Instinctual Memory epistemic report

Type: agentic-system-analyses/types/agentic-system-epistemic-report.md

Source-and-claim boundary

The question is whether converting sessions and instruction files into sourced facts, maintaining them, and delivering them to agents establishes warranted knowledge or only operationally usable candidates. Scope and external exclusions are see boundary; evidence identities are see Source register. All source inspection below is SRC-1 at the frozen commit. Implementation is not observed operation. No target execution was performed.

This overlay assesses acquisition, rule and model extraction, quote checks, publication, duplicate and lastingness judgments, corrections, withdrawal, retrieval ranking, host delivery, instruction writeback, retrieval evaluation, index derivation and timing measurement. Supplied runtime findings remain provisional evidence for paths not independently reread here. No memory or workflow member was supplied. Available support files not independently read are not inaccessible evidence.

Entry point or operation	Source path	Coverage and evidence basis	Limit
CLI dispatch, ingest and backfill	SRC-1 src/cli.rs	Direct dispatch read; acquisition mechanics see RT-RTE-journal-consolidation, supplied account	Adapter fidelity and real source truth not measured
Consolidate: rules, LLM, filters and publication	SRC-1 src/consolidate.rs	Direct reads of batch, rules, prompt, proposal handling, placement and quote normalization; EPI-RTE-rules-extraction, EPI-RTE-llm-proposal, EPI-RTE-candidate-gate; container see RT-RTE-journal-consolidation	Provider compliance and semantic preservation unresolved
Immediate remember, correct, forget	SRC-1 src/change.rs, src/ops.rs, src/cli.rs, src/repo.rs, src/validate.rs	Direct shared dispatch read; mechanism and alternatives supplied by RT-RTE-immediate-change	No independent source truth test or observed caller approval
Search, read and history, CLI/MCP/HTTP/shell	SRC-1 src/mcp.rs, src/ops.rs, src/search/rerank.rs	Direct registered tools and shared dispatch read; selection/fallback supplied by RT-RTE-memory-retrieval	Ranker internals and actual host reliance uninspected
Hook start, prompt, end and sync	SRC-1 src/hook.rs, src/setup.rs	CLI dispatch read; callbacks supplied by RT-RTE-host-hooks	Host scheduling, callback completion and activation unobserved
Writeback	SRC-1 src/cli.rs	Dispatch read; formatting, force and source exclusions supplied by RT-RTE-instruction-writeback	Actual instruction-file loading unobserved
Tidy preview/apply/keep	SRC-1 src/tidy.rs	Direct judgment, filtering and apply read; publication see RT-RTE-tidy-maintenance	No source evidence accompanies lastingness judgment
Erase	SRC-1 src/erasure.rs	Supplied RT-RTE-erasure-maintenance	Sequential effects and external copies prevent universal withdrawal guarantee
Evaluate and bench	SRC-1 src/evaluation.rs, src/cli.rs	Direct scoring and timing reads; RT-RTE-search-evaluation, EPI-RTE-search-benchmark	No supplied result, authenticated holdout or controlled causal experiment
Index rebuild	SRC-1 src/index.rs, src/cli.rs	Direct full index implementation; EPI-RTE-index-rebuild	Counts reflect stored Active statuses, not truth or endorsement
Tasks	SRC-1 src/ops.rs, src/mcp.rs	Direct registered dispatch; supplied RT-RTE-task-update	Titles are operational directives; individual truth-apt assertions not determinable without actual content
Setup and model selection	SRC-1 src/setup.rs, src/search/laya.rs	Supplied RT-RTE-integration-admission	No quality test gates component replacement
Init, formats, status, transport administration	SRC-1 src/cli.rs, src/ops.rs	Direct dispatch; store/status account supplied in runtime	No distinct claim acceptance; not an exhaustive audit of metadata accuracy

Public warrant claims are EPI-CLM-sourced-facts, EPI-CLM-lasting-facts and see RT-CLM-validated-publication. Missing real stores, event/proposal traces and provider responses prevent deciding which observed statements were copied, paraphrased, inferred or invented. Missing source authentication prevents asserting source truth. Missing host traces prevent activation claims. Missing rubric provenance and comparative results prevent retrieval-effectiveness or causal-benefit claims. Missing pinned weights and provider implementation prevent distributed-parametric process or training conclusions. These gaps do not erase the implemented candidate checks or wired later consumption.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-proposal	Name, location, preference, project rule or architecture statement; natural-language candidate plus symbolic routing and evidence fields	Durable candidate	SRC-1 src/consolidate.rs	Plausible standing statement need not follow from its quote
EPI-OBJ-source-ref	Quoted source assertion; source ID/role and link are symbolic	Source lineage	SRC-1 src/consolidate.rs	Attribution preserved; source truth and semantic entailment unknown
EPI-OBJ-fact	Stored assertion; separately, symbolic status/kind/timestamps govern eligibility	Curated fact	SRC-1 src/consolidate.rs	ExplicitAssertion/Active is not evidence-consuming truth acceptance
RT-OBJ-mem-skill	None as a truth evaluator; source-native procedure	Host review/use guidance: see record	SRC-1 skills/mem/SKILL.md	Host compliance unresolved
RT-OBJ-tasks	Not determinable for actual titles; task directives need not be truth-apt	Versioned tasks: see record	SRC-1 src/ops.rs	Version admission does not validate content
RT-CMP-llm-config	Opaque numerical model state; truth-apt localized content not determinable	Extraction and cleanup: see record	SRC-1 src/consolidate.rs	Requested semantic judgments not evidence of reliable execution
RT-CMP-jev-reranker	Opaque state; relevance scores are query/candidate judgments, not source verification	Ranking: see record	SRC-1 src/jev.rs	No independent truth license
RT-CMP-laya-reranker	Opaque state; cross-encoder output ranks candidates	Ranking: see record	SRC-1 src/search/laya.rs	No pinned weight digest or measured benefit

Raw journal assertions, cleanup findings, rubric checks and index counts are addressed by their route records rather than creating duplicate container objects. New objects describe proposal, lineage and eligibility parts absent from the supplied object inventory; their identities and forms are in Shared records.

Authority-route ledger

Each row is a consequential function, not a second declaration of a supplied route. Linked functions sharing a container retain independent epistemic assessments. Activation, timing, possible results and generic ownership are see the cited record; local license and mismatch assessments are this overlay's findings. No observed operation is established.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
RT-RTE-journal-consolidation	content transformation	implemented	EPI-OBJ-source-ref	truth-apt transformation: acquisition/import	Session/file text to journal and extractor input	Adapter/source role selection; not source authentication	see route	see route	Retains attributed text	Source warrant unknown; imported, not newly discovered truths	Makes eligible inputs available	see route	SRC-1 src/cli.rs	EPI-CLM-sourced-facts	none	Fidelity depends on adapters and original origin
EPI-RTE-rules-extraction	content transformation	implemented	EPI-OBJ-proposal	truth-apt transformation: indeterminate	Regex match to standing assertion	Fixed lexical patterns on User text	see route	see route	Proposes without semantic adjudication	Conditional paraphrase under intended reading; counterexample below	Proposals can become active facts	see route	SRC-1 src/consolidate.rs	EPI-CLM-lasting-facts	Regex matching is not a lastingness test	Quotation, negation and always/never distinction can change meaning
EPI-RTE-llm-proposal	content transformation	implemented	EPI-OBJ-proposal	truth-apt transformation: indeterminate	Bounded event text to statement/evidence/lasting	Chat model instructed to extract, not guess	see route	see route	Produces candidates and self-assessed lasting flag	No established entailment or independent semantic warrant	Feeds software gate	see route	SRC-1 src/consolidate.rs	EPI-CLM-lasting-facts	Requested judgment is not verified compliance	Copies, reshaping, entailed inference or conjecture require actual linked responses to distinguish
EPI-RTE-candidate-gate	check/evidence production	implemented	EPI-OBJ-proposal, EPI-OBJ-source-ref	no content change	Quote occurrence, flag, statement bounds, usable routing	Deterministic normalized ordered-fragment matching; model flag consumed as given	see route	see route	Enforcing eligibility gate	Licenses normalized quote occurrence in cited event, not entailment or source truth	Rejects malformed/unmatched candidates	see route	SRC-1 src/consolidate.rs	EPI-CLM-lasting-facts	Quote and statement are checked differently	Evidence fragments below eight characters are discarded; no statement-to-evidence semantic check
RT-RTE-journal-consolidation	disposition/acceptance	implemented	EPI-OBJ-fact	no content change	Candidate placement and publication	Duplicate/suppression gates, DomainValidator and CAS	see route	see route	Operational admission	No evidence-consuming acceptance of produced ampliative claim against truth criterion	Writes explicit assertions as Active	see route	SRC-1 src/consolidate.rs	RT-CLM-validated-publication	Structural validation is not semantic acceptance	Accepted event disposition can coexist with placement dropping a proposal
RT-RTE-journal-consolidation	retention	implemented	EPI-OBJ-fact, EPI-OBJ-source-ref	no content change	Facts, sources, checkpoint and dispositions	Git publication after candidate checks	see route	see route	Durable snapshot/reload	Preserves attribution and status, not added truth warrant	Later retrieval and consolidation can rely operationally on retained state	see route	SRC-1 src/consolidate.rs	EPI-CLM-sourced-facts	Retention is not lifecycle integration after epistemic acceptance	Linked evidence is useful lineage without verified entailment
RT-RTE-journal-consolidation	lineage/freshness/recovery	implemented	EPI-OBJ-fact	no content change	Old file-source versions to superseded status	Source ID version comparison, not independent truth check	see route	see route	Eligibility changes	Freshness relative to imported file version only	Removes old-source facts from active consumption	see route	SRC-1 src/consolidate.rs	EPI-CLM-lasting-facts	Changed file does not prove previous claim false	No semantic equivalence/contradiction test on changed file
RT-RTE-immediate-change	content transformation	implemented	EPI-OBJ-fact	truth-apt transformation: acquisition/import	Caller statement to explicit assertion/correction	Caller chooses content; software checks structure	see route	see route	Operationally admitted assertion	Source warrant unknown; no extraction quote gate	Adds/replaces/suppresses facts	see route	SRC-1 src/change.rs	RT-CLM-validated-publication	Correction label not a correctness test	Human authorship/approval unresolved; replay shortcut may report new input without applying it
RT-RTE-immediate-change	disposition/acceptance	implemented	EPI-OBJ-fact	no content change	Structural admission and targeted supersession/retraction	Request checks, validator/CAS, caller target	see route	see route	Enforcing current-read eligibility	Operational acceptance only	Remember/correct/forget current fact state	see route	SRC-1 src/change.rs	RT-CLM-validated-publication	Soft forget does not withdraw raw source assertions	Raw journal search and prior delivered context remain separate
RT-RTE-tidy-maintenance	check/evidence production	implemented	EPI-OBJ-fact	no content change	Duplicate statements and session-only usefulness	Word containment/Jaccard; optional LLM over IDs/statements	see route	see route	Findings only without apply	Lexical similarity and requested temporal usefulness judgments, not truth checks	Proposes removals	see route	SRC-1 src/tidy.rs	EPI-CLM-lasting-facts	Longest is not necessarily semantically most complete	Contradictory assertions can share almost all words; sources not given to judge
RT-RTE-tidy-maintenance	disposition/acceptance	implemented	EPI-OBJ-fact	no content change	Findings to suppression/retraction	Apply flag, keep exclusions, validator/CAS	see route	see route	Enforcing current fact eligibility	No positive truth acceptance of surviving assertions	Withdraws selected facts from future fact reads	see route	SRC-1 src/tidy.rs	none	Skill review is not executable approval; apply recomputes	Survivor status does not certify truth or lastingness
RT-RTE-memory-retrieval	operational admission/selection/consumption	implemented	EPI-OBJ-fact, raw events in see route	no content change	Query-relative hits/ranks	Lexical shortlist and optional JEV/local scorer	see route	see route	Ranking and truncated delivery	Query relevance, not verification; retrieved raw assertions retain unknown warrant	Determines supplied context	see RT-BAP-host-context	SRC-1 src/ops.rs	none	Ranking is not endorsement	Shortlisting/filtering/fallback constrain consumer visibility
RT-RTE-host-hooks	operational admission/selection/consumption	implemented	EPI-OBJ-fact	no content change	Preferences/matching facts to host context	Fact status, lexical overlap and delivery caps	see route	see route	Advisory unsolicited context	Hook warns to check against code if changed	Can shape later host action if interpreted	see RT-BAP-host-context	SRC-1 src/hook.rs	none	Delivery not activation	Already delivered context has no enforced revocation
RT-RTE-instruction-writeback	content transformation	implemented	EPI-OBJ-fact	truth-apt transformation: non-ampliative reshaping	Selected statements to managed instruction block	Fixed status/source/project selection and formatting	see route	see route	Persistent context placement	Copies claims, adds no truth warrant	Makes statements available through later instruction loading	see RT-BAP-instruction-file	SRC-1 src/cli.rs	EPI-CLM-sourced-facts	Channel may grant practical influence without verification	Host loading external; force and backfill feedback remain bounded by supplied route
RT-RTE-erasure-maintenance	disposition/acceptance	implemented	EPI-OBJ-fact, EPI-OBJ-source-ref	no content change	Caller-targeted content withdrawal	Target lookup and sequential store/history scrub	see route	see route	Destructive removal	Does not assess falsehood; removes lineage as well as assertions	Changes future fact/journal/history visibility if stages finish	see route	SRC-1 src/erasure.rs	none	Deletion completion flag precedes later effects	Partial failure/external copies prevent universal withdrawal
RT-RTE-search-evaluation	check/evidence production	implemented	Returned hits and caller rubric in see route	no content change	Expected IDs and required/forbidden text in top hits	Deterministic caller-supplied oracle, bounded retrieval cases	see route	see route	Report-only comparison	Pass means rubric satisfaction, not true answers, explanation validity or transfer	No automatic model/content promotion	Caller via report; informational, current evaluation	SRC-1 src/evaluation.rs	none	Held-out description is not authenticated independence	Empty expected/string lists permit vacuous per-case checks; snapshots/providers may vary between runs
EPI-RTE-index-rebuild	content transformation	implemented	Entity metadata and Active statuses in see route	truth-apt transformation: entailed derivation	Active counts and entity listing	Deterministic filtering/counting within stored-status domain	see route	see route	Published derived metadata	Correct only relative to parsed snapshot and predicate; no truth status implied	Supplies bounded navigation metadata	see route	SRC-1 src/index.rs	none	none	Counts are not suppression-aware truth counts; unparseable entities omitted
EPI-RTE-search-benchmark	check/evidence production	implemented	Timing observations in see route	no content change	Duration of selected search calls	Instant elapsed time and repeated call summaries	see route	see route	Report-only measurement	Selected-process latency, not semantic quality or component causality	No implemented promotion/update	Caller via stdout; informational, current benchmark	SRC-1 src/cli.rs	none	none	No actual run or environment controls supplied
RT-RTE-task-update	operational admission/selection/consumption	implemented	RT-OBJ-tasks	non-truth-apt policy/content update: task directive appended	Version-matched caller task	Lock and expected-version check	see route	see route	Durable task admission	No content truth license; actual titles could additionally carry assertions	Later host sees task title	see route	SRC-1 src/ops.rs	none	none	No task execution or improvement established
RT-RTE-integration-admission	behavior/policy adaptation	implemented	Components and installed skill in see route	non-truth-apt policy/content update: caller configuration	Host integration/model choice	Caller settings, supported names and load checks	see route	see route	Installs configuration	No measured semantic quality admission	Changes future available callbacks/inference machinery	see route	SRC-1 src/setup.rs	none	Configuration is not evidence-responsive self-improvement	Parameter updates not performed by inspected client; provider internals uninspected

No lifecycle-integration-after-truth-acceptance witness is established by these routes. Evidence links, publication, supersession and retrieval implement useful ledger functions without establishing the type's evidence-consuming acceptance of a produced ampliative claim. This is a bounded assessment of inspected/supplied routes, not an assertion about inaccessible model internals or every possible caller assertion.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-sourced-facts	Curated sourced facts with later search/read/writeback	SRC-1 README.md; doctrine/design	see claim	EPI-RTE-rules-extraction, EPI-RTE-llm-proposal, RT-RTE-journal-consolidation, RT-RTE-instruction-writeback	None supplied	No answer improvement experiment	Durable assertions with source linkage and consumption wiring	Sourced is not source-verified; direct ops assertions have no source refs
EPI-CLM-lasting-facts	Only durable facts with matching evidence survive extraction	SRC-1 README.md, src/consolidate.rs; doctrine/design	see claim	EPI-RTE-llm-proposal, EPI-RTE-candidate-gate, EPI-RTE-rules-extraction, RT-RTE-tidy-maintenance	None supplied	No independent temporal test or source-entailment adjudication	LLM route requests lastingness and enforces flag/quote checks; cleanup can revise eligibility	Rules have no equivalent lasting flag; word matches can misrepresent source intent
RT-CLM-validated-publication	Validated changes and no concurrent overwrite	SRC-1 README.md; doctrine/design	see claim	RT-RTE-journal-consolidation, RT-RTE-immediate-change	None supplied	Requires cooperating publisher/Git; custom-validator and other file effects differ	Structural validation/CAS, not truth evaluation	Replay short-circuit and validator limitations see supplied records; not superseded here

The evaluator's module calls itself held-out evaluation, but its inspected contract authenticates neither authorship nor independence of the caller rubric. A passing expected-ID/string check is evidence about returned hit composition only (SRC-1 src/evaluation.rs; see RT-RTE-search-evaluation).

Bounded conclusion

Instinctual Memory acquires externally supplied assertions, retains their attributed fragments and operational statuses, and later supplies selected content through search, hooks and instruction files. Those are independently supported capacities. Source warrant remains unknown: source linkage is not source authentication, and source occurrence is not semantic support for a rewritten statement. The LLM route requests extraction without guesses, but actual linked outputs are absent, so its content relation remains indeterminate rather than being classified automatically as conjecture or entailment.

The software quote gate licenses ordered normalized fragment presence. The model's lasting flag licenses operational admission under the product's temporal-usefulness policy; no independent temporal or truth oracle checks that flag. Rules are deterministic but can discard distinctions in source meaning. Direct assertions and corrections bypass extraction checks. None of these facts means that every admitted assertion is false or that the model cannot extract correctly; it means structural eligibility does not supply the missing claim warrant.

Tidy separately evaluates lexical duplication and model-judged session-only usefulness, then optionally suppresses candidates. Keep exclusions provide veto, while human review is a host procedure and apply recomputes findings. Surviving claims are not thereby accepted as true. Forget changes current fact eligibility, not raw source assertions; erase attempts broader sequential removal and can leave partially completed effects. Prior host context and external copies lie beyond enforced withdrawal.

Retrieval relevance judgments have ranking force, not verification force. Hook and writeback delivery can influence a host over later sessions, but no host trace demonstrates activation or benefit. Index counts are entailed within the stored Active-status domain, not evidence of factual correctness. Retrieval evaluation checks a caller rubric; timing measures selected searches. Both return information without an automatic consumer that promotes a model, revises knowledge or adapts policy.

Retention and explicit source attachment are established; acceptance and lifecycle integration of produced ampliative claims are not established by the assessed routes. Trace-fed writes and later reads afford behavioral adaptation without proving improved capacity. Static instructions, maintenance and caller configuration do not establish autonomous iterative criticism of the system's own method. Provider learning/reflection remains uninspected; no single conclusion about these independent properties follows from quote checks, curation or parameter opacity.

Shared records

Operative objects

EPI-OBJ-proposal

Label: Extraction proposal

Source-native identity: Proposal, with LLM input represented by LlmFact. Natural-language statement/evidence supply claim content; symbolic entity/predicate/ref/lasting fields control placement and admission. Storage substrate: transient Rust values/model JSON, before durable Fact publication. These objects are parts of the supplied consolidation route, not replacements for its full loop; no supplied object counterpart was declared. SRC-1 src/consolidate.rs, implementation.

/// A fact the extractor wants to record, before it becomes a [Fact]. --- src/consolidate.rs

EPI-OBJ-source-ref

Label: Fact source reference

Source-native identity: SourceRef as constructed by Extraction::place. Natural-language evidence is retained alongside symbolic event ID and role in entity YAML. Storage substrate: published entity file in bare Git. Distinct from proposal and stored statement: its consumer can identify source text without verifying the statement. Closest supplied route is RT-RTE-journal-consolidation; no source-reference object counterpart in supplied inventory. SRC-1 src/consolidate.rs, implementation.

sources: vec![SourceRef { event_id: event.event_id.clone(), role: fact_role(event.role), evidence: p.evidence, }], --- src/consolidate.rs

EPI-OBJ-fact

Label: Published explicit assertion

Source-native identity: Fact constructed in Extraction::place. Natural-language statement is interpreted by a host; symbolic kind/status/timestamps and ID determine software eligibility and references. Storage substrate: entity YAML/Markdown in Git snapshots. Proposal is transient, SourceRef carries lineage, and Fact carries published content and eligibility. Closest supplied record is RT-RTE-journal-consolidation, not an existing Fact object declaration. SRC-1 src/consolidate.rs, implementation.

statement: p.statement, kind: FactKind::ExplicitAssertion, status: FactStatus::Active, --- src/consolidate.rs

Routes

EPI-RTE-rules-extraction

Label: Rules extraction Part of: RT-RTE-journal-consolidation

Owner/endpoints: rules_extract takes bounded events and existing/blocked state, proposes name/location/preferences, and passes each Proposal to place. Trigger: batch chooses rules explicitly or Auto without LLM configuration. Software regexes propose and decide; caller selects mode, not per-fact content. No semantic answer oracle; open extraction mode. Guidance is fixed patterns, not requested lastingness model evaluation. Publication/rollback follow the parent route. Placement can veto suppression and near duplicates; ordinary matching cannot distinguish source assertion from reported speech. SRC-1 src/consolidate.rs, implementation.

Concrete code deduction: User input please always use tabs yields statement User requested: use tabs and evidence please always use tabs; please never use tabs yields the identical statement, predicate and entity, but opposite original directive. The capture omits always/never/don't, and the statement includes only cap[1]. The fixed entity/predicate/statement digest then identifies the same fact, with no sign bit distinguishing polarity. Input The example says my name is Alice also matches the name regex despite not asserting the actual user's name. Thus deterministic execution does not imply warranted semantic preservation. Normal literal name/location declarations may be faithful reshaping; across eligible User inputs, remaining alternatives include reshaping, entailed interpretation or unsupported attribution. Without contextual checks, relation is indeterminate. These are deductions from inspected code, not target runs.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Extraction with proposals placed, counts, accepted event sequences or quarantine reasons.
• Later read-back: Published statements and sources reach subsequent consolidation/retrieval/hooks/writeback through the parent route.
• Delegated visibility: No provider called; later readers see admitted statements and sources, not the original regex decision trace.
• Selection predicate: User role and matching name/live/preference pattern, then blocked/duplicate/space checks.
• Invalidation or expiry: Controls and parent supersession/withdrawal alter later eligibility.
• Activation or effect: Candidate generation/placement wired; host behavior and semantic quality unobserved.
• Evidence limits: Contextual interpretation is not checked; source assertions are not authenticated.

let pref_re = Regex::new(r"(?i)please (?:always|never|don't)\s+([a-z][^.!?]+)").unwrap(); --- src/consolidate.rs

statement: format!("User requested: {}", cap[1].trim()), --- src/consolidate.rs

EPI-RTE-llm-proposal

Label: LLM extraction proposal Part of: RT-RTE-journal-consolidation

Owner/endpoints: llm_extract selects event text, sends bounded batches and known entity summaries to RT-CMP-llm-config, receives LlmFact JSON and forwards in-batch refs to candidate gate. Guidance asks for standing facts useful next month, excludes session requests/guesses, and prioritizes distinct standing rules in memory files. Trigger and mode: configured LLM extraction of accumulated events, not held-out experiment. Model proposes statement and lastingness; software can veto; no human content approval or independent truth oracle. Errors abort rather than publish partially ready batch. Parent owns final publication/recovery. SRC-1 src/consolidate.rs, implementation and doctrine/design. Relation is indeterminate without linked response evidence: the prompt asks acquisition/reshaping, but neither guarantees output equivalence nor establishes non-entailment required for a conjecture classification.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Parsed candidate facts with event refs, lasting flag, entity, statement and evidence.
• Later read-back: Admitted output persists as facts; later batches receive existing-entity summary and downstream consumers receive statements.
• Delegated visibility: Provider receives selected user/Note text, project and known-entity context; other roles excluded.
• Selection predicate: Note text or filtered User text; bounded/chunked batches; output event ref must belong to batch.
• Invalidation or expiry: Failed gates/drop conditions prevent placement; later controls/freshness affect published facts.
• Activation or effect: Model request and proposal pipeline wired; actual compliant extraction unobserved.
• Evidence limits: Provider processing and weights uninspected; actual output lineage absent.

The test for every candidate: would a new teammate, starting fresh next month, need to know this? If it only describes what was happening in that session, it is not a fact. --- src/consolidate.rs

memory_file events are curated project instructions (AGENTS.md, CLAUDE.md, README.md): record every distinct standing rule, convention, architecture fact, and command in them, one fact each; skip only examples, history, and prose that states nothing."#; --- src/consolidate.rs

EPI-RTE-candidate-gate

Label: LLM candidate eligibility gate Part of: RT-RTE-journal-consolidation

Owner/endpoints: check_llm_fact consumes LlmFact/event/existing entities and returns Proposal or reason; quote_found checks evidence occurrence. It does not itself decide publication. Target/domain: normalized cited-event fragment occurrence, minimum evidence length, statement bounds, lasting flag and usable entity/predicate. Fixed program decides eligibility, model supplies flag/content; no human veto in this substep and no semantic oracle. Trigger: each in-batch LLM candidate; open extraction. Failed checks drop candidates. Parent placement/publication owns retention/recovery. SRC-1 src/consolidate.rs, implementation.

Concrete deduction: source Alpha beta gamma and evidence Alpha beta gamma...NO normalize to one qualifying fragment because NO is under eight characters. Matching succeeds despite the unmatched short fragment. The quote gate also lowercases and strips Markdown marks. A statement User lives in Paris. paired with a matching quote User lives in London. can pass the inspected gate if its other fields satisfy the structural conditions; no statement/quote entailment relation is tested. Neither deduction claims such a model output actually occurred.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Proposal or rejection reason; normalized matching Boolean is internal evidence.
• Later read-back: inapplicable — this gate itself retains no result; downstream published sources/status are owned by consolidation.
• Delegated visibility: No new model call; caller loop receives Proposal/reason, optional verbose errors expose rejection text.
• Selection predicate: Lasting true; evidence at least eight characters; statement 8–600 characters; matching qualifying fragments; usable routing.
• Invalidation or expiry: inapplicable — current candidate check is not a durable grant; future publication is separate.
• Activation or effect: Enforces eligibility before placement; does not license source truth or semantic entailment.
• Evidence limits: Flag is self-assessment; short fragments ignored; no independent semantic or temporal check.

if !fact.lasting { return Err("not lasting (true for one session only)"); } --- src/consolidate.rs

if !quote_found(&event.content, evidence) { return Err("evidence not found in the cited event"); } --- src/consolidate.rs

.map(normalise) .filter(|f| f.chars().count() >= 8) .collect(); --- src/consolidate.rs

EPI-RTE-index-rebuild

Label: Index metadata derivation

No supplied counterpart after comparison with runtime's routes: cmd_index dispatch is mentioned there, but standalone index rebuild/count derivation was not declared. This route concerns only metadata publication, not consolidation's full loop.

Owner/endpoints: CLI cmd_index calls index::rebuild; current head's parseable entities produce ID/type/title/Active-count rows, sorted and bounded at 60,000 bytes, with omitted-entity count appended. Program derives counts from stored predicates; no external truth oracle. Trigger: caller command; open maintenance. Caller starts rebuilding, fixed program chooses rows, DomainValidator/CAS can reject publication. It reuses old head when unchanged. Guidance is canonical index schema; Git prior revisions permit external recovery, not semantic rollback. SRC-1 src/index.rs, src/cli.rs, implementation. Entailed derivation is bounded to parsed entity facts' stored status: it does not mean true, unsuppressed, endorsed or useful facts. Example deduction: an entity with one Active and one Retracted fact renders one active fact; no fact's proposition is tested.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Unchanged head or published index revision; CLI reports which.
• Later read-back: Later snapshot/file readers can read retained INDEX.md metadata.
• Delegated visibility: No model delegation; index readers see bounded rows and omitted-entity count, not all source evidence.
• Selection predicate: Eligible entity files that parse; sort by entity ID; budget stops listing; counts filter Active status.
• Invalidation or expiry: Rebuild replaces derived index from current head; other publishers also refresh it.
• Activation or effect: Metadata publication wired; exercised navigation use unobserved.
• Evidence limits: Does not establish entity/source truth; malformed entities silently omitted from derivation.

let active = entity.facts.iter().filter(|f| f.status == FactStatus::Active).count(); --- src/index.rs

if current.as_deref() == Some(body.as_str()) { return Ok(base); } --- src/index.rs

EPI-RTE-search-benchmark

Label: Search timing benchmark

Distinct from RT-RTE-search-evaluation: shares combined_search, but evaluates elapsed duration rather than caller expected hits/text. Runtime locates bench without declaring this measurement route.

Owner/endpoints: caller Bench queries/options invoke repeated combined_search calls; Instant measures elapsed milliseconds and print_bench_row summarizes durations/ranker/hits. Target/domain: calls in this process/store/environment, not source truth or answer correctness. Program measures; caller supplies cases and repetition settings. No reference-answer oracle or content veto. Mode: bounded caller-selected timing comparison, not production improvement trigger. It prints lexical and optional reranked summaries, admits no knowledge/model change and supplies no automatic winner consumer. SRC-1 src/cli.rs, implementation.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Printed timing summaries, ranker and hit counts; errors reject completion.
• Later read-back: inapplicable — route retains no report for production adaptation.
• Delegated visibility: Search rankers receive usual query/candidate inputs; caller sees aggregate timings.
• Selection predicate: Caller queries/project/limit and iteration counts; lexical always, reranked when rerank_iterations exceeds zero.
• Invalidation or expiry: inapplicable — report-only measurements have no retained grant.
• Activation or effect: Timing/report code wired; no observed measurements or policy updates.
• Evidence limits: Sequential comparisons, caching, provider and environment changes can confound attribution; no actual result supplied.

ms.push(start.elapsed().as_secs_f64() * 1000.0); --- src/cli.rs

Claims

EPI-CLM-sourced-facts

Label: Curated sourced facts

Claimed operation: imported conversations/instruction files become curated sourced facts with later search/read/change/writeback. SRC-1 README.md, doctrine/design. Distinct from RT-CLM-validated-publication, which claims validation/CAS rather than source-backed content. The source term facts is assessed as a product claim, not assumed epistemic acceptance.

mem reads the conversations you have already had with Claude Code, Codex, and OpenCode, plus your AGENTS.md/CLAUDE.md/README.md files, and turns them into a small set of curated, sourced facts stored in a plain Git repository. Agents search that memory (facts and raw sessions together), read it over MCP, change it through a validated path, and get it written back into each project's AGENTS.md. --- README.md

EPI-CLM-lasting-facts

Label: Lasting fact admission

Claimed operation: extraction keeps only content still true/useful next month and evidence appearing in cited source. SRC-1 README.md, src/consolidate.rs, doctrine/design. Related to EPI-CLM-sourced-facts but distinct criterion; no supplied lastingness claim counterpart. Prompt and quote gate establish requested model judgment and lexical enforcement, not reliable semantic adjudication.

A fact is worth keeping only if it will still be true and useful in a session weeks from now, in a different task: the user's identity and role, standing preferences and working rules, project facts (what it is, stack, architecture, conventions, commands, decisions and why), and named people or organisations with a stated relationship. --- src/consolidate.rs

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; supplied RT-BAP-host-context and RT-BAP-instruction-file cover the consequential external channels.