analyses/retained/iusztinpaul--ai-research-os-workshop/overview.md: c12c66037ae00c2d3742f9466e8bda3a2968fc30942f492dd4c0598aac20b129; artifact kb/agentic-system-analyses/retained/iusztinpaul--ai-research-os-workshop/ARTIFACT.yaml: 6a6cb417037e071e5ae6cfd6e984a6e389f667ade169299cdd65d6980840a733; source https://github.com/iusztinpaul/ai-research-os-workshop at dc6660555266239cd36483cf51b5ab39febacf02.
• Profile revision 2; kb/agentic-system-analyses/retained/alexdcd--ai-context-os/overview.md: 42e40fd0884fd1ccff1dfd0f8d54b56d6b2df0191084734fec5db5c7919d53f4; artifact kb/agentic-system-analyses/retained/alexdcd--ai-context-os/ARTIFACT.yaml: 2548f4276155708e1a8ae8f7589ed23be509f79c505ed75f1ab03d2b47131a4c; source https://github.com/alexdcd/AI-Context-OS at 9ed263080cadf98702bebb42d5260b45bd526f6c.
• Profile revision 2; kb/agentic-system-analyses/retained/zelixag--ai-memex-cli/overview.md: 62e693f6d9ad7e3197ea3f7f9dc577ae586bc79ba75e1af37f9a471b6f147b3f; artifact kb/agentic-system-analyses/retained/zelixag--ai-memex-cli/ARTIFACT.yaml: bebb8fbfec0eecf5bd5a6173abe684c5dcda26c27793b534fef61e9f49b98cf0; source https://github.com/zelixag/ai-memex-cli at 4adb274a33cbb9ff0f03a6aeeb9c8d2525a5ac3f.
• Profile revision 2; kb/agentic-system-analyses/retained/theafh--ai-modules/overview.md: 5ba7ea6f602ed02c3a4617be00db86fc91889c2e6c08fff27b53f2eca87b918b; artifact kb/agentic-system-analyses/retained/theafh--ai-modules/ARTIFACT.yaml: 0ee1e6fd3e74fe1535294fdb946b5f4a0a03828e9f4f6f731af3654caa38f204; source https://github.com/TheAFH/ai-modules at 0cf0227c416f75a79d7317f603a209a63e7a3c8f.
• Profile revision 2; kb/agentic-system-analyses/retained/2609-26457/overview.md: c485f86f11f36c938660fd8ea2d15c9181b1e14849a587c2dd4a9c93a4d2fc09; artifact kb/agentic-system-analyses/retained/2609-26457/ARTIFACT.yaml: 23837ced2ce4224904df8aa9473eacbd92223cd1769cbeef7f28a2c0fa047a5e; source https://arxiv.org/abs/2609.26457 at arxiv-2609.26457v1-2026-10-14.
• Profile revision 2; kb/agentic-system-analyses/retained/apache--maka/overview.md: 46bebb1e99d80e23d386cd99b3cb5992e7730b1289bf9643681dc80b7a757436; artifact kb/agentic-system-analyses/retained/apache--maka/ARTIFACT.yaml: 32b05d352ffa14e291e40d8032c3bbad05135d577755cbb97abc068d4a194c50; source https://github.com/apache/maka at 1751c399239b0f6b717e98e2a5e2b875c3dd11ce.
• Profile revision 2; kb/agentic-system-analyses/retained/gchamon--archie/overview.md: c32ba4ba06571a362861c026a2de08621efb8c9ed3af014cad38dd39d0d824d8; artifact kb/agentic-system-analyses/retained/gchamon--archie/ARTIFACT.yaml: c11ccd950f9ef023c1a3c5a924bc266881598738c2268d3e6819d4d74e222aec; source https://github.com/gchamon/archie at 76caf94548cf76f2399d9a7446da15e0daa22685.
• Profile revision 2; kb/agentic-system-analyses/retained/airi-institute--arigraph/overview.md: 541dba3699db24c5d19b0af05d163ac81abd36a105d21690489e9d001dc41a79; artifact kb/agentic-system-analyses/retained/airi-institute--arigraph/ARTIFACT.yaml: d5613598acb6801f05571fa145f5efd19dd15d6f26cf76f24fb47b4196de7985; source https://github.com/AIRI-Institute/AriGraph at e884b76d7fa5185a3a8a55e5a67393b5a43f5ef2.
• Profile revision 2; kb/agentic-system-analyses/retained/agenticnotetaking--arscontexta/overview.md: 12e89518a49a8d98ac9d052dbb0531c440c1c1cf4f72476ae88707c9bbd61427; artifact kb/agentic-system-analyses/retained/agenticnotetaking--arscontexta/ARTIFACT.yaml: b0d32b6aa2860b486fb93569334b6e7010bc0a1ac2082d6522fbe606dc26ce08; source https://github.com/agenticnotetaking/arscontexta at 2acfd5cc4473c4d06c46be63df748e77e00e2746.
• Profile revision 2; kb/agentic-system-analyses/retained/arsumbris--arsumbris/overview.md: 93a83f93c887a12c3a399c6fe90c75792f8e24a60ce5a0457f444fb3a81604e4; artifact kb/agentic-system-analyses/retained/arsumbris--arsumbris/ARTIFACT.yaml: 6d5063eddc8671199ad6daaeaf4c352aaeaff7480deb57f1d320478fcbd434f2; source https://github.com/arsumbris/arsumbris at 716bb14f2bf50e249d414525c8e07c50478fc786.
• Profile revision 2; kb/agentic-system-analyses/retained/cargofy--atlas/overview.md: 8f40115f02405f7c40560c976c1319b66d40daaa89acad786b45b283b8fbc3ac; artifact kb/agentic-system-analyses/retained/cargofy--atlas/ARTIFACT.yaml: 8ef4cdbbac44beca0d1d1463cd01c15f051aff3d19dd5b3d192afd56ee44c08e; source https://github.com/cargofy/ATLAS at 65b91b4fd5d703e8a82ad8fa2a9b8d2f38754988.
• Profile revision 2; kb/agentic-system-analyses/retained/kenforthewin--atomic/overview.md: e44bc3bd38879c945dd25831580853f17c1abaf7c07e357a2fa20cdcb3c7eb7c; artifact kb/agentic-system-analyses/retained/kenforthewin--atomic/ARTIFACT.yaml: c99b574674ed34dbc1ae3e0b2b9a8498383c2db732f3978449419c76184fb5e8; source https://github.com/kenforthewin/atomic at 983bea6baf449d060514c94b3b15b5f85f77ba9b.
• Profile revision 2; kb/agentic-system-analyses/retained/wanshuiyin--auto-claude-code-research-in-sleep/overview.md: a1fe9200b1b8f6826053ad6c4b3d249a247dbcf750a3a239f4ce267af9b919b0; artifact kb/agentic-system-analyses/retained/wanshuiyin--auto-claude-code-research-in-sleep/ARTIFACT.yaml: 79c1ba21b7c049ab896b244a7550f9bcfcf196a94957fec3f0bec1828fb6cb05; source https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep at 26b95cfa0d8747078e9e43b42e20952709e561b8.
• Profile revision 2; kb/agentic-system-analyses/retained/neosigmaai--auto-harness/overview.md: a60dc315bdf5023fd758734ca5bc0667bc5bd77ce3d42899dbe9154948987809; artifact kb/agentic-system-analyses/retained/neosigmaai--auto-harness/ARTIFACT.yaml: 187bfb8d1adde8d63805e8ea289b7c41de9b57ea22cc6bebfce2019112f3ef29; source https://github.com/neosigmaai/auto-harness at 52360dbbeb8cea8886d90778e634a145c51da735.
• Profile revision 2; kb/agentic-system-analyses/retained/greyhaven-ai--autocontext/overview.md: 2be2c7acb43199738ac14f86d8e5c6736e31658525a0969380d475f61fea1e66; artifact kb/agentic-system-analyses/retained/greyhaven-ai--autocontext/ARTIFACT.yaml: aa11208f3e92f5c8873c682b8711281a6e64acef6d65081d354d336ad75f9b39; source https://github.com/greyhaven-ai/autocontext at f72c154152397f54d55407508b153d28861888a5.
• Profile revision 2; kb/agentic-system-analyses/retained/skyllwt--autosci/overview.md: cf7ddb6d9aec502ce798890ce84ef7263dab43eb0f47eef25d00385049b8d009; artifact kb/agentic-system-analyses/retained/skyllwt--autosci/ARTIFACT.yaml: 68bb190ac5e68d2185847e0692261ea8ba4829eaa961bf6dcc0531cc0dc5743a; source https://github.com/skyllwt/AutoSci at f5c2fe009c6683b6aa5b668de759a0610c938aaf.
• Profile revision 2; kb/agentic-system-analyses/retained/basicmachines-co--basic-memory/overview.md: d9242a31b078770136dbb2f40dbb37784feaab2f22eff565ee37a11c5f4e16f0; artifact kb/agentic-system-analyses/retained/basicmachines-co--basic-memory/ARTIFACT.yaml: 448fe9f15d133c1fe461b309be467c6e541536cb770aa74b28b4dd14f336b4e5; source https://github.com/basicmachines-co/basic-memory at a72751c5d626380dedeed30f55f7a80ba3dc710d.
• Profile revision 2; kb/agentic-system-analyses/retained/dicklesworthstone--beads_rust/overview.md: cd1355d81ccaa365664957c6c23a865e82e06b9a51adaa20d8ad47768c925d21; artifact kb/agentic-system-analyses/retained/dicklesworthstone--beads_rust/ARTIFACT.yaml: e198b71ce8966d5b885fb547d61ed78b4bbd78bc6e76ed91c00a42f6e2f818e2; source https://github.com/Dicklesworthstone/beads_rust at 53f7a3d30aca8cab3c3e4abdefeade459e8a3089.
• Profile revision 2; kb/agentic-system-analyses/retained/beever-ai--beever-atlas/overview.md: 0101072edaa8a1abaa60a262ab499d892b8bf461c90b81472369f88609459bb4; artifact kb/agentic-system-analyses/retained/beever-ai--beever-atlas/ARTIFACT.yaml: ac433d28946b78193e1305f3f01af12c7005fb56f285998f64097f156aa59a00; source https://github.com/Beever-AI/beever-atlas at 7d791af27ef4aa19644cb385c7ca24619bc4f848.
• Profile revision 2; kb/agentic-system-analyses/retained/mpazik--binder/overview.md: 0adebe4242de60a6691fff3c4899f0b0a7399638cd2a93a3e7202c715fe3a5bb; artifact kb/agentic-system-analyses/retained/mpazik--binder/ARTIFACT.yaml: 28f70a39a871aba664e4dfc2ea065493048b80c6942e538b185cfcaefdb80758; source https://github.com/mpazik/binder at e4066d750ac5c6918706f44b12daa261b261106e.
• Profile revision 2; kb/agentic-system-analyses/retained/viharikanukollu--browzy.ai/overview.md: 842e94ccf8f61ecfe67e1ecc1a722476bbcc13bd7ece971757b6dc8b02a4c7c2; artifact kb/agentic-system-analyses/retained/viharikanukollu--browzy.ai/ARTIFACT.yaml: d660710030b7c5d27fb99a1e9dd03b12874567b8ada8cfa59a18db1f1c51498b; source https://github.com/VihariKanukollu/browzy.ai at 56c253042041ee2f483a5e9b824174d746891cf4.
• Profile revision 2; kb/agentic-system-analyses/retained/campfirein--byterover-cli/overview.md: 9eb6ce305cf3b4280cd7ef87285a3faed20744e179c8368899664cdc39a43c1d; artifact kb/agentic-system-analyses/retained/campfirein--byterover-cli/ARTIFACT.yaml: 34b2f8ac8c7ea7b5c6c1479a5bada39cf421dec91924c21a7dce00df56e357ad; source https://github.com/campfirein/byterover-cli at 1052ac1a5dd0fde4da8693d4712064f7876c269c.
• Profile revision 2; kb/agentic-system-analyses/retained/dicklesworthstone--cass_memory_system/overview.md: dffb244a21995b5f1418d613efb25751db05ae242967c25ee65b1841fab06f50; artifact kb/agentic-system-analyses/retained/dicklesworthstone--cass_memory_system/ARTIFACT.yaml: 2dcdb811124058cacf2dcde3b50b127dc826f2a8e603fb06ef92531b910c4729; source https://github.com/Dicklesworthstone/cass_memory_system at 8aca3bbb6732321dca041b32c155617784687897.
• Profile revision 2; kb/agentic-system-analyses/retained/atreiou--claude-context-guard/overview.md: 608ea2cfffc93057b2572996d7d439dd6d2f569a99d6a2ca16626b10463bae30; artifact kb/agentic-system-analyses/retained/atreiou--claude-context-guard/ARTIFACT.yaml: 044f9f2009026f0ad0dc36c3f5108d1daef58254a2819953357d712f3c4f4f33; source https://github.com/atreiou/claude-context-guard at efcd0a1815887483a862b32bfb3fbfd6aac79aa6.
• Profile revision 2; kb/agentic-system-analyses/retained/workflows/overview.md: a48fd9cafa804ad2712917e3715e074d68cbe669caea1bfecc73d451794f63cb; artifact kb/agentic-system-analyses/retained/workflows/ARTIFACT.yaml: fa3ae207ce9bbb9b10ee46f593e5ea5cee4e8c6a581b4e54c2f239e1608c05fd; source https://code.claude.com/docs/en/workflows at workflows-2026-10-14.
• Profile revision 2; kb/agentic-system-analyses/retained/agricidaniel--claude-obsidian/overview.md: 90a070fd19221d237373faa1ed4e6a22eebee82a87fa3e07c7f67fc06b0038ee; artifact kb/agentic-system-analyses/retained/agricidaniel--claude-obsidian/ARTIFACT.yaml: 4a5f47d874c1878bcc76b503e373957ae224b4bc5c45082589f2ec14b099a1c3; source https://github.com/AgriciDaniel/claude-obsidian at 32ac5a02c4e082e4a5628ca810776375e134708e.
• Profile revision 2; kb/agentic-system-analyses/retained/christophera--claude-workstream-kit/overview.md: 6dddc88635dc3efbac288b7cfe2433fb545345931a0d19659799c4502b8cba53; artifact kb/agentic-system-analyses/retained/christophera--claude-workstream-kit/ARTIFACT.yaml: 51e4288f82f2a93fa00c95fc8eee749b02e5420010994c4bea2f81c632e52a32; source https://github.com/ChristopherA/claude-workstream-kit at 15705c38a4fffb206b868b3e6bc5ffb4fd8b9d64.
• Profile revision 2; kb/agentic-system-analyses/retained/faltz009--closure-sdk/overview.md: b87201a4192725feb324e04f1ee479f46d4e65daaef240b875d8390b498b0c6c; artifact kb/agentic-system-analyses/retained/faltz009--closure-sdk/ARTIFACT.yaml: 125c18d0f05c298e80246d3bb3daf80d2fafe971ab44d35635067a76ffa0d128; source https://github.com/faltz009/Closure-SDK at 9735f48a6ddd00b6f429a91443489531b3a115ef.
• Profile revision 2; kb/agentic-system-analyses/retained/sebbsssss--cludebot/overview.md: ee5f7e42852c744c83e4889a948d39596939f4283319a9da92e2ca3be00e05ee; artifact kb/agentic-system-analyses/retained/sebbsssss--cludebot/ARTIFACT.yaml: 60aa311e50d0e19e1ab3878e9c3d035b3f7c349bad01a1cc79af6fc90a8b3322; source https://github.com/sebbsssss/cludebot at cecf234f10397b1e21dec2ff65ded864982f3564.
• Profile revision 2; kb/agentic-system-analyses/retained/cocoindex-io--cocoindex/overview.md: 5e6aa0254df11dd2f55c9ff9dedcab4491754f8eb264dd8c4f2325615e590cf6; artifact kb/agentic-system-analyses/retained/cocoindex-io--cocoindex/ARTIFACT.yaml: ab625e32e38f9eb69262fb54b0f4113622d1a4c60ca02f98ba15886af8447255; source https://github.com/cocoindex-io/cocoindex at 1639793b3602408a9cb220728fb30e306d433473.
• Profile revision 2; kb/agentic-system-analyses/retained/topoteretes--cognee/overview.md: 782351dae7b3d8021818123733aafba24718547590aadb1ef7ae3d98f75f08e9; artifact kb/agentic-system-analyses/retained/topoteretes--cognee/ARTIFACT.yaml: 6ea43b55baf253fe8eb49349736f0c8dc91003502c1047e08f511fe7af31c0c1; source https://github.com/topoteretes/cognee at f154b32a9e544dac73e7ddf2010327cf9988edcd.
• Profile revision 2; kb/agentic-system-analyses/retained/supermemoryai--company-brain/overview.md: 1e9ecdaf647e76f33e79c8176f99cdbe29047172caff0e778888a8c9d32e7641; artifact kb/agentic-system-analyses/retained/supermemoryai--company-brain/ARTIFACT.yaml: eabee5220a6cfb8234d42edfd56c186964f07a6e2639df2fbbd8144dc7802e6f; source https://github.com/supermemoryai/company-brain at ef8a45ef3aed97b32599c7d8007edefab6d14e9a.
• Profile revision 2; kb/agentic-system-analyses/retained/everyinc--compound-engineering-plugin/overview.md: d6d734acff70b7a3fe56c65a7f1cc71e82c3e3284b9668e2692a2dd06f186b2e; artifact kb/agentic-system-analyses/retained/everyinc--compound-engineering-plugin/ARTIFACT.yaml: 08972accd12da87c60f512bf41d9b3006beacddfa75c91d852beb1e1a026a5d6; source https://github.com/EveryInc/compound-engineering-plugin at 50532b00de523980010d97fe950965b35382a3b7.
• Profile revision 2; kb/agentic-system-analyses/retained/letta-ai--context-constitution/overview.md: 5e93ac78e423cf4bb38c0a19be44ec4f15e6af6d00af1d895ee5bf93feaf1725; artifact kb/agentic-system-analyses/retained/letta-ai--context-constitution/ARTIFACT.yaml: cebc390deac36ce442b484b2bc2b2c3187bf2acea56ac510192e8969905db421; source https://github.com/letta-ai/context-constitution at 964f30137ace9eade1aef112992ed6539b61bbb0.
• Profile revision 2; kb/agentic-system-analyses/retained/uziiuzair--continuity/overview.md: 8ea7ca5fc53311689cdcbc2eaa3442fc122c4ef04bec5eb8b5d398f58d9426fe; artifact kb/agentic-system-analyses/retained/uziiuzair--continuity/ARTIFACT.yaml: 3ce9146352f5a85daa2e056ff853df9155cbda884a80df000dc9ac0dd45cf90d; source https://github.com/uziiuzair/continuity at 4ca8f6b4108aa4494e3861ed33c8019dbd662c67.
• Profile revision 2; kb/agentic-system-analyses/retained/human-agent-society--coral/overview.md: fda9db6f2b7b5f39291dc024799f5d4cfdf443cd9e221043a47abd55bbd7881c; artifact kb/agentic-system-analyses/retained/human-agent-society--coral/ARTIFACT.yaml: 06a10aa0db92bd9d04885194fa8c3dd2cc01383cd7f1e1d5c3f74b94f821015b; source https://github.com/Human-Agent-Society/CORAL at 0123dfb939b35228cf2c1fde224cd0561e727408.
• Profile revision 2; kb/agentic-system-analyses/retained/abbacusgroup--cortex/overview.md: ab7904263770d2585748826137944adeb36b6a2ffb8833e23f85c2e011db4111; artifact kb/agentic-system-analyses/retained/abbacusgroup--cortex/ARTIFACT.yaml: 8b71da8a485833ffea4af81d02af62f5be8edbdfc8fe3fa92c4d390f196f2584; source https://github.com/abbacusgroup/cortex at a51c2209eb6d26812c19c97aaade3cb64ccfc23c.
• Profile revision 2; kb/agentic-system-analyses/retained/mozilla-ai--cq/overview.md: 3db14a26e0996ba95942fda9dbdb4c0c7a87e667bf4a3fabddf56858545906a4; artifact kb/agentic-system-analyses/retained/mozilla-ai--cq/ARTIFACT.yaml: 8d2fcd66aa8707273424c810f716e9203b9f7c90be7296a5b1a10666c7d415aa; source https://github.com/mozilla-ai/cq at 5765494a3840e5f5a7223c42101f41e85cb9d898.
• Profile revision 2; kb/agentic-system-analyses/retained/crewaiinc--crewai/overview.md: c2fc98ab8254eb47178ff9946e1df70ede234b7b86ab1ec29257dd3b776e13ba; artifact kb/agentic-system-analyses/retained/crewaiinc--crewai/ARTIFACT.yaml: af3805c06f70412b8f9f87e7f0dce2a9b4218e295d68801ba270dbae1cefaeaa; source https://github.com/crewAIInc/crewAI at 3da2d4e9932b5f2fa9aa73ea9245e0f8b5c154f8.
• Profile revision 2; kb/agentic-system-analyses/retained/decapodlabs--decapod/overview.md: 0f82b1a28067848483d9af423b1f8de9c557c12a2754fc106608f39e5f3c2850; artifact kb/agentic-system-analyses/retained/decapodlabs--decapod/ARTIFACT.yaml: 0176264acdc450c7977e1680a6550144470e5794a141a512d30cb971943d1869; source https://github.com/DecapodLabs/decapod at 067b8f55071e13dd103779998168b4c9ffd31a79.
• Profile revision 2; kb/agentic-system-analyses/retained/vshulcz--deja-vu/overview.md: 80ddf1c7129d53e65c5259e6ed49006fccabf7302b7f9082b18229251f9a0b23; artifact kb/agentic-system-analyses/retained/vshulcz--deja-vu/ARTIFACT.yaml: cb8d0b6a9a953203123aaa2a58ddc8455a9ca9d20d332e484a30bcc2fc8d07e2; source https://github.com/vshulcz/deja-vu at 4c08fb26ad8a8b8e656a120f06ca4175a7078148.
• Profile revision 2; kb/agentic-system-analyses/retained/markhuangai--dense-mem/overview.md: 21cbdefe01557426b13c1fd596d57d81e6221ae73714e76ee2889cbf1aaec061; artifact kb/agentic-system-analyses/retained/markhuangai--dense-mem/ARTIFACT.yaml: 760ea00c8225102e2eb1c89eb73fe1163ec288170bd60154b6a216096f5a518c; source https://github.com/markhuangai/dense-mem at bd736450454e237f1ab6d138f4ae9755916d73f6.
• Profile revision 2; kb/agentic-system-analyses/retained/jetxu-llm--docmason/overview.md: 22738319af66d3ae3c46b2252bd1ba4f77890dfa7d095fba1a88107cc2382fa9; artifact kb/agentic-system-analyses/retained/jetxu-llm--docmason/ARTIFACT.yaml: f832cb91cd31db729318d85c41910cf9918d19a706f553a5711f6491ddf56300; source https://github.com/JetXu-LLM/DocMason at f3970939593194a158ab7ea1d49995f06cbccc86.
• Profile revision 2; kb/agentic-system-analyses/retained/suzgunmirac--dynamic-cheatsheet/overview.md: 0019966a29c5e23c376f6d2a53bf68c27ac02bd23cf26df58d5cd26ccc32d2d0; artifact kb/agentic-system-analyses/retained/suzgunmirac--dynamic-cheatsheet/ARTIFACT.yaml: adc74cbcaffde604ae4df6294af2be3917d8db5a683171d544ce3b0611732645; source https://github.com/suzgunmirac/dynamic-cheatsheet at 5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9.
• Profile revision 2; kb/agentic-system-analyses/retained/tommasocerruti--eal-bench/overview.md: 91dbe4573529a81bd28729057a00a45a640718904262f2f1d19e14846a72247e; artifact kb/agentic-system-analyses/retained/tommasocerruti--eal-bench/ARTIFACT.yaml: a496014440dd8369548f5e1397e789033030e3950cd60845e9e6a4d6bf76ec78; source https://github.com/tommasocerruti/eal-bench at 51648690bc52d7a9c7ac080a2c67a784fe9d56cb.
• Profile revision 2; kb/agentic-system-analyses/retained/cuiyu-ai--ecdysis/overview.md: 26f388db44d82907ae36f28d183df31b028232f337cc49e1ead65f620941f897; artifact kb/agentic-system-analyses/retained/cuiyu-ai--ecdysis/ARTIFACT.yaml: a1d858a2316d73f01caaeb5fbc71d1a1afaa74a11df6202fc6adfd5eebe2b024; source https://github.com/cuiyu-ai/Ecdysis at cf93866d545b0974dbb0bc83b39c31fbbdeeecb8.
• Profile revision 2; kb/agentic-system-analyses/retained/alirezabbasi--echel/overview.md: 410e3e3ab4caa5675f9194ec82e25bb7f46678c031d1464ef68b356bf7679b03; artifact kb/agentic-system-analyses/retained/alirezabbasi--echel/ARTIFACT.yaml: 53f8d1dc491ebb71b63c25e9304d71caede91193e5cebec1cbc680f40130a18d; source https://github.com/alirezabbasi/echel at aee36de4d2d1c404fb36e9ad68cc0a920784ee61.
• Profile revision 2; kb/agentic-system-analyses/retained/psinetron--echoes-vault-opencode/overview.md: 7de9cb0f464985e89c8bf46b46b57ef91ddcbc1d6aaa57692671115e0e40b14e; artifact kb/agentic-system-analyses/retained/psinetron--echoes-vault-opencode/ARTIFACT.yaml: 464e014b4b432bec1c80c4e46f9021df8c1dc70eefaaddac5700b56596bb9b65; source https://github.com/psinetron/echoes-vault-opencode at 875c45aa45559529b87a7481b0c21e6dfd8c5eb0.
• Profile revision 2; kb/agentic-system-analyses/retained/mohammadmaso--echowiki/overview.md: 1e45be20712bd36a2b7750d98eab5e634bb01be50995ea6f912878da9595f389; artifact kb/agentic-system-analyses/retained/mohammadmaso--echowiki/ARTIFACT.yaml: 1fb41151b3d722115f2975faa550c77d4f661f60fe6134e09cf644e7fcb0de4e; source https://github.com/mohammadmaso/echowiki at 11446997423e88177dba62d6a3d5c1e8b4886c62.
• Profile revision 2; kb/agentic-system-analyses/retained/larikoz--eidetic/overview.md: 77fa52cdf600c260f868a75296bd6dbce2f7254478cc13c848014e7963eb100c; artifact kb/agentic-system-analyses/retained/larikoz--eidetic/ARTIFACT.yaml: a6bbe3727265fda7553691b63af40840308705137fd1b63b8a8b8687f439b0ae; source https://github.com/LARIkoz/eidetic at 67506b62fdd0944c243537c3cb10938f7343b6b2.
• Profile revision 2; kb/agentic-system-analyses/retained/eigenius--eigenius/overview.md: f61d68e9f702566d53762bd2c70fd4b8aceb8a978582e8a5cfbb9b534f53bd9b; artifact kb/agentic-system-analyses/retained/eigenius--eigenius/ARTIFACT.yaml: af41a35f70d90d41fd2e2434daaa47d2b235ab3387a38999ba46cc8786183eb0; source https://github.com/eigenius/eigenius at 836498ebd99733067d45aa6f7b09d4ff5035dd99.
• Profile revision 2; kb/agentic-system-analyses/retained/devwhodevs--engraph/overview.md: 86561343a1d63f5f0f66f86fea07dca503e0e15ce60278fd5c63ed3e0f9beabc; artifact kb/agentic-system-analyses/retained/devwhodevs--engraph/ARTIFACT.yaml: 04de40777115a44fa3c6ed27e99db04c08d6fa3b6ec074449748a4764c9dc05f; source https://github.com/devwhodevs/engraph at f9a95bc96accc792c02ee384d9e6bf768a88c8c8.
• Profile revision 2; kb/agentic-system-analyses/retained/our-ark--enoch/overview.md: 48e41d92194898700cd1fe8dd109195bf726cf046f103d43edac605f17c3b566; artifact kb/agentic-system-analyses/retained/our-ark--enoch/ARTIFACT.yaml: 69d4f301f2c468c3275febe79fbe53bd006cd0851094297fa40f512efde2bbf2; source https://github.com/our-ark/enoch at 0934ad59bd253adfffb8263c65e4bf0f4e2d7a9f.
• Profile revision 2; kb/agentic-system-analyses/retained/sbknana--equipa/overview.md: d41ec54edd5f3c890e55acf0538c7c59207fda387925cefbc00581acc935b694; artifact kb/agentic-system-analyses/retained/sbknana--equipa/ARTIFACT.yaml: 31c0aa3eaa5e9007ff3b2649188df28acd4e3371bf77c753156ce278fbf2db02; source https://github.com/sbknana/equipa at fa8f41a6948584ee53e804edbb08b7449d05cd14.
• Profile revision 2; kb/agentic-system-analyses/retained/ruc-datalab--evoontology/overview.md: 0c17ba9af8ff30a10fb52e7035520f7030b261edc271c166650ed89ab461c8d5; artifact kb/agentic-system-analyses/retained/ruc-datalab--evoontology/ARTIFACT.yaml: 0d461a64e7f29e7f9fe23f85920ce313f6ece77b1243691bf0d7bfef577431ac; source https://github.com/ruc-datalab/EvoOntology at c94389b764394379086ff25e50d0e0f4396167b4.
• Profile revision 2; kb/agentic-system-analyses/retained/exoharness--exo/overview.md: adfe591d19c5cd8b990d8ce388efe0b6eb834fb82dea6207c9c4e489003a4db1; artifact kb/agentic-system-analyses/retained/exoharness--exo/ARTIFACT.yaml: 044fc5b76dd84a40adda24251f933be497d9aefedaf3af969a043627b9584ebf; source https://github.com/exoharness/exo at 423d7d0f4394a1ddc9955d13c73c9f8a9c4d4e7e.
• Profile revision 2; kb/agentic-system-analyses/retained/cookiengineer--exocomp/overview.md: 503ed5ad827bf4b9125052523d6daadbc81b47e81dd32e9c7cc7f5cb1aa9f26b; artifact kb/agentic-system-analyses/retained/cookiengineer--exocomp/ARTIFACT.yaml: a135de34dacbeba580aa9d5587d18a5c9d288b95857fb150bea13cab570aced1; source https://github.com/cookiengineer/exocomp at b36f7ba1f3e2eee25bbcd2711db7d2d5d4824251.
• Profile revision 2; kb/agentic-system-analyses/retained/leaplabthu--expel/overview.md: ce38d356fe83cc42bacf801cc18befb6abc1d4c35e26a33729e308aa11943bf5; artifact kb/agentic-system-analyses/retained/leaplabthu--expel/ARTIFACT.yaml: 9e227f93fce1be1f574c0c6ce0347114e5d9f0d892bfe7788f9c270815534330; source https://github.com/LeapLabTHU/ExpeL at e41ec9a24823e7b560c561ab191441b56d9bcefc.
• Profile revision 2; kb/agentic-system-analyses/retained/trampoline-ai--fractal/overview.md: 4cc9464c005bd27743bc275d38220a7d54f053ef647ab1e551cd32d9159dbf03; artifact kb/agentic-system-analyses/retained/trampoline-ai--fractal/ARTIFACT.yaml: 23a0357f1b780a960d6d07c9679e6d022f536c8b0a93bfe690d899be54cc1d83; source https://github.com/trampoline-ai/fractal at 6a226e7d539e229b5c7cc9ffa3b05ab1c8f00385.
• Profile revision 2; kb/agentic-system-analyses/retained/from-agent-memory-to-portable-skills/overview.md: 190907bc824fd5f163fb6ba2a5c35a2055a94d34f09c7a82c5f0f56593409e82; artifact kb/agentic-system-analyses/retained/from-agent-memory-to-portable-skills/ARTIFACT.yaml: 9ca29001bd53c96550f60a25c43751c98bcedfca11956af15e73457d8b715c6e; source https://neo4j.com/blog/genai/from-agent-memory-to-portable-skills at neo4j-article-2026-10-18-99f91d4da959.
• Profile revision 2; kb/agentic-system-analyses/retained/ulyssestenn--funes/overview.md: 9efbd3a89526c30364a0fe0c04941eb39a3ec1e6d71040b2bbba06019da21498; artifact kb/agentic-system-analyses/retained/ulyssestenn--funes/ARTIFACT.yaml: 21c218b7b0e3ed1f4094ed8a0053ff56a3c2d0d96166aeba110dc9aca98e4f02; source https://github.com/ulyssestenn/funes at 3fb448b8e0d1ab5f6de3c6de61d8d556dab4ee11.
• Profile revision 2; kb/agentic-system-analyses/retained/garrytan--gbrain/overview.md: 24ed198a847680c994907be876db6b4055128d7e9fe47552daaf473bb41694c9; artifact kb/agentic-system-analyses/retained/garrytan--gbrain/ARTIFACT.yaml: fd3deedf9ec21974e7fe3ec47b4675ca91684c95e5a3c8a73d1a7e12fbb03b7d; source https://github.com/garrytan/gbrain at da434385572e0dc0d825649179aabed4107d6e60.
• Profile revision 2; kb/agentic-system-analyses/retained/bingreeky--gmemory/overview.md: 990b8d1f006f1e3e5a4cd350eb0206024f2a502031c79112ffad5b0cf3fbb9e9; artifact kb/agentic-system-analyses/retained/bingreeky--gmemory/ARTIFACT.yaml: fe55f29c053ae84242ad0c6ca0245548bc593ebe45a7938c0a616079157592c1; source https://github.com/bingreeky/GMemory at 7b581c51d993bd600df14691d101d7e601040cc6.
• Profile revision 2; kb/agentic-system-analyses/retained/skorokithakis--gnosis/overview.md: 08ee6d4013a60d906f32b809f94aff859de468b4e16d39015b27a581de33799e; artifact kb/agentic-system-analyses/retained/skorokithakis--gnosis/ARTIFACT.yaml: a1191bdc2af1782823e22a9a44e2147581d8e4ff87f4e8a73734532996548307; source https://github.com/skorokithakis/gnosis at 429528bf98e424b40837d2dbe3631d0fb42da6fd.
• Profile revision 2; kb/agentic-system-analyses/retained/getzep--graphiti/overview.md: b6187cb266927e84bb4a641a67d4e033faedd7d98e6676adb99b44799afd373a; artifact kb/agentic-system-analyses/retained/getzep--graphiti/ARTIFACT.yaml: 1423f4c0399121a5b2ddbbfcf19dd06c32b0dfb3efdcba4f513b8d7cfd15c65f; source https://github.com/getzep/graphiti at 964bb11d1e7020e3cc67fab244268b0736b6c219.
• Profile revision 2; kb/agentic-system-analyses/retained/context-labs--halo/overview.md: 4cca3b27d8e13fc639d205d352e8d3c2244106e13d9060d4a7ea0f61a1a38340; artifact kb/agentic-system-analyses/retained/context-labs--halo/ARTIFACT.yaml: b6bc3f1f1d2856c512d0b0d292eb1d3b985d60920e67fe5753a7eb0080f4a4e2; source https://github.com/context-labs/halo at b7f8509745d67b499b4e80efe20ea37c03426a74.
• Profile revision 2; kb/agentic-system-analyses/retained/nousresearch--hermes-agent/overview.md: a080b7140d19eb659f5ca4b009dc3b61af486675a3c1604f854b485a1d5e398f; artifact kb/agentic-system-analyses/retained/nousresearch--hermes-agent/ARTIFACT.yaml: a1c1c9c6d66287eecb408af1bea873f3a1a1f0a126eda4f970f1ebf45bd9bfcf; source https://github.com/nousresearch/hermes-agent at 0e427d788d253630d38a3a51cd76849e73cd6925.
• Profile revision 2; kb/agentic-system-analyses/retained/vectorize-io--hindsight/overview.md: fad59dace6f2c6d44bf18a2c410c7e33d46502281e9afaf50fc6264413d64dea; artifact kb/agentic-system-analyses/retained/vectorize-io--hindsight/ARTIFACT.yaml: 3e9e67f75bff2a28a6ffc32ceb894615058008641f960da1e24bc9d70fb38579; source https://github.com/vectorize-io/hindsight at 610c135b6f2d0d8b972da94146ae635943f36271.
• Profile revision 2; kb/agentic-system-analyses/retained/osu-nlp-group--hipporag/overview.md: 147db35fd80de9eaa4e23f1817424122836e1bb74b695f21b6d8e7f9f16c5595; artifact kb/agentic-system-analyses/retained/osu-nlp-group--hipporag/ARTIFACT.yaml: de44b02ef5c0f216be03b89ca8a0091275c6ed9569e82bc0d1778340e712fc7e; source https://github.com/OSU-NLP-Group/HippoRAG at d5c8329422e0a0b834a15874545cb6a74b4f9b26.
• Profile revision 2; kb/agentic-system-analyses/retained/ractive--hyalo/overview.md: 78c3163610d12f70b53b1d55ff4208198cc2d19443300e7d04ecd5d0a27ff5fd; artifact kb/agentic-system-analyses/retained/ractive--hyalo/ARTIFACT.yaml: cf390d8c799cb7ab15d5d92e64c26b416cdb1fc9d56dc8607e947e3983b20c84; source https://github.com/ractive/hyalo at 697f215fe8a90a5c760fceb686f2a321c9db7221.
• Profile revision 2; kb/agentic-system-analyses/retained/facebookresearch--hyperagents/overview.md: 4837193aac5ecf8b64243106ec0e25d875170c86c0b3c43e0862ad85afc50f10; artifact kb/agentic-system-analyses/retained/facebookresearch--hyperagents/ARTIFACT.yaml: c6d0d1e2328565277d91843426d5234648b1f63bbd972656f5194b2c4eec0abb; source https://github.com/facebookresearch/Hyperagents at 59a68f672dfb92c74aeb7e61535d776fb36e172d.
• Profile revision 2; kb/agentic-system-analyses/retained/fki198-94-pdf/overview.md: 5081c087d9d2285c06ce19b3a31491040af040e1593b2f74d3ad28a304143f95; artifact kb/agentic-system-analyses/retained/fki198-94-pdf/ARTIFACT.yaml: d84be1fef01799b209bbd93529db2afb6ff70b1e8f259eb0cf4747b64cee40bd; source https://people.idsia.ch/~juergen/fki198-94.pdf at fki198-94-revised-1995-01-31-captured-2026-10-14-text.
• Profile revision 2; kb/agentic-system-analyses/retained/jasonkneen--instinctual-memory/overview.md: 9399b2a03fa4ae0159e08937e2cc6cae98ac8efd21b6ed67759e0e56ceda49c4; artifact kb/agentic-system-analyses/retained/jasonkneen--instinctual-memory/ARTIFACT.yaml: ec04cc0b86052879dc738f9a737c02c1e98c3a6182b422008fb1ab6903f7c00d; source https://github.com/jasonkneen/instinctual-memory at 6acb13dc35765bf5ccfc87e445dd09c480f1c28a.
• Profile revision 2; kb/agentic-system-analyses/retained/shilren--interview-doc-agent/overview.md: 48daf382b47bc928ff05c10a8db04175585ba30bd88fdaa5d23a611921d3fc75; artifact kb/agentic-system-analyses/retained/shilren--interview-doc-agent/ARTIFACT.yaml: 768ae040f321b12f356cfcc1b6e98db7dd71fe7ed65b0cfa54a67778914ee4a2; source https://github.com/Shilren/interview-doc-agent at 6ebfe9aba989d7949dadea0fc4af6fb64892c6aa.
• Profile revision 2; kb/agentic-system-analyses/retained/noirblue--isaacclupus_mnemosyn_spec/overview.md: 3a10027a4fffc186fa2908cefa2477e4979494507728b29a750112246d0cca94; artifact kb/agentic-system-analyses/retained/noirblue--isaacclupus_mnemosyn_spec/ARTIFACT.yaml: 29f6cf530ccdf62f8bf5a7f5202cbde0658bce2af58b66db29ecdf313f90d66e; source https://github.com/noirblue/IsaacCLupus_mnemosyn_spec at 0fda14eb33d2fc01a4e3969a9ab487b22c4f9d92.
• Profile revision 2; kb/agentic-system-analyses/retained/jaz-lang--jaz-evals/overview.md: fb25e920c1a2dc75645e49a9b36f1fbbfa8a033e62c5dc78921c84156dc51ef7; artifact kb/agentic-system-analyses/retained/jaz-lang--jaz-evals/ARTIFACT.yaml: f8e9ba3ff3836e26ae130a2f3ab525c14ef214a44dd6d15c1cdf54cfc654e85b; source https://github.com/jaz-lang/jaz-evals at 2ffe7b5e6792a76fccb26f9ce188c88fb95ffa72.
• Profile revision 2; kb/agentic-system-analyses/retained/gen-verse--jepa-anything/overview.md: 2697b29ff7720e5091d464f7c2e8f680250aaaf184aa7879f783065af8680f59; artifact kb/agentic-system-analyses/retained/gen-verse--jepa-anything/ARTIFACT.yaml: bc57d17e6429686f73a53bd1ca84b169a6a1dc7ef9444cba2b05d4ffa5640506; source https://github.com/Gen-Verse/JEPA-Anything at c6e6c88f3ef75a4ce7acd660d6fa5779d995512c.
• Profile revision 2; kb/agentic-system-analyses/retained/microsoft--kblam/overview.md: c7093691e8a513383d6eaeaef523d762e14e169a13839da9e3c8601c7a55cc29; artifact kb/agentic-system-analyses/retained/microsoft--kblam/ARTIFACT.yaml: 2d969c81661ac87a1f2939128d41fef3feab761c5617936bd2fc0838a03ea4f3; source https://github.com/microsoft/KBLaM at 69683ad096e022b58854fd88c0c58968c7d8c470.
• Profile revision 2; kb/agentic-system-analyses/retained/kgaidev--kgai/overview.md: 6fbdeb8bc9b2e516147ba74fcd07b8e73020ab3812bb3e0f47273f4fddc36d24; artifact kb/agentic-system-analyses/retained/kgaidev--kgai/ARTIFACT.yaml: d47537a7fc0e1cc06b44d2688f5c89d06e8c373734681fa029e324838633bc14; source https://github.com/kgaidev/kgai at 0ac7f1fae4528805c1d5f504c1bfd6cc34ac953b.
• Profile revision 2; kb/agentic-system-analyses/retained/tuirk--kompl/overview.md: 365ff8e4f6b7608b246f62fb6ba99a22f04ee8a782d7c39cb2a9477dbac577e3; artifact kb/agentic-system-analyses/retained/tuirk--kompl/ARTIFACT.yaml: df1109f4bac34b734f83ce71d40944f90f96f2f820e79387650ade8bedd87bc6; source https://github.com/tuirk/Kompl at 17970c969b0c2576bf722840e0e2db3c6416a7aa.
• Profile revision 2; kb/agentic-system-analyses/retained/0xnyk--lacp/overview.md: 7b60b4579cb5a146e6d60dc52aa1694c7f8d033c4e54b6fd0c47d777efefa771; artifact kb/agentic-system-analyses/retained/0xnyk--lacp/ARTIFACT.yaml: f7194bae7af80f34073fbf0a7439145357a147fb1b545977da52f7669b355fcc; source https://github.com/0xNyk/lacp at e31f134b689afd6a71d94b0420682d1f528d6fa1.
• Profile revision 2; kb/agentic-system-analyses/retained/letta-ai--letta-code/overview.md: 9dcf71a1c6baa91d62eafadda488b9bc4b5e95746fed5135f23bb0ea261f1552; artifact kb/agentic-system-analyses/retained/letta-ai--letta-code/ARTIFACT.yaml: ad17a61b420acae420da29804d86a8ab19d4aaa44c70f401f85a45642fb40eba; source https://github.com/letta-ai/letta-code at 271f119dcf37d58649a18e6c733386b3c710a456.
• Profile revision 2; kb/agentic-system-analyses/retained/zli12321--lhtb/overview.md: 29a648057ea71e470af21f56015bf5725bc424c9316aa5bc0d5a5bb0d67d3fe2; artifact kb/agentic-system-analyses/retained/zli12321--lhtb/ARTIFACT.yaml: 9dc6086d63c5b1efd37a2701f46208a5321bc5a6bf0a0448a5f1812bb667a351; source https://github.com/zli12321/lhtb at 7ca0cacea9bf6a0694eaad7306bed794eaaaabce.
• Profile revision 2; kb/agentic-system-analyses/retained/gowtham0992--link/overview.md: ea51a61cfddea311d0e2038067ff11f583172bacad19b62df3ac594a020d8ca0; artifact kb/agentic-system-analyses/retained/gowtham0992--link/ARTIFACT.yaml: 435360a696174275474796718a29f2e0c56433d177359cb3903a7d82488f5c63; source https://github.com/gowtham0992/link at cc92f3632121b04063b24eaaf3a8c5429eaf2b57.
• Profile revision 2; kb/agentic-system-analyses/retained/asakin--llm-context-base/overview.md: 4073f72d7d059597878963091bc87f899dc837018dd39505a9159aad42a04cd2; artifact kb/agentic-system-analyses/retained/asakin--llm-context-base/ARTIFACT.yaml: cc73ff16d61391b69e5d6edda3cd8e1b3053b274989f8f18f396918a97470d2c; source https://github.com/asakin/llm-context-base at 6d01cba8e2c22f9ca2519c70073a05cd54378a8c.
• Profile revision 2; kb/agentic-system-analyses/retained/akash-r34--llm-project-wiki/overview.md: 39ed989ac99aeacef655d172ddef2e7ea7bcfa9b3af63eab89249b035fa0acb3; artifact kb/agentic-system-analyses/retained/akash-r34--llm-project-wiki/ARTIFACT.yaml: 348bbe1994a12da3110f56e882fe8b4dad39dd5d6f19f0a548d3c3832fe287a8; source https://github.com/akash-r34/llm-project-wiki at 1fdfba9129a46afd18aa68a83f12d6716ab694c6.
• Profile revision 2; kb/agentic-system-analyses/retained/cobusgreyling--llm-wiki/overview.md: 1e57555cccab93940049ef6b57f4bdce0c179fe113922f5929160360be5200c5; artifact kb/agentic-system-analyses/retained/cobusgreyling--llm-wiki/ARTIFACT.yaml: aa5a4c67fb9c7f49255c926ce277f9453abe158c1d6cf6417b79f04575b80e5d; source https://github.com/cobusgreyling/llm-wiki at a424e6875aa2d5736d9c0d03d7c28bdbe32892a5.
• Profile revision 2; kb/agentic-system-analyses/retained/pratiyush--llm-wiki/overview.md: fae1c390e6f95f5f20d2e334b57929d580a1c338655d0b7045a6967162eeca13; artifact kb/agentic-system-analyses/retained/pratiyush--llm-wiki/ARTIFACT.yaml: 5ca98e67f857b438592801cfaa0b186fe8579988971834cb1c04a0f9de1c036e; source https://github.com/Pratiyush/llm-wiki at b1088890ee0743810a92577aecad946c6b3eb2d2.
• Profile revision 2; kb/agentic-system-analyses/retained/kenhuangus--llm-wiki/overview.md: 371ab78989dcbbf8f17b0ff59956a4dd847dc9fd2c4f7db0225ae5633f0d6a29; artifact kb/agentic-system-analyses/retained/kenhuangus--llm-wiki/ARTIFACT.yaml: 10c31410e4033f33aea82617e49eb93bbb884ce24611405ec63e8b8435d455f0; source https://github.com/kenhuangus/llm-wiki at d2bd485921f6c7b6071afbb389f44ab8e4e83c69.
• Profile revision 2; kb/agentic-system-analyses/retained/nvk--llm-wiki/overview.md: 704ac08912a58b02e2a8c565319c5d070646f48b8fdccfb5340b322600ad50c5; artifact kb/agentic-system-analyses/retained/nvk--llm-wiki/ARTIFACT.yaml: a48c9f87b0a37203f2a329f1f59d724dd31599928e7ad8151ed8f323bd530682; source https://github.com/nvk/llm-wiki at 95a042c68998559305bc69ac8c482c4e1a415ad3.
• Profile revision 2; kb/agentic-system-analyses/retained/aevyra--llm-wiki-coordination/overview.md: 16a599c4a1d167039471f49f656c751ab5c90d4f77b38505022be9de0b556ace; artifact kb/agentic-system-analyses/retained/aevyra--llm-wiki-coordination/ARTIFACT.yaml: 1f34524471c313120792ca3b11bd1b24e9bb5ed10c0b43698611761c0b960275; source https://github.com/AEVYRA/llm-wiki-coordination at 126749634d1c8c2fd6141f59711c882c2d629699.
• Profile revision 2; kb/agentic-system-analyses/retained/electro-resonance--llm-wiki-mcp/overview.md: 7457ef3c4c0a4da211a52bd9c4869a65353b7a05029390e65208081f3d816ee2; artifact kb/agentic-system-analyses/retained/electro-resonance--llm-wiki-mcp/ARTIFACT.yaml: e6fbcf39cf40a3a57ce404c6754ef7659245be9a7fce5e5ddef28d3a1b2d4766; source https://github.com/Electro-resonance/LLM-WIKI-MCP at 25efb7631082fa3732723fee08e24d39b43311fa.
• Profile revision 2; kb/agentic-system-analyses/retained/vvvvvivekkk--llm-wiki-v3/overview.md: a65df3598acf4369246ce4540eebd82233605b3aa670403183f7af3e304f05fb; artifact kb/agentic-system-analyses/retained/vvvvvivekkk--llm-wiki-v3/ARTIFACT.yaml: fcdc5957cb1804a331c07c822fba1fe809630b4932a5c5c7263aaf53649981dd; source https://github.com/vvvvvivekkk/LLM-Wiki-v3 at 267541db838cbca757a112bd0cfa47fa61eb5fb2.
• Profile revision 2; kb/agentic-system-analyses/retained/clod--llmwiki-marimo/overview.md: 2f5b077884613a31f839065b62f4b58bb3f1cc6e350c9049666963b4cd6ed90b; artifact kb/agentic-system-analyses/retained/clod--llmwiki-marimo/ARTIFACT.yaml: 78afc0e09a802bff495983aad1a33a48b80c3ff1cea984a5fc9d5236142100a4; source https://github.com/Clod/llmwiki-marimo at 0df06b0499a6242a4a5edfb1495e1d55a3e3c8b7.
• Profile revision 2; kb/agentic-system-analyses/retained/mem0ai--mem0/overview.md: 0cecd07e11561a3f9b35df13dea85771516d63851726b56d7a0cfedb01442ea9; artifact kb/agentic-system-analyses/retained/mem0ai--mem0/ARTIFACT.yaml: 5d064c4100f3634646d08a826c54a7130611c2de3e7af2674a0c9d6cbef6e844; source https://github.com/mem0ai/mem0 at 0724291bd864cf807e88ee8981c8226644e8fb42.
• Profile revision 2; kb/agentic-system-analyses/retained/wastedcode--memex/overview.md: 41cafb17c638e17fe9f9a3e53b01e2d7b9670da264122c096a5efc7dd19b26f2; artifact kb/agentic-system-analyses/retained/wastedcode--memex/ARTIFACT.yaml: 8c01b4a25b69822259274c9f8dd08452c1c4f5d68c855649f61e82357df8c82a; source https://github.com/wastedcode/memex at 1e223b6045be31b70b1ab9e4b28dd403150d29e7.
• Profile revision 2; kb/agentic-system-analyses/retained/memorilabs--memori/overview.md: 027b0b1f3326ee0e98ad166bd3d01fab4c16c218169854deecd443d7ac3cf3c7; artifact kb/agentic-system-analyses/retained/memorilabs--memori/ARTIFACT.yaml: 03ab4d563da84bcff4b4b910b579ede41d2bdc395bbe670094cf8e215dd035f1; source https://github.com/MemoriLabs/Memori at 574b1ea3e876f100ef82c37817d603eb7e258e59.
• Profile revision 2; kb/agentic-system-analyses/retained/bai-lab--memoryos/overview.md: e0f08cd709952ca301b98cc2c642572b3191e00f55106d6aa78a59840180a084; artifact kb/agentic-system-analyses/retained/bai-lab--memoryos/ARTIFACT.yaml: 939d6250b2a98381a7ccae7d8fd8537b502deec1f962239c76750e25f1b05190; source https://github.com/BAI-LAB/MemoryOS at 587ed7755c7aed179965792830ff1b5ad9a6fa92.
• Profile revision 2; kb/agentic-system-analyses/retained/milla-jovovich--mempalace/overview.md: a6d36e4a6c7f011c29e62fc24ac3552514290ed9f951eded39f108c7be26bafc; artifact kb/agentic-system-analyses/retained/milla-jovovich--mempalace/ARTIFACT.yaml: 6990f81cb552b1d56b65f9e4d1a35dad5aaac05d3868655ff4185a523371b2eb; source https://github.com/milla-jovovich/mempalace at 1392e1810c3ca67776bd7fe3e198788a96dbd6d1.
• Profile revision 2; kb/agentic-system-analyses/retained/hereisswapnil--memwiki/overview.md: 31800a0e3771110d29977b3ee21395682bea391b656fc1065a50a002816652d0; artifact kb/agentic-system-analyses/retained/hereisswapnil--memwiki/ARTIFACT.yaml: f15e6631931d5086e380d64632b00acfba545bcde6db47a63627b3bfc08c18cb; source https://github.com/hereisSwapnil/memwiki at 8034a3da991ac2639b87875172a9572903ecf1d5.
• Profile revision 2; kb/agentic-system-analyses/retained/cloudllm-ai--mentisdb/overview.md: 54bd4a93f89d61aa12b5d2e951d703d64a6dd083329db44ad56d76c9efb7d34f; artifact kb/agentic-system-analyses/retained/cloudllm-ai--mentisdb/ARTIFACT.yaml: 8d48f0abd0f6d1728e17dc88283aa65bd468bde89ac9620d2ea3d84175be5259; source https://github.com/cloudllm-ai/mentisdb at 4c9b95041da68c5a26dece9f8776ff54b05b910c.
• Profile revision 2; kb/agentic-system-analyses/retained/khancold--merchantbench/overview.md: 06ec426f9345cf22c621e9b96849a38e5c1d2d02b170d4b2e481ab724c593e7a; artifact kb/agentic-system-analyses/retained/khancold--merchantbench/ARTIFACT.yaml: 30e579ea3d31ca7a9489598c386db713f77dbf27523364a365ff2314bcb71fbc; source https://github.com/khancold/MerchantBench at d9e5bd7519ab100193c21778aecfb59e30f24651.
• Profile revision 2; kb/agentic-system-analyses/retained/stanford-iris-lab--meta-harness/overview.md: 22c4450acdfc4ea626406321ac08157f0f7041496d2eeb5fcca49449f8384ce6; artifact kb/agentic-system-analyses/retained/stanford-iris-lab--meta-harness/ARTIFACT.yaml: 2441c2b7985ac28ea28c3ba8e43b11e568c3ba9909cb80af36b98054af772b99; source https://github.com/stanford-iris-lab/meta-harness at 8123ccabe2b19fa1123090b2e5f1bacc20963ce8.
• Profile revision 2; kb/agentic-system-analyses/retained/minnesotanlp--meta-n/overview.md: f02a6e80b629da173fa1954c38c5cacb039e38c27fb1add2e4379b6520e7e742; artifact kb/agentic-system-analyses/retained/minnesotanlp--meta-n/ARTIFACT.yaml: d109a8d936edde7a1bd8ccf3ca59053c7cc76311d2f4aa71f161a87075522596; source https://github.com/MinnesotaNLP/Meta-n at b7081843d3c7b0e0f418ca10aaf2ccbff856e7f8.
• Profile revision 2; kb/agentic-system-analyses/retained/aaronjmars--miroshark/overview.md: 983df2517ace39b30ec4f40bb16c750170bd722255df884443ab8de131a6e694; artifact kb/agentic-system-analyses/retained/aaronjmars--miroshark/ARTIFACT.yaml: 7676863d753d0e7dd7d776f85bf40c6e59d4ad17e634ade8778e560c9a825d3f; source https://github.com/aaronjmars/MiroShark at e95d3d8c2a2206b8b6695648e8504cf23436d3de.
• Profile revision 2; kb/agentic-system-analyses/retained/iquestlab--modularrsi/overview.md: db5fd9bb99ede394249371f9aeddce1893d1fc23bff441a99bd834a158363ef6; artifact kb/agentic-system-analyses/retained/iquestlab--modularrsi/ARTIFACT.yaml: ca0b3d2529d111f1104a250d970aaa246c0596f1bebb79ca5fcd011c816c6528; source https://github.com/IQuestLab/ModularRSI at b5c72c36b0d08ff93f00ee202a8fbdebe849dfb9.
• Profile revision 2; kb/agentic-system-analyses/retained/getnao--nao/overview.md: 8a61ece07e27f6d7cf2ded9e5244e41a43c52aaa75a8981c3a9ebc35503c31db; artifact kb/agentic-system-analyses/retained/getnao--nao/ARTIFACT.yaml: 70c7a3430431680eb137772a82d898e91208b95b301b43402771d2e96f4e7f70; source https://github.com/getnao/nao at 0085625a8fd22e4f404b6a3971d153f6b6699479.
• Profile revision 2; kb/agentic-system-analyses/retained/michaelliv--napkin/overview.md: 57a6ce86883be249d7e0bf770ad803211c4926c0fa50d1871c311c31be7d32ab; artifact kb/agentic-system-analyses/retained/michaelliv--napkin/ARTIFACT.yaml: 54279eca29b4b0add0039a606353184cb53779f122952ee17816e10b0db6e26b; source https://github.com/Michaelliv/napkin at 7582d6a46f5a11995956e60a59c41a5b242109f1.
• Profile revision 2; kb/agentic-system-analyses/retained/neovertex1--nuggets/overview.md: f93ed138ef021ed8fd3a9142f5e32c0b8536b91534e786323688c1de0d3022c8; artifact kb/agentic-system-analyses/retained/neovertex1--nuggets/ARTIFACT.yaml: fe49991761e2ecfc3cdc063ecf0c628f7a98363713e762277c66c029117dc2cf; source https://github.com/NeoVertex1/nuggets at 714cab8a3b1fb843aa98dfb51584d2c07a6739f3.
• Profile revision 2; kb/agentic-system-analyses/retained/jahala--o-o/overview.md: 09dc576357fb78765fc7742c74fcdf9bb136c5527412f4bb8e437d8a2ef70028; artifact kb/agentic-system-analyses/retained/jahala--o-o/ARTIFACT.yaml: d196163a06346da3faefd87e3b28e660c91f1caf8333732ebe2c0466e56af871; source https://github.com/jahala/o-o at b0d5063e37b4eafdbc23c1899a31f1836168b989.
• Profile revision 2; kb/agentic-system-analyses/retained/can1357--oh-my-pi/overview.md: e5b7e9fecab0b651cdc33976d2155e5a3b4d7ab7c1e8f000002a296bba531980; artifact kb/agentic-system-analyses/retained/can1357--oh-my-pi/ARTIFACT.yaml: f87005334083848d64ce9fbfa7863f6b63cff87068bb5f4c9902c1cb070e089c; source https://github.com/can1357/oh-my-pi at 2431e8a8716a4da73ae327a412f6d54c22f681e5.
• Profile revision 2; kb/agentic-system-analyses/retained/pumblus--okf-harness/overview.md: d194ef43784ed119933b1eef1c3ed4a2c789f1a78016819eb49f0ef930bd3c52; artifact kb/agentic-system-analyses/retained/pumblus--okf-harness/ARTIFACT.yaml: fb75519590f40f938abc1a8063a30ff609a66d1e3f94e0c791f7280ba03b5772; source https://github.com/pumblus/okf-harness at 0dd4f806c454a68088c5a3e64b4cbebbe2c216be.
• Profile revision 2; kb/agentic-system-analyses/retained/yazanabuashour--openclerk/overview.md: a521d8ae74398129e6e1ce95ec5969b935d7561bb0e1dd9fd1c3521d8f0b4e43; artifact kb/agentic-system-analyses/retained/yazanabuashour--openclerk/ARTIFACT.yaml: 142b3690661849956b3461844c39bd406ffa593d7f63df30e763641fb907c635; source https://github.com/yazanabuashour/openclerk at 092fb98d243c2d9d95867ad04618ebce88983552.
• Profile revision 2; kb/agentic-system-analyses/retained/opensage-agent--opensage/overview.md: b48a8b412ecb3f59359d529b04de7d6995bc44fc4ba4094a40cde46e68495db3; artifact kb/agentic-system-analyses/retained/opensage-agent--opensage/ARTIFACT.yaml: a6578c68728e4c10a59adf265b3f892d0222a6b8fb67f177a3cc8c55a7f775ab; source https://github.com/opensage-agent/OpenSage at 54d64702e18d195c8f0fed711c6db5dc1d62ab29.
• Profile revision 2; kb/agentic-system-analyses/retained/volcengine--openviking/overview.md: 24a883e6d319029380f5c9e6c14280e845d8a2ae261bb027b59e0225622e0300; artifact kb/agentic-system-analyses/retained/volcengine--openviking/ARTIFACT.yaml: a8a1926652236a41186ed7c10c6bb7a09754a7aa5c1b8bfa295e60ef5ef9920e; source https://github.com/volcengine/OpenViking at 5242d28d7792411e1a6eeecdc2b1d94d7fc98e6b.
• Profile revision 2; kb/agentic-system-analyses/retained/langchain-ai--openwiki/overview.md: 92bcf81b3125b5ce253d786da24da6bfc7bdcaf22f1c00578ceaf713948a69c4; artifact kb/agentic-system-analyses/retained/langchain-ai--openwiki/ARTIFACT.yaml: aecb0b5f4d3c539c85a5ff396973effbbf8421272982b8bcd8e96b285f0df7fc; source https://github.com/langchain-ai/openwiki at 7c469e59584bfac3f97f5c32a67604de46805999.
• Profile revision 2; kb/agentic-system-analyses/retained/7xuanlu--origin/overview.md: d693ee0c39150f483e868e2da6926b4f919e37e569086996c75eeafd05e274ab; artifact kb/agentic-system-analyses/retained/7xuanlu--origin/ARTIFACT.yaml: c0cd6799d58409d299bd6fa9cbae9225b9996ed3fdac6ecda6c408b171174385; source https://github.com/7xuanlu/origin at 61b462ea6adf019b3e27d3f58ac5a0fe055f76b1.
• Profile revision 2; kb/agentic-system-analyses/retained/os-copilot--os-copilot/overview.md: fa3160bfa6892dd342ec5a01c9b3e196dfcfddc9de9d7ea39dacfb3ddb8e864c; artifact kb/agentic-system-analyses/retained/os-copilot--os-copilot/ARTIFACT.yaml: c3a1ac1cc92d7d625e7bf42322e4996ba79ea470b23d6c4b49bd1395bddfff93; source https://github.com/OS-Copilot/OS-Copilot at f720af8807e49a92dda64572d2c6bc6c0ac7ee7e.
• Profile revision 2; kb/agentic-system-analyses/retained/vectifyai--pageindex/overview.md: bf56ced083624ed3cbc2bc6fce1f2475771d01125aa634319b8e9bbccbf2faf0; artifact kb/agentic-system-analyses/retained/vectifyai--pageindex/ARTIFACT.yaml: d3cd20ff471e98f637401d1ed21e5788321d3be2a4708abffe03a9b2e5597853; source https://github.com/VectifyAI/PageIndex at af92d9658a6a9c0ae31171b43198db28951af2f4.
• Profile revision 2; kb/agentic-system-analyses/retained/agno-agi--pal/overview.md: 67bb39181a156ad35c799643418debc30ab1fc54711865b375a3a8eb206560e7; artifact kb/agentic-system-analyses/retained/agno-agi--pal/ARTIFACT.yaml: 799613381135e3cdf8cc95652270879af3843f16c919a22be8654557705cc32e; source https://github.com/agno-agi/pal at 6516b8ede0c085e48f39f3bd04cb85b475a855dc.
• Profile revision 2; kb/agentic-system-analyses/retained/ghostwright--phantom/overview.md: 8f2d36388349422c878ed62a56d40c2a877525628435d8660e6b7adc99ed7c12; artifact kb/agentic-system-analyses/retained/ghostwright--phantom/ARTIFACT.yaml: 5c0840c80c8bdeb0cdce4d4a82919654eacb8b9ad41bc4d43109a540e94bbaa3; source https://github.com/ghostwright/phantom at f8c7ab42d885936ee54abc785528000260f4acc5.
• Profile revision 2; kb/agentic-system-analyses/retained/fitchmultz--pi-posthorse/overview.md: 171d651585bab6ef1fd7768c6d04b32a4e146fc57ee2b7cc55360b6c99bbd75f; artifact kb/agentic-system-analyses/retained/fitchmultz--pi-posthorse/ARTIFACT.yaml: a48c38cd1809c550b489808fc4d55c1f770fe9d1820b763efa9672c5553f95d9; source https://github.com/fitchmultz/pi-posthorse at 307ef6bf05bc8bd8d718725761ffdf2f2ef72539.
• Profile revision 2; kb/agentic-system-analyses/retained/mcollina--pi-self-learning/overview.md: c4bdbcfb471ec383849c0a9b049379c7db2dae16d5533dbff958742a863e3fc1; artifact kb/agentic-system-analyses/retained/mcollina--pi-self-learning/ARTIFACT.yaml: 903a6b625bf0d805a25d222dec241e857cf9bf8d0fe8677b009ff5e1832e0984; source https://github.com/mcollina/pi-self-learning at e0d76a93195b11b688ca71da8a8b5700c6541542.
• Profile revision 2; kb/agentic-system-analyses/retained/triblespace--playground/overview.md: 086823692c221825b63646e3373a7d9e7d94d6b5da8f349c3c9e6ccd8977ad9a; artifact kb/agentic-system-analyses/retained/triblespace--playground/ARTIFACT.yaml: f72e3be6a848a3f48dec899453f6d4c740c4c6ac23259f9deebd8f1ef1b50dce; source https://github.com/triblespace/playground at de28ac8fdda30c7a5c2ee099de089e65a685195f.
• Profile revision 2; kb/agentic-system-analyses/retained/tenequm--pond/overview.md: 7a17949d5f0764550a3c92d81f2214bc566335674ad0ac47c82d2d5e8315a4e8; artifact kb/agentic-system-analyses/retained/tenequm--pond/ARTIFACT.yaml: 577e8717f25958d04924eaae70b48b7344f4531b97312c4def1d64f87b4a381d; source https://github.com/tenequm/pond at 0c440e130faf44052d0d896a585e572090079f1f.
• Profile revision 2; kb/agentic-system-analyses/retained/primeintellect-ai--prime-agent/overview.md: 51a9a03ca790f69d92dfa996f513bc3b3f8a4ff1e3fcf9256ba72cbaa6e51d39; artifact kb/agentic-system-analyses/retained/primeintellect-ai--prime-agent/ARTIFACT.yaml: 60b1e629ad606631ec615e2ae649185b6db8b4ae1fb03bf7b523fece732fcd49; source https://github.com/PrimeIntellect-ai/prime-agent at 33da15a2c15c4e903c372f69ae8e1156d77d53da.
• Profile revision 2; kb/agentic-system-analyses/retained/henri-xyu02--primescientist/overview.md: 841f221b0e79ca133b8c1ce167013df7363ac76b6f9ddb6b8072d45a25bda657; artifact kb/agentic-system-analyses/retained/henri-xyu02--primescientist/ARTIFACT.yaml: 172d9bcc4b2afd3441ec1cdc095f138e5160b52bc6241c6095bc8ebd9ceb90e9; source https://github.com/henri-xyu02/PrimeScientist at 29971beac6f4f4b41309b1326762e1b83ceece98.
• Profile revision 2; kb/agentic-system-analyses/retained/prove2me--prove2me_workspace/overview.md: fbcb04e3e2610649395eed1397f2a7ae23865e54306304e1aa0968ccb3129d48; artifact kb/agentic-system-analyses/retained/prove2me--prove2me_workspace/ARTIFACT.yaml: 299bf835ee296bd7f3cc681ecaf3ed9608f9ed1a8971de336f9a80a69e011aa9; source https://github.com/prove2me/Prove2Me_workspace at 38e12f9d215cb103a0f9173ec230a614fcd93b61.
• Profile revision 2; kb/agentic-system-analyses/retained/anzal1--quicky-wiki/overview.md: b7de0f666a7dd1eb1838c0a66558c8a279419bf1f8aa959f21bb06fc6050ddf6; artifact kb/agentic-system-analyses/retained/anzal1--quicky-wiki/ARTIFACT.yaml: 1e72d48f631dedc3957b18f37533796ee0fffc8f328d88aabb5d70467494e0f9; source https://github.com/anzal1/quicky-wiki at cfeecbffce5b6d0f16533b258a1e88edd86f731e.
• Profile revision 2; kb/agentic-system-analyses/retained/niceirene--qwenpaw/overview.md: 7a1f9d54746b19cb0801854e330faac7fcabc930b8509ce8d0eb86c039426f44; artifact kb/agentic-system-analyses/retained/niceirene--qwenpaw/ARTIFACT.yaml: 48450a2600703af15cf1c5d04c3e16de6aa66718efea8e2d7f9c784d7fb95295; source https://github.com/niceirene/qwenpaw at 16b2e712d198166c9359e77055b13bf2668bf5ae.
• Profile revision 2; kb/agentic-system-analyses/retained/google-research--reasoning-bank/overview.md: 67ac081df99b99fd4649f9bd4cdefe317719c161d99acfe8651bf70ae606c650; artifact kb/agentic-system-analyses/retained/google-research--reasoning-bank/ARTIFACT.yaml: 968919ebbaaf768d4532cd3c33ac27a86352864da6c2df340aa09c074e558092; source https://github.com/google-research/reasoning-bank at ed80611788292ea739f1effd31f16c53823b8a0d.
• Profile revision 2; kb/agentic-system-analyses/retained/noahshinn--reflexion/overview.md: 49a0d7ef9e90cfab0926796c74da9b1d0ab3d67809c6b5869d798a4f25281961; artifact kb/agentic-system-analyses/retained/noahshinn--reflexion/ARTIFACT.yaml: bf3df4fa59ed377759416672ab7c139a922a03a637ea296795a46efab7dab8fe; source https://github.com/noahshinn/reflexion at 218cf0ef1df84b05ce379dd4a8e47f17766733a0.
• Profile revision 2; kb/agentic-system-analyses/retained/jubjub727--reframeweb/overview.md: 25c2b351dea96c3638deef84dea8b3fc5878e7d8d6cb26428728c1c59217dd8f; artifact kb/agentic-system-analyses/retained/jubjub727--reframeweb/ARTIFACT.yaml: df8889d4c4a8060f280c62fbc9bd99266b89faf4ba8a7bc2481fd3750b29075a; source https://github.com/jubjub727/ReframeWeb at 318987dcde9a016f8e3db819ed9a73abf6b3ad81.
• Profile revision 2; kb/agentic-system-analyses/retained/satyammistari--rem/overview.md: 5c84a05ab6fb16229555c2ac70083fa2924960a05ac781d49ef03e6fec9af1e4; artifact kb/agentic-system-analyses/retained/satyammistari--rem/ARTIFACT.yaml: 0e45470aeb4b227b6deb7877ab75e9cc4db0199c9c6cf5aa590506243d3a3dee; source https://github.com/satyammistari/REM at 935e8be0a1fca5b23dbabfe16c48b562c1cd24cc.
• Profile revision 2; kb/agentic-system-analyses/retained/aetherlabsai--rsiagent/overview.md: 33f95d4aa9450c89872f3bf70ab3356728d0d1706d24fd27ab666a1ca511cd02; artifact kb/agentic-system-analyses/retained/aetherlabsai--rsiagent/ARTIFACT.yaml: 7bb35f1fd2684ffdcd0762b7b53d3ffbac7abdea3e332820de1dfa4c45fffe73; source https://github.com/AetherLabsAI/RSIAgent at 4d7ef43770dda341fccd9a5743db92407687c615.
• Profile revision 2; kb/agentic-system-analyses/retained/amazon-science--sage/overview.md: 669f6d1a30cd3c9721c0e0d9e5157c2f31c89c945f9916af738b0182ce2826b4; artifact kb/agentic-system-analyses/retained/amazon-science--sage/ARTIFACT.yaml: 6f6d316bbf60d689ae2f234211cc5fdd332a1fa347c736b472cedb86ffc04553; source https://github.com/amazon-science/SAGE at 3c9244e82244abb1adc5467ee601a03ba0f433a0.
• Profile revision 2; kb/agentic-system-analyses/retained/l33tdawg--sage/overview.md: a657f4edfaf66f924ec324d91e979c2a04c038dcf3f17326ca815085e9030074; artifact kb/agentic-system-analyses/retained/l33tdawg--sage/ARTIFACT.yaml: e52bdc72ae32264bbb4376f7e5f97740b2fc277b77f2b3fa98b6cc5e64c763a8; source https://github.com/l33tdawg/sage at e2e0242a421afbd356ed135d7d754a35b0ad5f14.
• Profile revision 2; kb/agentic-system-analyses/retained/xoai--sage-wiki/overview.md: 8472bbd1876dabf6c04c62b08457b24b36a6287e52ce2a6a074f6c31875f85d1; artifact kb/agentic-system-analyses/retained/xoai--sage-wiki/ARTIFACT.yaml: db6ef74809788e33bc00bef443a1141d3b5e3f95f63fe8cde3681a23210f3b8b; source https://github.com/xoai/sage-wiki at d3ef3a1346405c82eac01b4d7c493c94c1c59169.
• Profile revision 2; kb/agentic-system-analyses/retained/nicobleh--secure-llm-wiki/overview.md: 82b4dc22fe1f927c3a20df1df500976e07673de9f7c763d3c1bf5518b7622760; artifact kb/agentic-system-analyses/retained/nicobleh--secure-llm-wiki/ARTIFACT.yaml: 73f299e7ae900c9e5d0deae2569e1a1fec1f251551e40a0b9b810680ad06e0d9; source https://github.com/nicobleh/secure-llm-wiki at dfbc60e37487b61e9a5cbce1271a13576a776d51.
• Profile revision 2; kb/agentic-system-analyses/retained/wj210--self-training-llm/overview.md: 6891f80e1a83f9f96de66682e523d1b23f4bbb1afe8cb44fe4c6db6fb663db7c; artifact kb/agentic-system-analyses/retained/wj210--self-training-llm/ARTIFACT.yaml: 482984bbaa1087e4d44953665493d3f63e02ae359ed65ca6778446ea449f3682; source https://github.com/wj210/Self-Training-LLM at 97839b29d0fd8bb474f5549fa3e9d6ca504732e0.
• Profile revision 2; kb/agentic-system-analyses/retained/open-research-institute--semantic-engine/overview.md: 343c4b97d0b6939053d55afe826172b8f7fcea681bde03afdd1bbeb92c425eb5; artifact kb/agentic-system-analyses/retained/open-research-institute--semantic-engine/ARTIFACT.yaml: 65e4c65f48a12eeabf89c137018767c96687a039e08d150b09b90b72bb84c0ee; source https://github.com/Open-Research-Institute/semantic-engine at 7e2481671cad01f0b4c89f2a1877d6733d7e552c.
• Profile revision 2; kb/agentic-system-analyses/retained/the-ai-alliance--semiont/overview.md: ab16af0efc7d4cab1c81396a17996fad42c3bb0e1d362125c8cb1d00ec8b6f99; artifact kb/agentic-system-analyses/retained/the-ai-alliance--semiont/ARTIFACT.yaml: 0be58dc0e4f341a4f3df477623735912eae37ce65ddfe9d833d89a4f32974b21; source https://github.com/The-AI-Alliance/semiont at a920067ac4ea63a57b1855b7d263485fbd22bab4.
• Profile revision 2; kb/agentic-system-analyses/retained/getsentry--skills/overview.md: dae8f1fe005042d19d1847f908944978f6527a293e42b4fb587bc1cdc02aa424; artifact kb/agentic-system-analyses/retained/getsentry--skills/ARTIFACT.yaml: fa0f932c44859f1983779927916d0fe89a66e35ed25302657fcbaa314aa804f5; source https://github.com/getsentry/skills at d18b7aa8ba878354e5c348310230e652f7690f9c.
• Profile revision 2; kb/agentic-system-analyses/retained/viperrcrypto--siftly/overview.md: ad6b763199980c76a7ae247941a3321e437afc84a48762bfdc4d0f04dc8a8d09; artifact kb/agentic-system-analyses/retained/viperrcrypto--siftly/ARTIFACT.yaml: 9a8de7382c1ec0e7c1dde0ea394ce750ec15d3788489419f51968c7756a9952a; source https://github.com/ViperrCrypto/Siftly at b25daa45b858f4be096b5c368671c34db4407d8e.
• Profile revision 2; kb/agentic-system-analyses/retained/sig-ai-app--sig-releases/overview.md: f08bd5d071377f0f47e9a7dcb1ebb1a5a83315d9f1ef3149875979e36a395329; artifact kb/agentic-system-analyses/retained/sig-ai-app--sig-releases/ARTIFACT.yaml: 3283a1ebedb8cc34979524db8b4dd3f1626238bef403ad2a845de0ab4b03f934; source https://github.com/sig-ai-app/sig-releases at 314bbea9d1cef911cea0f201eec66a57c2e59020.
• Profile revision 2; kb/agentic-system-analyses/retained/signet-ai--signetai/overview.md: 8fb4be55688e1b73c6a39634859a10a82dad1d28d5b9e1aab150e1001c44fc1f; artifact kb/agentic-system-analyses/retained/signet-ai--signetai/ARTIFACT.yaml: 4eb6367fe3884963d0c083b7c276bef866a089d4bcc8a1f0680bf1d7401bece9; source https://github.com/signet-ai/signetai at 12dba1ce753babda735f2ba7b6f150c6c0b7f95e.
• Profile revision 2; kb/agentic-system-analyses/retained/walter-mitty-pro--skilllift/overview.md: ba44c1c387ffd949527ff8f82e069a20a338b71761230eef4f535159f7d3a116; artifact kb/agentic-system-analyses/retained/walter-mitty-pro--skilllift/ARTIFACT.yaml: 0cf21bdd9cafc4250bc329a15734ed3682635104f17a2d0fb1a0657ca192f294; source https://github.com/walter-mitty-pro/SkillLift at 599358b4d4c4a27c0e004df8228ab93026600653.
• Profile revision 2; kb/agentic-system-analyses/retained/luna-prompts--skillnote/overview.md: 401876170e5d53528c25971cda6436016a2b430f9e3b2e316075ad6741dc8ccc; artifact kb/agentic-system-analyses/retained/luna-prompts--skillnote/ARTIFACT.yaml: 32bbbef3f072571983687caf0f53a9640cb54f8a308e14392f9fa48e48930244; source https://github.com/luna-prompts/SkillNote at 7ea2fc7a8435af64d1c080d8765c0eabba1121ea.
• Profile revision 2; kb/agentic-system-analyses/retained/aiming-lab--skillrl/overview.md: 90e0ea011f1a39591ef7160efc8186e82273fb7d509e0683b4c000d6e9e0738e; artifact kb/agentic-system-analyses/retained/aiming-lab--skillrl/ARTIFACT.yaml: 9771fa25ae76ce88c6efdfc88fb4e126b1abe0bd3fac23ae510761d4acb361f4; source https://github.com/aiming-lab/SkillRL at 4d984d455b9b02f0bad283cbe9ab3d66a7e5b89d.
• Profile revision 2; kb/agentic-system-analyses/retained/osu-nlp-group--skillweaver/overview.md: 220acad5c246b5f4340587bef58e373c8479b76bfceb7f281b0b99a431b38c90; artifact kb/agentic-system-analyses/retained/osu-nlp-group--skillweaver/ARTIFACT.yaml: eaa5d002869db0b44af3b0e137a79fc39652ac4fdbeb3fb6f4ce69d799d29f6e; source https://github.com/OSU-NLP-Group/SkillWeaver at f2a63d65d0f6ff46ac30e817cede8797f8f25b97.
• Profile revision 2; kb/agentic-system-analyses/retained/zjunlp--skillx/overview.md: 9a5f85ca9ad6d471c33c5ccb0c595fe3675d767821a6fb69f726c81ebe8455ba; artifact kb/agentic-system-analyses/retained/zjunlp--skillx/ARTIFACT.yaml: f9dc66741f36be94a78ed1373f9053aed3089a4efccfefcd23b6d63e4d21f0ff; source https://github.com/zjunlp/SkillX at b0047c848dbbcf019363d6d56eef8c558f69aefc.
• Profile revision 2; kb/agentic-system-analyses/retained/deepak-bhardwaj-ps--smriti-mcp/overview.md: dc221dd53a1aa4a7e04ae8c1d51bcde30d869f734f3fa4170dd5430b6e0f2dba; artifact kb/agentic-system-analyses/retained/deepak-bhardwaj-ps--smriti-mcp/ARTIFACT.yaml: 35c1e0a7706816f3cdf0680d2eb18f39e87639ed10482467c684828d0e359731; source https://github.com/Deepak-Bhardwaj-PS/smriti-mcp at 8264bf9103477d9521e57b27c81dcc2e943a3230.
• Profile revision 2; kb/agentic-system-analyses/retained/nvlabs--sol-pi/overview.md: 1e3a2324e9381c780cc4226ca06132e490692706ba33a33a48c307556813fc5c; artifact kb/agentic-system-analyses/retained/nvlabs--sol-pi/ARTIFACT.yaml: 5d66c4d3818a8dbf67203d5c748edd308997fb8df8e33fce711f480610119c34; source https://github.com/NVlabs/SoL-Pi at 41fa7e0e230bedc5e639aed2e7b9cc5698726db1.
• Profile revision 2; kb/agentic-system-analyses/retained/spacedriveapp--spacebot/overview.md: 38e7237eb294e624090523dcb0317d42ca95d89a95cd8234b5d315b227e4f004; artifact kb/agentic-system-analyses/retained/spacedriveapp--spacebot/ARTIFACT.yaml: 484a042cfd4306978e187ddbdabd5d9b6c98230c4032cc48cd7248bd85c1e5a7; source https://github.com/spacedriveapp/spacebot at 5c2ee47b397e09c956dc15000ea576ecd28e92bb.
• Profile revision 2; kb/agentic-system-analyses/retained/yogirk--sparks/overview.md: bac64229b165376412f1d81f6f71a6d6b9e1623342f4e214d861fdcdf6eb838b; artifact kb/agentic-system-analyses/retained/yogirk--sparks/ARTIFACT.yaml: b63432733774919defd8d579d6ee42b1cf197157b9dffaf03806c474a39847b4; source https://github.com/YogiRK/sparks at 49abddc848ea55f78518bfbbbe04c3d37a5f2705.
• Profile revision 2; kb/agentic-system-analyses/retained/alash3al--stash/overview.md: 1b8c9207ad5db33491f2148850e062aaec275e7dc76dc65a117d980aa0784d78; artifact kb/agentic-system-analyses/retained/alash3al--stash/ARTIFACT.yaml: 2d2914eb866ce9c8aaee7bccdf3f48bdd0ae358fed850bdb7eb6868e1f6082a5; source https://github.com/alash3al/stash at d34ed430c4348ef741ef95146b488aea8616f4fb.
• Profile revision 2; kb/agentic-system-analyses/retained/supermemoryai--supermemory/overview.md: b4dbad4a29ef46cd8dd199b3321a788318c07c33aecdf0b65714f0ebb7a577cc; artifact kb/agentic-system-analyses/retained/supermemoryai--supermemory/ARTIFACT.yaml: f743a44aa45d6ff27ad6c4c161c25f7f06238ad2e6f39e268a96a0f14d960821; source https://github.com/supermemoryai/supermemory at f81fad00e4bfc42610159cf6bd7bdb5fc4ca7cbe.
• Profile revision 2; kb/agentic-system-analyses/retained/swamp-club--swamp/overview.md: fe6432fbeb2537025777b7820cf0a1ce7ca8e3fbf3c10f8be6a17f35d90adc09; artifact kb/agentic-system-analyses/retained/swamp-club--swamp/ARTIFACT.yaml: d7060856022b3c9e0c3e11b3be4f09ab7a193441a945fd2b69d4f7ccac2734e5; source https://github.com/swamp-club/swamp at 380e3c307164501c7d705f7707a864f4612b2785.
• Profile revision 2; kb/agentic-system-analyses/retained/lamm-mit--swarmworld/overview.md: 5117d80d7429fe6f03603c289b49e61a0f40c1d5bc37c7f90adabc2dbadfb193; artifact kb/agentic-system-analyses/retained/lamm-mit--swarmworld/ARTIFACT.yaml: 05cffed3d02ef96834acfe0d3b34e3ce3592bd7e2d30228411f62045c9114e91; source https://github.com/lamm-mit/SwarmWorld at 6af7ae9fa36d98b07b0492cf139658e8af1f6eab.
• Profile revision 2; kb/agentic-system-analyses/retained/appcuarium--synapptic/overview.md: 3f86130681c65e9a85a593c3db930ce4b7173ff6ebf1b018973a83ab0d4d9b2e; artifact kb/agentic-system-analyses/retained/appcuarium--synapptic/ARTIFACT.yaml: 6a89387e5536dd6e3eafcc61c0041fed26dd06d3ff6b62932c74df5d30d9dbdd; source https://github.com/appcuarium/synapptic at bcde026770daa5cd9a8ca18ca7182b0aeaa87d1f.
• Profile revision 2; kb/agentic-system-analyses/retained/axoviq-ai--synthadoc/overview.md: 074011fdde5d4d6918e252c432f384169432b17bde3187096c69fcfb269c07a2; artifact kb/agentic-system-analyses/retained/axoviq-ai--synthadoc/ARTIFACT.yaml: 4621d6a86efe5b5f54d6d8c74e8143b85e62bba7c0b03852988d4647b00e4fa1; source https://github.com/axoviq-ai/synthadoc at b80d9388a3013e8051baf2854c7250942c46e91a.
• Profile revision 2; kb/agentic-system-analyses/retained/kytmanov--synto/overview.md: 2e7f35ee664fb322e4786623d8dfae81e57ff8372b8b390687dea6969cfd9188; artifact kb/agentic-system-analyses/retained/kytmanov--synto/ARTIFACT.yaml: 8d6c3de62c3a5f28792b250b2290faa249621e663e0d3fbb3886d71acaa59c54; source https://github.com/kytmanov/synto at 91ac6a3681df0552b6c9baef159439c7ebd780c9.
• Profile revision 2; kb/agentic-system-analyses/retained/rociiu--talorys/overview.md: 26333dbc36fda310f39076ff5b58a1bd947d6d943ca512432b623365e1cf7f7d; artifact kb/agentic-system-analyses/retained/rociiu--talorys/ARTIFACT.yaml: 9eb4d9bde902f40d0ce2b9302f4f2fc16ae5b0cad95fa318acadf8a2177afc92; source https://github.com/rociiu/talorys at c5da440010402ed4c1b9789ae22eab96a8f0f379.
• Profile revision 2; kb/agentic-system-analyses/retained/clavia-labs--tardigrade/overview.md: 6faf7471bfb653a4d00f0621b0c08b5a842a113d5074a9263c7b04013d540e92; artifact kb/agentic-system-analyses/retained/clavia-labs--tardigrade/ARTIFACT.yaml: 868d5bf3a06fd476d21c0fac6a3eb2c144f06c080f5dd35228cfbec90cd0cd91; source https://github.com/clavia-labs/tardigrade at 1b9f1e5760d1d4d4a63eb1f87af9e5a2341de7e8.
• Profile revision 2; kb/agentic-system-analyses/retained/serverless-dna--tendril/overview.md: bee90c9f895444119b697760c9909f23e0266405f74e0cdb1ab8978c8c7f98bb; artifact kb/agentic-system-analyses/retained/serverless-dna--tendril/ARTIFACT.yaml: 6b56f25854226e7738fce2d3622101a3c707d7dc82d718be06248b119f841bf0; source https://github.com/serverless-dna/tendril at e671a4143d28de68289efd81580002041bb4cb6a.
• Profile revision 2; kb/agentic-system-analyses/retained/rejot-dev--thalo/overview.md: 5ff173dac649dbdebd047e9d64244a240ac65a22c8a8f560e1cccb82cc95619e; artifact kb/agentic-system-analyses/retained/rejot-dev--thalo/ARTIFACT.yaml: ba119d67e693ef41bda3efefe6d93cb1f2019709bfd142aff834d89ac79846d9; source https://github.com/rejot-dev/thalo at cdb9aae983e6bc0b75eff1606bc99b088c3aebff.
• Profile revision 2; kb/agentic-system-analyses/retained/badwally--theknowledge/overview.md: 5c4c9090bb7fe9344309eccefca4bc5f94a525cd703cd47295b0d22c6d28f982; artifact kb/agentic-system-analyses/retained/badwally--theknowledge/ARTIFACT.yaml: d1dad164680c9afc6b10fa1c530297c2efc91a0a19acc3301813eea2dcce0505; source https://github.com/BadWally/TheKnowledge at aca6e29e69c6a72b697b8178841d03e6296ee3fc.
• Profile revision 2; kb/agentic-system-analyses/retained/refactoringhq--tolaria/overview.md: e021b55172de202e712766e7d4bd88b4ca96235253086ac55bcdcd265962d61e; artifact kb/agentic-system-analyses/retained/refactoringhq--tolaria/ARTIFACT.yaml: a93e9d5f36b1347b4dc8443aba66bdd4055acb6c231844cf5cc26ed920c31ac5; source https://github.com/refactoringhq/tolaria at 48f2a70844c1e8005a7aa5f1499c66f7678360de.
• Profile revision 2; kb/agentic-system-analyses/retained/arrmlet--tracecraft/overview.md: 3b0c12093e80cb348de8eb9ddb951a7e9c97b2449f9b7ba477e46a3d168d2157; artifact kb/agentic-system-analyses/retained/arrmlet--tracecraft/ARTIFACT.yaml: ca2c75f8ea1e98fb2523dc01c1d1eb3ed64bf9485646b29885f424f339fa3a4f; source https://github.com/arrmlet/tracecraft at 97c6b7447f8e696f0116357d9f86bb0ece25e5b1.
• Profile revision 2; kb/agentic-system-analyses/retained/2603-10600v1/overview.md: f22ccddee5acd7d24af0783eff153b4d836f4ccf682628c16265675ec2c67522; artifact kb/agentic-system-analyses/retained/2603-10600v1/ARTIFACT.yaml: 2f1b566fe633b5907013b677e1263140f8e572bf3223f63c4d7042a3a93c0200; source https://arxiv.org/html/2603.10600v1 at arxiv-2603.10600v1-html-captured-2026-10-14.
• Profile revision 2; kb/agentic-system-analyses/retained/virtual-context--virtual-context/overview.md: 75a8e85edb68f6ee45aacf6be6e5063d7cb5f4ed095e959626c9642001995689; artifact kb/agentic-system-analyses/retained/virtual-context--virtual-context/ARTIFACT.yaml: 3178fc27dd319844d7406c56d5704f8bdbddc101376b32fe5e618986eb4398da; source https://github.com/virtual-context/virtual-context at 74ccc570c40bb551820fcd330c42e0dfc0588f05.
• Profile revision 2; kb/agentic-system-analyses/retained/veniveci--vlm-wiki/overview.md: 42aea7c359e15c87a83e62c724bf7d3fa62ccf2cc897b419069d3e313af215dc; artifact kb/agentic-system-analyses/retained/veniveci--vlm-wiki/ARTIFACT.yaml: 299c240c99e22bc102b9f102737c47007f3d1c1c7f83c822b8086b6690e3354c; source https://github.com/veniVeci/VLM-wiki at 503c43f98b5d1e21c74df384fdffd3e0652573b3.
• Profile revision 2; kb/agentic-system-analyses/retained/voidenhq--voiden/overview.md: 23c0e462fa7a8cb2abf6199fb45f76f0aa69f4d9e6e352699fd6d00e460677c9; artifact kb/agentic-system-analyses/retained/voidenhq--voiden/ARTIFACT.yaml: 9fdb3a33709b76ead1a9e7fa62ae7f89e73ca732bcc82b5cee07825eef76ef7c; source https://github.com/voidenhq/voiden at d7a0d05c40acb6cc483000643c700f17aecba0ef.
• Profile revision 2; kb/agentic-system-analyses/retained/minedojo--voyager/overview.md: 888f5a6b9e5ce040ae8fb5af9e35e7766e78af7a3eedb4d5f67c8ce700c888e9; artifact kb/agentic-system-analyses/retained/minedojo--voyager/ARTIFACT.yaml: 8945164b8997d2c4a642a741eb43aae6c71ce7191bdf4f20563bb02c058c2f06; source https://github.com/MineDojo/Voyager at 55e45a880755d0c8c66ca7fb5fe7962ac8974f89.
• Profile revision 2; kb/agentic-system-analyses/retained/tencent--weknora/overview.md: 6e95d68207c573cbb197612fa732257265d2e223ec1b1e64f6fd8edc56fd6afc; artifact kb/agentic-system-analyses/retained/tencent--weknora/ARTIFACT.yaml: addb8db1b26afb41cd22e70a93af952dc9aee195efa51804aae407becd2c74d2; source https://github.com/Tencent/WeKnora at 234aa948db29c10f5ed4b7939285c7b3711c5d4c.
• Profile revision 2; kb/agentic-system-analyses/retained/2608-27454/overview.md: 4f1ebde65f822a7582b0b3800c27c9c4402b7a430754847203e93b13540deaa8; artifact kb/agentic-system-analyses/retained/2608-27454/ARTIFACT.yaml: b9a9f7aa822cf356890c826dcce5695f25c9d99269763660a40948b9d1a68b87; source https://arxiv.org/abs/2608.27454 at arxiv-2608.27454v1-text-2026-10-15.
• Profile revision 2; kb/agentic-system-analyses/retained/stahl-g--wikiskill/overview.md: 0f39d983c3aac9aa6ada5d37c5c870862addedb87e9b44aef21628a28c922caf; artifact kb/agentic-system-analyses/retained/stahl-g--wikiskill/ARTIFACT.yaml: f6f84759d6d8cb0d19f53eb5a2f1294571e42bc8ec648892e942d401a03a294f; source https://github.com/Stahl-G/WikiSkill at da6984caca09d5f81b2336cf3067b59c0c18bb1e.
• Profile revision 2; kb/agentic-system-analyses/retained/nex-crm--wuphf/overview.md: c14cebf882b6133315b7b3f6bc9451125b4b910ac6c9dac184ab29cd08f743ea; artifact kb/agentic-system-analyses/retained/nex-crm--wuphf/ARTIFACT.yaml: 96f3121179ac29cfd6a6e2542e3b1c7ee222ff5081a2b942fd7a8f0133f62821; source https://github.com/nex-crm/wuphf at 88023d4dac7491b6e71dbc1e86d274c435b9cb2c.
• Profile revision 2; kb/agentic-system-analyses/retained/hu-xiaobai--xmemory/overview.md: 4b129bda6b8ebf456be5b20f762032de0ee6cb3b9a30b9f05b8a3c5b8812a944; artifact kb/agentic-system-analyses/retained/hu-xiaobai--xmemory/ARTIFACT.yaml: 5cec4aa10578233cfb710dc78ce8026f8d2c6f39a5087f467d0b8ca9c45b3c48; source https://github.com/hu-xiaobai/xMemory at 1a4f9dd8abd7f2da86bfe6afe4293a1ea1c63608.
• Profile revision 2; kb/agentic-system-analyses/retained/amanhij--zikkaron/overview.md: 8916ad57c59c24efb44db4f4ad7e21196af466b7d5032be4e3b7103a5996112a; artifact kb/agentic-system-analyses/retained/amanhij--zikkaron/ARTIFACT.yaml: 603416b5652c1406fb2b87f83f4b09e381bb27fb1b2d257a64c03c18fa97945d; source https://github.com/amanhij/zikkaron at dda34a5d903d04ecb5517af214d437873c833302.