RightMemory epistemic report

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

Source-and-claim boundary

This report asks how submitted evidence becomes reusable context or guidance, what checks license reliance, and what later consumers can do with it. Scope and exclusions are see boundary; sources are see Source register. SRC-1 supplies implementation, doctrine/design and reported operation, not observed runs or causal experiments. Provider weights, actual host behavior, private memory roots and private benchmark traces remain unavailable. This prevents conclusions about exercised acceptance, factual accuracy, causal task benefit and improved capacity.

Runtime findings are provisional supplied evidence, not an inspection ceiling. Direct reads added the module admission rules, all four Update/Retrieve/Pruner/Insight prompts, builder prompt, guidance capture/review, shared-view builder/approval/question operations, structural validation body, entry dispatch and MCP registration. All are SRC-1 at the frozen commit. The countercase search deliberately included manual curation, sync, imports and live provider questions: an ordinary Update gate cannot characterize all admission, and internal source rendering cannot characterize every answer route.

Consequential design claims are EPI-CLM-memory-admission, EPI-CLM-correction-cases and EPI-CLM-retrieval-coverage; publication's narrower claim is see RT-CLM-native-publication. No claim of a universal factual oracle is inferred from these.

Entry point or operation	Source path, SRC-1	Coverage and records / conclusion prevented
Installed dispatch, validation, MCP	rightmemory/entrypoint.py, rightmemory/mcp.py	Direct dispatch and three registered MCP tools inspected; Retrieve/submit route findings see RT-RTE-agent-retrieval, RT-RTE-isolated-update; capture adds EPI-RTE-guidance-capture.
Native assistant prepare/submit/complete	rightmemory/assistant.py, rightmemory/assistant_context.py, rightmemory/assistant_write.py	Supplied findings see RT-RTE-native-assistant; no independent host trace.
Agent and embedding retrieval	rightmemory/retrieve_selection.py, rightmemory/embedding_retrieval.py	Supplied implementation findings see RT-RTE-agent-retrieval, RT-RTE-embedding-retrieval; direct Retrieve prompt read. No exhaustive renderer audit added.
Update and structural gate	rightmemory/prompts/update.md, rightmemory/reference/MEMORY_RULES.md, rightmemory/tools.py	Direct semantic criteria and validation body; supplied publication details see RT-RTE-isolated-update.
Pruner, Insight, Historian	rightmemory/prompts/pruner.md, rightmemory/prompts/insight.md, rightmemory/runtime.py	Direct first two prompts; supplied scheduler/executor/history findings see RT-RTE-maintenance-cycles. Historical completeness not independently assessed.
Maintenance and Pursuit editor	skills/maintain-rightmemory/SKILL.md, rightmemory/pursuit_store.py	Supplied findings see RT-RTE-manual-maintenance, RT-RTE-pursuit-editor; no inference of universal user approval.
Guidance inbox and reviewed curation	rightmemory/guidance.py, skills/review-agent-guidance-inbox/SKILL.md	Direct capture and approval procedure; EPI-RTE-guidance-capture, EPI-RTE-guidance-review.
Edit feedback	rightmemory/reference/RIGHTMEMORY_EDIT_CORRECTION_RULES.md	Direct rules and Update's late-check request; EPI-OBJ-corrections. Actual user correction artifacts unavailable.
Sync and shared-file imports	rightmemory/sync.py, rightmemory/shared_view_files.py	Supplied findings see RT-RTE-sync-admission, RT-RTE-file-import; all Git transport, repair and revocation combinations not independently audited. Prevents complete withdrawal guarantee.
Shared-view creation/refresh/approval	rightmemory/shared_view_builder.py, rightmemory/shared_view_files.py, rightmemory/prompts/shared-view-builder.md	Direct relevant build, refresh, extractive render and approval bodies; EPI-RTE-view-build, EPI-RTE-view-approval.
Live provider questions	rightmemory/shared_view_questions.py	Direct request, provider execution, timeout/error and answer wrapping; EPI-RTE-provider-question. External providers remain opaque.
Retrieval evaluators	experiments/README.md	Direct full report read, attributed operation only; EPI-CLM-retrieval-coverage. Harness implementation not independently audited; raw labels/traces unavailable.
Agent-file collection	rightmemory/agent_files.py	Direct catalog discovery/hash grouping inspected. No traced later semantic consumer here; catalog availability does not establish authority. Remaining CLI/storage operations unassessed.
Other installation/hooks, manager/onboarding, hub and web frontend paths	rightmemory/install_core.py, skills/, rightmemory/hub/, rightmemory/web/	Boundary inventory or supplied runtime coverage only; detailed independent hook, host prompt and authorization audits not performed. Prevents system-complete negative about additional evaluators or behavioral routes.

Available but uninspected branches are not access gaps or evidenced absences. This is a sparse epistemic assessment of named paths, not an exhaustive absence proof.

Epistemic-object inventory

object ID	candidate truth-apt content or none	claimed role	evidence source ID and local anchor	gap/limit
EPI-OBJ-memory	Facts, interpretations, scoped decisions and rules; questions are not assertions	see EPI-CLM-memory-admission	SRC-1 rightmemory/reference/MEMORY_RULES.md	No actual stored claim or initial warrant supplied.
EPI-OBJ-agent-corrections	Prior attempt/redirection/outcome contrast and applicability	see EPI-CLM-correction-cases	SRC-1 rightmemory/reference/AGENT_CORRECTION_MEMORY_RULES.md	A contrast is not a causal test of a generalized lesson.
EPI-OBJ-corrections	Exact candidate and rejected/accepted edit contrast	Operational curation feedback	SRC-1 rightmemory/reference/RIGHTMEMORY_EDIT_CORRECTION_RULES.md	No independent semantic adjudicator or executed correction supplied.
EPI-OBJ-guidance-inbox	User redirection or direct-guidance evidence	Pending, not formal guidance	SRC-1 rightmemory/guidance.py	Caller interpretation and completeness unknown.
EPI-OBJ-view-document	Selected or rewritten provider memory	Shareable context	SRC-1 rightmemory/shared_view_files.py	Extractive and generative lineage differ; approval is permission, not truth.
EPI-OBJ-retriever-policy	Mostly policy; factual assumptions may be embedded	Provider-side question instructions	SRC-1 rightmemory/shared_view_questions.py	Native policy text is not independently checked for semantic fidelity.
EPI-OBJ-view-permission	No truth-apt content update from approval metadata	Permission to distribute or answer	SRC-1 rightmemory/shared_view_files.py	A flag is not a factual acceptance decision.
EPI-OBJ-question-answer	Provider response, potentially truth-apt	Answered/unavailable result	SRC-1 rightmemory/shared_view_questions.py	Concrete proposition not determinable without response; status does not establish correctness.
EPI-OBJ-insight-logs	Patterns, tensions and possible recommendations	Reflection for future work	SRC-1 rightmemory/prompts/insight.md	No produced instance or acceptance decision supplied.
RT-OBJ-role-instructions	Mostly policy rather than assertions about a user's world	see RT-OBJ-role-instructions	SRC-1 rightmemory/prompt.py	Model interpretation not compliance.
RT-OBJ-runtime-config, RT-OBJ-semantic-operation	No world-claim change from dispatch/phase state alone	see cited records	SRC-1 rightmemory/config.py, rightmemory/runtime.py	State labels do not grant semantic warrant.
RT-CMP-role-model, RT-CMP-embedding-model, RT-CMP-reranker-model	Distributed-parametric inference/ranking, not individually recoverable propositions	see cited records	SRC-1 rightmemory/runtime.py, rightmemory/embedding_models.py	No accessible weights, training evidence or exact deployed identity; no operational parameter update established in inspected callers.

Authority-route ledger

Each row is a separate consequential function; repeated route citations identify different phases of the same declared referent, not new declarations. Record-owned endpoints, activation, recovery and evidence fields are see the cited record. Here “unknown warrant” means the source claim's factual warrant has not been established; faithful text transport does not improve it.

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
EPI-RTE-guidance-capture	retention	implemented	EPI-OBJ-guidance-inbox	truth-apt transformation: acquisition/import	Incoming evidence text	Nonempty session/evidence and inbox parser; syntactic domain	see EPI-RTE-guidance-capture	Pending inbox entry	Enforcing write transaction	Unknown source warrant; no formal guidance acceptance	Save evidence for later review	Review host; inbox text; advisory evidence; later explicit review	SRC-1 rightmemory/guidance.py	none	none	No observed run.
EPI-RTE-guidance-review	disposition/acceptance	doctrine only	EPI-OBJ-guidance-inbox	no content change	Proposed meaning, scope and destination	Human explicitly approves host proposal against module rules; intended use is future behavior guidance	see EPI-RTE-guidance-review	Approve, reject, already covered, defer	Procedural human veto	Approval licenses requested guidance within approved scope; factual truth not independently checked	Authorize only approved semantic changes	Host editor; explicit approval; permissive; current curation	SRC-1 skills/review-agent-guidance-inbox/SKILL.md	none	none	Protocol, not enforced per-write gate.
EPI-RTE-guidance-review	lifecycle integration	doctrine only	EPI-OBJ-memory; EPI-OBJ-agent-corrections	truth-apt transformation: indeterminate	Approved semantic proposal → scoped canonical guidance and inbox removal	Host applies approved meaning and validates in clean worktree	see EPI-RTE-guidance-review	Merged/replaced guidance and evidence disposition	Skill protocol	Conditional on prior approval; new generalized wording needs separate warrant	Land exact validated change; preserve deferred evidence	Future retrieval then host; source text; advisory; cross-session	SRC-1 skills/review-agent-guidance-inbox/SKILL.md	none	none	Semantic preservation versus new generalization depends on unavailable proposal/output.
RT-RTE-isolated-update	check/evidence production	implemented	EPI-OBJ-memory; EPI-OBJ-agent-corrections	no content change	Durability, usefulness, clear scope, overlap, evidence clarity and stale claims	Role model under module rules and late corrections check; semantic relevance to future work	see RT-RTE-isolated-update	Keep/merge/skip/edit judgment, question or uncertain state	Natural-language criteria requested	Fallible evidence-informed candidate assessment, not independent corroboration	Shape proposal; model can reject all candidates	Update model; prompt plus candidates/current state; advisory instructions; current operation	SRC-1 rightmemory/prompts/update.md	EPI-CLM-memory-admission	none	No observed run.
RT-RTE-isolated-update	content transformation	implemented	EPI-OBJ-memory; EPI-OBJ-agent-corrections	truth-apt transformation: indeterminate	Candidates/current state → revised wording	Role model; no supplied input/output pair	see RT-RTE-isolated-update	Revision, compression, new scoped inference or no change	Model edits in candidate tree	No concrete entailed or ampliative claim established without instances	Produce candidate commit	Supervisor; candidate files; proposed state; current operation	SRC-1 rightmemory/prompts/update.md	none	none	Reshaping, entailment and conjecture remain possible; non-entailment not evidenced.
RT-RTE-isolated-update	check/evidence production	implemented	EPI-OBJ-memory; EPI-OBJ-agent-corrections	no content change	Committed paths, file modes, semantic base and graph/correction structure	Supervisor and MemoryTools.validate_memory; structural domain	see RT-RTE-isolated-update	Validation pass/failure	Enforcing validation at publication	Structure only; no factual oracle; correction capacity exception is see record	Report eligibility or reject candidate	Supervisor; commit/validation result; enforcing; publication attempt	SRC-1 rightmemory/isolated_write.py	none	none	No observed run.
RT-RTE-isolated-update	operational admission/selection/consumption	implemented	EPI-OBJ-memory; EPI-OBJ-agent-corrections	no content change	Validated proposed state → active root	Supervisor, applicable validation and base fencing	see RT-RTE-isolated-update	Landed/no-change/rejected/prepared	Enforcing publication checks	Operational admission is not evidence of accepted ampliative knowledge	Allow state to reach later retrieval	Active root and retriever; Git files; permissive admission then advisory delivery; later invocations	SRC-1 rightmemory/isolated_write.py	none	none	No observed run.
RT-RTE-native-assistant	operational admission/selection/consumption	implemented	EPI-OBJ-memory	no content change	Native prepared edits → completion receipt	Host semantic judgment, program validates completion; same structural/factual distinction	see RT-RTE-native-assistant	Receipt, reject, retry	Program enforces inspected completion checks	Receipt attests publication, not correctness	Make committed content available; native retrieval independently host-selected	Native host; live files/handoff; advisory context; current and later tasks	SRC-1 rightmemory/assistant_write.py	RT-CLM-native-publication	none	No observed run.
RT-RTE-agent-retrieval	content transformation	implemented	EPI-OBJ-memory; EPI-OBJ-agent-corrections	truth-apt transformation: non-ampliative reshaping	Selected IDs/ranges → source-derived return	Model relevance judgment; symbolic source resolver and output-size constraint	see RT-RTE-agent-retrieval	Rendered content or failed selector	Selection plus structural enforcement	Preserves selected source wording, unknown initial warrant; relevance not proven	Supply context rather than synthesize an answer	Host; retrieval return; advisory; current task and repeat retrieval	SRC-1 rightmemory/retrieve_selection.py	none	none	No observed run.
RT-RTE-embedding-retrieval	operational admission/selection/consumption	implemented	EPI-OBJ-memory	no content change	Ranked source entries → returned set	Embedding similarity then reranker/count limits; ranking domain	see RT-RTE-embedding-retrieval	Top entries even without stored answer	Ranking	Neither similarity nor answered-need guarantee is truth warrant	Prioritize source context	Host; selected original text; ranking/advisory; query	SRC-1 rightmemory/embedding_retrieval.py	EPI-CLM-retrieval-coverage	none	No observed run.
RT-RTE-maintenance-cycles	operational admission/selection/consumption	implemented	EPI-OBJ-memory	no content change	Prune stale active surface	Due code then model comparison of boundary/current meaning, active patterns, Pursuit references and grace	see RT-RTE-maintenance-cycles	Remove or preserve; checkpoint ledger	Deletion admission plus procedural semantic judgment	Unchanged text is not false text; age is not refutation	Withdraw active retrieval availability; keep grace/history route	Future retriever; remaining files and historian; enforcing active-set change; generations	SRC-1 rightmemory/prompts/pruner.md	none	none	No observed run.
RT-RTE-maintenance-cycles	content transformation	implemented	EPI-OBJ-insight-logs	truth-apt transformation: indeterminate	Memory/prior logs → possible reflection	Insight model judges meaningfulness; source limited to memory root	see RT-RTE-maintenance-cycles	New log or no-op	Role request and scoped publication	Possible conjecture; no evidenced accepted ampliative claim	Retain separate logs, not promote to active Memory	Later Insight model; prior logs; advisory; later cycles	SRC-1 rightmemory/prompts/insight.md	none	none	No content-directed criticism/revision trace or concrete new proposition supplied.
RT-RTE-manual-maintenance	operational admission/selection/consumption	doctrine only	EPI-OBJ-memory	truth-apt transformation: indeterminate	Explicit requested maintenance edits	Host/user-scoped judgment plus instructed validation and base fences	see RT-RTE-manual-maintenance	Changed canonical files or stop	Skill protocol	Requested edit authorization not universal truth acceptance	Afford direct state revision outside Update	Host then retriever; worktree/files; permissive then advisory; later sessions	SRC-1 skills/maintain-rightmemory/SKILL.md	none	none	Supplied findings only; host compliance unobserved.
RT-RTE-pursuit-editor	behavior/policy adaptation	implemented	Pursuit state — see RT-RTE-pursuit-editor	non-truth-apt policy/content update: human direction-map changes	User direction operations → canonical Pursuit map	Allowed symbolic operations, revision fence and graph validator; not a factual evaluator	see RT-RTE-pursuit-editor	New intended directions and opening context	Enforcing allowed map changes	User authority over intent, not world truth	Change direction supplied to later agent work	Host; opening-context/Pursuit text; advisory; future tasks	SRC-1 rightmemory/pursuit_store.py	none	none	No observed host adoption.
RT-RTE-sync-admission	operational admission/selection/consumption	implemented	EPI-OBJ-memory	truth-apt transformation: acquisition/import	Upstream state → validated active candidate	Allowlist/graph validation and possible reconciler; structural domain	see RT-RTE-sync-admission	Admitted state or repair/rejection	Enforcing structural admission	Imported warrant unknown; local Update semantic judgment not required	Change sources available to retrieval	Retriever/host; synchronized files; permissive then advisory; later calls	SRC-1 rightmemory/sync.py	none	none	No observed run.
RT-RTE-file-import	lineage/freshness/recovery	implemented	EPI-OBJ-shared-view	truth-apt transformation: acquisition/import	External package → persistent local import	Expected identity/namespace/package validation; failed refresh may keep old valid import	see RT-RTE-file-import	Fresh/stale/unavailable	Enforcing package checks, freshness status	Provider attribution, unknown factual warrant; stale does not imply false	Continue serving valid stale source or replace	Readers/host; imported files and status; advisory; until refresh	SRC-1 rightmemory/shared_view_files.py	none	none	No observed run.
EPI-RTE-view-build	content transformation	implemented	EPI-OBJ-shared-view	truth-apt transformation: non-ampliative reshaping	Extractive selected blocks → view distribution	Symbolic block selector, renderer and structural validator	see EPI-RTE-view-build	Source lines selected with ancestry and exclusions	Enforcing deterministic extraction	Selected text preserved; omitted context can limit interpretation; no new fact derived	Construct shareable subset	Package consumers; dist Markdown; advisory context; after publication/import	SRC-1 rightmemory/shared_view_files.py	none	none	No observed run.
EPI-RTE-view-build	content transformation	implemented	EPI-OBJ-shared-view	truth-apt transformation: indeterminate	Generative rewrite of mixed private/shareable memory	Builder model plus schema-valid compiler; not semantic equivalence check	see EPI-RTE-view-build	Rewritten dist or tool failure	Model generation and structural admission	No demonstrated faithful meaning, new entailment or accepted novelty	Prepare shareable candidate; refresh can retain approval	External retrievers; generated Markdown; advisory; subsequent publication	SRC-1 rightmemory/prompts/shared-view-builder.md	none	none	No concrete output pair; missing semantic checks cannot itself prove non-entailment.
EPI-RTE-view-approval	operational admission/selection/consumption	implemented	EPI-OBJ-shared-view	non-truth-apt policy/content update: approved flag	Structurally valid view → publication permission	Caller invokes approval; code validates then sets approved; no per-proposition evidence decision	see EPI-RTE-view-approval	Approved or structural error	Permissive metadata with code checks	Distribution authorization only	Enable sharing and later approved refresh publication	Publishers; recipe approved flag; permissive; until changed; refresh preserves old approval	SRC-1 rightmemory/shared_view_files.py	none	none	No observed run.
EPI-RTE-provider-question	operational admission/selection/consumption	implemented	EPI-OBJ-question-answer	truth-apt transformation: acquisition/import	Provider response → consumer-visible answer	Approved configuration, token gate at token verification, started/answer timeouts and status parsing; availability domain	see EPI-RTE-provider-question	Answered/unavailable	Availability enforcement; no factual acceptance gate	Answered means completed provider response, not verified truth	Return response for caller interpretation	Requesting caller; formatted answer/status; advisory; current question	SRC-1 rightmemory/shared_view_questions.py	none	none	Provider role may reshape or select; content relation inside provider is indeterminate.

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-memory-admission	see claim	SRC-1 rightmemory/reference/MEMORY_RULES.md, doctrine/design	see claim	RT-RTE-isolated-update	None supplied	No causal comparison; model interprets all semantic criteria	Program requests scoped evidence-grounded context and preserves useful uncertainty	Policy is not proof that every stored assertion passes a truth test.
EPI-CLM-correction-cases	see claim	SRC-1 rightmemory/reference/AGENT_CORRECTION_MEMORY_RULES.md, doctrine/design	see claim	RT-RTE-isolated-update, EPI-RTE-guidance-review	None supplied	User contrast and recurrence judgment do not identify causal transfer	Concrete cases and generalized behavior rules have distinct admission criteria	No measured improvement or automatic guarantee of recurrence.
EPI-CLM-retrieval-coverage	see claim	SRC-1 experiments/README.md, reported operation	see claim	RT-RTE-embedding-retrieval	No inspectable raw trace registered	Private, agent-labelled small corpus; model/count/render/interface comparisons differ; component effects not independently identified by this analysis	Attributed required-ID coverage and latency, not independent factual-answer accuracy	Fixed-count selection still returns results on no-answer cases; source explicitly acknowledges this.
RT-CLM-native-publication	see claim	SRC-1 skills/rightmemory-assistant/SKILL.md, doctrine/design	see claim	RT-RTE-native-assistant	None supplied	Publication receipt not semantic accuracy or exercised host honesty	Prepared work and landed state differ	No warrant for content truth follows from completion.

Bounded conclusion

RightMemory has several distinct checks, not one epistemic evaluator. Update requests an evidence-consuming semantic judgment against durability, future usefulness, scope, duplication and evidence clarity. Its model may skip, revise, retain uncertainty or preserve questions. A separate deterministic gate checks publishable structure and scope. Neither a structural pass nor a landed commit establishes the factual correctness of a stored claim. No concrete produced accepted ampliative claim is evidenced by these source instructions alone. See RT-RTE-isolated-update, EPI-CLM-memory-admission and EPI-CLM-correction-cases.

The ordinary agent path returns resolved stored text, not free-form model answers. Embedding/reranking selects source entries but supplies no answer-absence or truth oracle. Those paths preserve selected wording without creating initial warrant; missing source provenance stays unknown. Native retrieval has a different host-authored result path, and provider questions explicitly return live provider results with answered/unavailable labels. Thus “retrieval never synthesizes anything” is too broad for the whole shipped system. See RT-RTE-agent-retrieval, RT-RTE-embedding-retrieval, RT-RTE-native-assistant and EPI-RTE-provider-question.

Correction cases, generalized behavioral Memory and edit feedback differ. The first preserves a scoped user-redirection contrast; the second prescribes future behavior; the third preserves rejected-versus-accepted curation examples for a late check by future Update work. Guidance capture alone retains unsettled evidence. Explicit inbox review adds a user approval criterion for meaning/scope and then connects approved changes to canonical guidance. That is a conditional procedural acceptance/integration route, not evidence it operated or that ordinary automatic Update requires human approval. See EPI-OBJ-corrections, EPI-RTE-guidance-capture and EPI-RTE-guidance-review.

Pruner withdraws active availability using unchanged meaning, active use and grace criteria, not a test that an assertion is false. Insight can propose patterns in a separate retained log; later Insight reads are wired, but this does not establish acceptance or active-Memory promotion. Sync, imported views, direct maintenance and the Pursuit editor have separate authority. No inspected universal withdrawal propagation or rollback guarantee covers every external consumer. See RT-RTE-maintenance-cycles, RT-RTE-sync-admission, RT-RTE-file-import, RT-RTE-manual-maintenance and RT-RTE-pursuit-editor.

Extractive views mechanically reshape selected source text. Generative views request rewritten shareable text; their semantic preservation is indeterminate without actual inputs and outputs. Approval licenses distribution, and refresh preserves prior approval rather than requiring a new per-output human decision. This is operational authority, not evidence-consuming acceptance of every generated proposition. See EPI-RTE-view-build and EPI-RTE-view-approval.

Retained guidance can change later action through model interpretation, but actual activation and improved capacity are unobserved. No supplied trace establishes content-directed criticism iterating on a formulated explanation, successful reflection, autonomous theory building or exercised self-improvement. These conclusions remain independent: persistent trace-fed correction capacity is supported at the instruction/update boundary; factual warrant, learned benefit and causal self-change do not follow. Model-weight changes are not established in inspected callers; provider internals remain uninspected. See RT-CMP-role-model, RT-BAP-role-policy and RT-BAP-host-context. The coverage limits above prevent an unqualified system-wide negative.

Shared records

Operative objects

EPI-OBJ-memory

Label: Memory

Natural-language facts, preferences, interpretations, scoped rules and unresolved questions in Git-backed MEMORY.md, MEMORY_*.md and linked Markdown/skill assets. This is semantic state, distinct from runtime configuration, operation metadata and package policy in RT-OBJ-runtime-config, RT-OBJ-semantic-operation and RT-OBJ-role-instructions. No supplied operative-object record names this same retained semantic state. SRC-1: rightmemory/reference/MEMORY_RULES.md.

A Memory item should express one reusable fact, preference, rule, decision, conclusion, or unresolved question at the smallest scope that remains independently meaningful. --- rightmemory/reference/MEMORY_RULES.md

EPI-OBJ-agent-corrections

Label: Agent Corrections

Natural-language scoped cases stored in the two fixed Git-backed correction Markdown collections. Distinct from the generic update route RT-RTE-isolated-update and package instructions RT-OBJ-role-instructions: this record names its semantic products, not execution or policy. SRC-1: rightmemory/reference/AGENT_CORRECTION_MEMORY_RULES.md.

A stored Agent Correction is not the same thing as a user redirection in conversation. A redirection is evidence; Update or an explicit review workflow decides whether that evidence deserves a reusable entry. --- rightmemory/reference/AGENT_CORRECTION_MEMORY_RULES.md

EPI-OBJ-corrections

Label: RightMemory Edit Feedback

Natural-language Candidate/Proposed edit/Accepted edit sections in optional root corrections.md, stored in the memory repository. Distinct from Agent Corrections and the supplied role-instructions object; Update reads this operational feedback after tentative planning. Its exact-evidence requirement supports lineage within the example, not generalized truth. SRC-1: rightmemory/reference/RIGHTMEMORY_EDIT_CORRECTION_RULES.md.

Keep evidence exact. Do not derive general behavior lessons inside this collection. --- rightmemory/reference/RIGHTMEMORY_EDIT_CORRECTION_RULES.md

EPI-OBJ-guidance-inbox

Label: Pending Agent Guidance

Natural-language evidence bodies with symbolic entry/session/time fields in Git-backed AGENT_GUIDANCE_INBOX.md. No counterpart object in supplied runtime: RT-RTE-isolated-update processes a different submission route; the inbox explicitly uses user-reviewed direct curation. SRC-1: rightmemory/guidance.py.

GUIDANCE_INBOX_PATH = "AGENT_GUIDANCE_INBOX.md" --- rightmemory/guidance.py

EPI-OBJ-shared-view

Label: Shared-view artifacts

Container whose operative parts are split below by consumption: dist Markdown is natural-language provider content; retriever.md is natural-language provider policy; recipe.toml/question.toml contain symbolic selection, approval and endpoint metadata. Stored under shared_views// and imported runtime directories. This names provider-produced artifacts; RT-RTE-file-import instead names acquisition of file packages. SRC-1: rightmemory/prompts/shared-view-builder.md.

Build only provider-owned shared-view artifacts under shared_views/<view-id>/. --- rightmemory/prompts/shared-view-builder.md

EPI-OBJ-view-document

Label: Shared Memory document

Part of: EPI-OBJ-shared-view

Natural-language dist/MEMORY.md and extractively copied backing documents, stored in provider distribution files and consumer imports. The part is source content, not recipe permission or question instructions. Extractive selection preserves surviving text; generative compiler accepts a complete document subject to structure checks. Compared with RT-RTE-file-import, this object names what is transported, not its acquisition route. SRC-1: rightmemory/shared_view_files.py.

_write_text(dist / "MEMORY.md", memory_document.strip() + "\n") --- rightmemory/shared_view_files.py

EPI-OBJ-retriever-policy

Label: Provider retriever instructions

Part of: EPI-OBJ-shared-view

Natural-language retriever.md in a provider view directory. The builder supplies policy text; answer_question_view concatenates it with the current question before invoking the configured provider role. This policy has behavioral authority at that model invocation, separate from consumer-facing dist content and from package role policy RT-OBJ-role-instructions. No meaning-equivalence validator for generated instructions is established by the inspected wrapper. SRC-1: rightmemory/shared_view_questions.py.

retriever.read_text(encoding="utf-8").strip(), --- rightmemory/shared_view_questions.py

EPI-OBJ-view-permission

Label: Shared-view approval metadata

Part of: EPI-OBJ-shared-view

Symbolic approved fields in recipe.toml and question.toml, stored in provider files. File approval validates then writes a flag; question answering refuses an unapproved config. Their scope is distribution or execution permission, not semantic acceptance. The narrower file-view change is EPI-RTE-view-approval; question control is EPI-RTE-provider-question. Distinct from RT-OBJ-runtime-config: view-level content permission, not executor/backend settings. SRC-1: rightmemory/shared_view_files.py, rightmemory/shared_view_questions.py.

if not config.approved: return _format_unavailable(clean_view_id, "question view is not approved") --- rightmemory/shared_view_questions.py

EPI-OBJ-question-answer

Label: Shared question answer

Natural-language provider result with symbolic status interpreted by wrappers, transient in the returned string/HTTP payload. Distinct from RT-RTE-agent-retrieval: a separately exposed live provider question path, not ordinary terminal-selection retrieval. SRC-1: rightmemory/shared_view_questions.py.

return f"Shared question: {clean_view_id}\nStatus: answered\nAnswer:\n{answer.strip()}\n" --- rightmemory/shared_view_questions.py

EPI-OBJ-insight-logs

Label: Insight logs

Natural-language reflections in timestamped Git-backed insight_logs/*.md. Distinct product of the Insight part of RT-RTE-maintenance-cycles, not a replacement for that execution container. Prior logs are explicitly sources for later cycles; no automatic Memory promotion found in inspected prompt. SRC-1: rightmemory/prompts/insight.md.

Do not append to existing logs. --- rightmemory/prompts/insight.md

Routes

EPI-RTE-guidance-capture

Label: Guidance evidence capture

Caller/session evidence → submit_guidance → parsed/rendered inbox entry → committed temporary worktree → base-fenced fast-forward. Owner is capture code, not Update. Nonempty text and parse validity decide retention; caller supplies content, and code vetoes malformed evidence. No expected-answer oracle, per-entry human approval or semantic admission is imposed. This extends coverage beyond RT-RTE-isolated-update with a distinct durable pending object. Operating mode is open capture, not an improvement experiment. SRC-1: rightmemory/guidance.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: New pending entry identity; MCP caller ordinarily receives only warning state.
• Later read-back: Explicit review host reads inbox and proposes formal guidance; pending evidence is not automatically promoted.
• Delegated visibility: CLI/MCP capture caller and review host; inspected tools do not grant this inbox to ordinary RightMemory role reads.
• Selection predicate: Nonempty evidence/session and valid inbox structure; no semantic truth test.
• Invalidation or expiry: Explicit review may remove approved/rejected entries; deferred entries stay; HEAD conflict aborts fast-forward.
• Activation or effect: Durably retains unresolved evidence; later behavioral change requires approved curation and retrieval.
• Evidence limits: No real captured entry or review run supplied.

_git(root, "merge", "--ff-only", _git(worktree, "rev-parse", "HEAD")) --- rightmemory/guidance.py

EPI-RTE-guidance-review

Label: User-reviewed guidance curation

Pending inbox → host proposal naming evidence, intended meaning/scope/destination → explicit user approval → host worktree edits and inbox removals → instructed validation and exact landing. Owner is host/human under the skill. Human can veto or defer individual proposals; host drafts meaning, not exact final wording before approval. Criterion is approved scoped guidance under module admission rules, intended for future behavior or reusable cases; approval does not independently corroborate factual claims. Operating mode is explicit user-reviewed curation; no answer oracle. Recovery: stop on dirty/changed active base; leave deferred entries; no universal rollback after protocol violation. Compared with RT-RTE-manual-maintenance, this is a distinct inbox-specific proposal/approval/removal workflow, not general direct maintenance. SRC-1: skills/review-agent-guidance-inbox/SKILL.md.

• implementation conclusion status: afforded
• operation conclusion status: uninspected
• Immediate return: Applied proposals, remaining inbox, validation/commit report or no commit.
• Later read-back: Canonical guidance becomes available to later retrieval; inbox removal disposes reviewed evidence.
• Delegated visibility: Host/user see proposal evidence and scope; later agents receive admitted guidance rather than necessarily all original evidence.
• Selection predicate: Explicitly approved proposals only; reject/already-covered/defer dispositions are available.
• Invalidation or expiry: Deferred/unreviewed entries unchanged; module rules allow future narrowing/removal; active-base change blocks landing.
• Activation or effect: Approval can authorize semantic changes; skill affords lifecycle integration after that approval, with no observed execution.
• Evidence limits: Natural-language procedure, not mandatory program gate; no factual oracle or benefit trace.

Do not present exact final wording, graph ids, or a patch before approval. Apply only explicitly approved proposals. --- skills/review-agent-guidance-inbox/SKILL.md

EPI-RTE-view-build

Label: Shared-view build and refresh

Caller intent/current active memory → shared-view-builder model → extractive selection or generative complete Memory document → compiler artifacts → applicable schema/dist checks. Owner is builder wrapper and role model. Model proposes mode/selection/refined intent, tools can fail and request correction; no per-content human acceptance in build itself. Operating mode is open sharing request or due/forced refresh. No expected-answer oracle checks meaning. File build/refresh backs up view directory and restores it on failure; refresh records time/active commit, preserves old approval and can publish when configured. Distinct from RT-RTE-file-import: this is provider construction, not consumer acquisition. The provider model follows the fixity limits of RT-CMP-role-model. SRC-1: rightmemory/shared_view_builder.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Builder summary or failure; refresh can report not due.
• Later read-back: Recipe/dist/retriever policy persist for later publication, import or question invocation.
• Delegated visibility: Builder sees provider active memory; consumers get selected/rewritten distribution, not all private source context.
• Selection predicate: Model's shareability/clarity judgment; extractive recipes select/exclude IDs/files; generative compiler checks schema, not semantic equivalence.
• Invalidation or expiry: Refresh due requires changed active commit and configured elapsed interval unless forced; failed file build/refresh restores view directory.
• Activation or effect: Prepares context and question policy; later publication/import supplies them to consumers; no exercised benefit supplied.
• Evidence limits: No actual generative output pair; no proof of privacy-preserving semantics or factual completeness; all repair/publication branches not exhaustively audited. Generated question instructions become EPI-OBJ-retriever-policy, consumed as provider policy, whereas dist text becomes EPI-OBJ-view-document.

approved=old_recipe.approved, --- rightmemory/shared_view_builder.py

EPI-RTE-view-approval

Label: File-view approval permission

Caller → validate recipe → render or validate generated output → validate dist → persist approved flag. Code owns this API's gate. Caller requests approval; code can reject structure. Caller identity alone does not establish human review or an evidence-consuming per-claim acceptance decision. No answer oracle. Operating mode is explicit view permission, distinct from RT-RTE-file-import and EPI-RTE-view-build. Publication consumes approval, including configured refresh retaining it; approval is not a truth evaluation. SRC-1: rightmemory/shared_view_files.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Approved-view status or validation exception.
• Later read-back: Persisted recipe flag licenses later publish paths; dist becomes external context after publication/import.
• Delegated visibility: Publisher consumes flag; later readers consume distribution, not this permission as factual evidence.
• Selection predicate: Valid recipe/selection and generated-or-extractive dist checks before approved=True.
• Invalidation or expiry: Recipe revisions can alter permission; automatic refresh deliberately preserves old approval; universal remote revocation is uninspected.
• Activation or effect: Permission changes publication eligibility, not confidence in source propositions.
• Evidence limits: No supplied reviewer identity, inspected per-proposition evidence decision or external consumer withdrawal trace.

_write_text(_view_dir(root, recipe.view_id) / "recipe.toml", _render_recipe_toml(_replace_recipe(recipe, approved=True))) --- rightmemory/shared_view_files.py

EPI-RTE-provider-question

Label: Live provider question response

Consumer ask_question_view → configured HTTP target → provider answer_question_view → approved view/retriever prompt plus question → configured runtime role → wrapped answer or unavailable. Provider code owns availability checks and caller code parses/relabels results. Model/provider determines content under generated retriever policy; no answer oracle, independent semantic rejection or accepted-claim record is added here. Operating mode is open question answering. Token verification checks authorization and approval, not truth. Timeouts/errors withhold an answer status; successful return permits consumer reliance without another truth check. Compared with RT-RTE-agent-retrieval and RT-RTE-file-import, this is a distinct live question endpoint. Configured runtime model fixity is see RT-CMP-role-model. SRC-1: rightmemory/shared_view_questions.py.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Formatted answered/unavailable response; provider error/start/answer timeout becomes unavailable.
• Later read-back: uninspected — wrapper does not itself persist the answer for future retrieval; configured runtime role/session may retain its own history.
• Delegated visibility: Provider sees retriever policy and question; consumer receives wrapped response, not full internal work.
• Selection predicate: Configured question connection, approved provider view, retriever file, authentication where token verifier is used, successful provider completion.
• Invalidation or expiry: Request timing bounds availability; token verifier rejects unapproved configuration; no general recall/withdrawal of already returned content is shown.
• Activation or effect: Answered status changes availability and delivers provider content, not factual endorsement.
• Evidence limits: No deployed provider prompt/output or inspected provider trace; internal transformation may select, rewrite or infer and is indeterminate.

if not config.approved: return _format_unavailable(clean_view_id, "question view is not approved") --- rightmemory/shared_view_questions.py

Implementation deduction for EPI-RTE-view-build: in the inspected extractive renderer, suppose a selected full block has node line - [x] latency is 20 ms and its sibling is excluded. Given a parsed node and full inclusion, _render_block returns that exact line; the excluded sibling returns the empty string. The parent joins included pieces; render strips managed examples and adds a terminal newline before dist validation/promotion. The surviving proposition is unchanged, but nothing checks whether 20 ms is true or whether omitted surrounding material was necessary warrant. This is a code-based hypothetical trace, not target execution or an observed stored assertion. SRC-1: rightmemory/shared_view_files.py.

if block.kind == "node": return block.line if full or key in self.exact_nodes else "" --- rightmemory/shared_view_files.py

if self._is_excluded(key): return "" --- rightmemory/shared_view_files.py

text = _strip_managed_examples(text).strip() --- rightmemory/shared_view_files.py

Claims

EPI-CLM-memory-admission

Label: Durable context with scoped evidence

Memory rules claim a future-use admission policy, not storage of everything. Module policy treats useful uncertainty explicitly and forbids converting speculation into fact. Update asks revalidation of possibly stale claims. This is a design claim about intended handling, distinct from the publication receipt claim RT-CLM-native-publication. SRC-1: rightmemory/reference/MEMORY_RULES.md, rightmemory/prompts/update.md.

Useful unsettled context may be stored as Uncertain: when the uncertainty itself affects future work. Do not convert speculation into declarative fact. --- rightmemory/reference/MEMORY_RULES.md

Revalidate claims that may have become stale instead of copying candidates as fact. --- rightmemory/prompts/update.md

EPI-CLM-correction-cases

Label: Reusable user-redirection cases

Agent Corrections claims to preserve scoped contrast that generalized rules would weaken. Its six admission conditions require identifiable attempt, clear redirection/outcome, recurrence or cost, distinct value and clear applicability. A different result alone is explicitly insufficient. Distinct from the supplied publication claim: this is a semantic content/admission claim. SRC-1: rightmemory/reference/AGENT_CORRECTION_MEMORY_RULES.md.

Agent Corrections is RightMemory's bounded library of reusable cases in which a user redirected prior agent work. It preserves concrete contrast that would be weakened or lost if reduced to a generalized preference or rule. --- rightmemory/reference/AGENT_CORRECTION_MEMORY_RULES.md

A different result alone is insufficient. --- rightmemory/reference/AGENT_CORRECTION_MEMORY_RULES.md

EPI-CLM-retrieval-coverage

Label: Reported labelled retrieval coverage

The repository reports 28/28 positive required-ID cases retained across three integrated-backend passes, under the documented small frozen benchmark and environment. This is attributed reported operation, not observed-run evidence in this analysis. Unlike RT-CLM-native-publication, the claim concerns retrieval measurements, not publication. It supplies no independent end-to-end answer judgment or causal component attribution. SRC-1: experiments/README.md.

This is still a small, agent-labelled coverage test, not independent answer-quality evaluation. --- experiments/README.md

The four no-answer queries also receive ten candidates; deciding their usefulness belongs to the caller. --- experiments/README.md

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; downstream host interpretation and package role policy are see RT-BAP-host-context and RT-BAP-role-policy. New-route-specific channels, force and horizons are stated in their declarations and ledger rows.