Instinctual Memory epistemic report

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

Source-and-claim boundary

This lens traces truth-apt content entering the journal and fact store, candidate generation and checks, explicit fact changes, retrieval ranking, and writeback. It assesses the full repository boundary described in boundary.md, using SRC-1 and the runtime baseline records OBJ-1–OBJ-3, RTE-1–RTE-7, CLM-1–CLM-2, and BAP-1–BAP-3; the Source register in boundary.md defines the frozen source scope. The assessed route families are ingestion, extraction/consolidation, direct remember/correct/forget/erase, search and reranking, retention/read-back, and writeback. Actual installed-host calls, agent use of returned material, provider-side model behavior, and production outcomes are uninspected; those gaps prevent observed-use and end-to-end behavioral conclusions. No observed run or causal experiment was supplied (SRC-1; RTE-1–RTE-7).

The inspected design and implementation describe durable facts and their operational handling, but no explicit criterion for accepting a fact as true, or broader knowledge-production/warrant claim, was identified in the inspected claims and routes. CLM-1 and CLM-2 concern hook failure behavior and publication protocol, respectively; neither claims factual correctness. This is a boundary-limited finding, not a claim that no such statement exists elsewhere or in uninspected operation.

Epistemic-object inventory

object ID	candidate truth-apt content or none	claimed role	evidence source ID and local anchor	gap/limit
OBJ-1	Imported session, user, and file content that can express propositions	Journal input for later consolidation and journal search	SRC-1 src/journal.rs; see OBJ-1	Ingestion preserves event content as source material, but the truth of a statement and its source warrant are not independently established. No observed event was supplied.
EPI-OBJ-1	Fact statements proposed by rule or LLM consolidation, including source-event links	Candidate durable facts derived from journal/file inputs; subset of the facts in OBJ-2	SRC-1 src/consolidate.rs; RTE-4	Possible duplicate of the fact content within OBJ-2: both refer to persisted entity facts; this record isolates the extraction-produced subset and its evidence checks. No observed candidate artifact or trace was supplied.
EPI-OBJ-2	A one-sentence fact supplied by a user or agent for remember/correct	Explicitly requested durable fact; subset of facts in OBJ-2	SRC-1 src/change.rs; RTE-5	Possible duplicate of the fact content within OBJ-2: this isolates facts whose proposition is supplied directly by a caller. The caller's assertion/reason is not independent truth evidence; no observed request was supplied.
EPI-OBJ-3	No truth-apt output; query-relative relevance scores/order over retrieved fact statements	Search ranking result used to select or order items for return	SRC-1 src/jev.rs; RTE-2, RTE-3	External scoring operation was not observed. A relevance judgment is not an assessment of whether the fact is true.
EPI-OBJ-4	Fact statements rendered in a host instruction-file managed block	Writeback of eligible stored facts for a host to load	SRC-1 src/cli.rs; RTE-6	Possible duplicate of OBJ-3's writeback content: this isolates the generated instruction-file form. Host loading, precedence, and use were not inspected.

Authority-route ledger

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
RTE-4	content transformation	implemented	OBJ-1 → EPI-OBJ-1	truth-apt transformation: indeterminate; candidates may be paraphrase/entailed, reshaping, or ampliative; the inspected checks do not decide which. The evidence link may preserve provenance to an event, but proposition warrant is unknown.	Proposed fact statement against cited event content	Rules or configured LLM proposes; check_llm_fact checks lasting flag, length, evidence occurrence, and structure; DomainValidator applies domain constraints. These do not test proposition truth or entailment.	Manual consolidation or thresholded session-end consolidation; conditional on events after checkpoint, configured extractor, and (for LLM) provider/key	Code can admit a proposal into a batch or quarantine it; no candidate instance observed	Publication makes the fact available to store consumers; no epistemic acceptance force	No truth warrant granted by source-quote matching or domain validation. Source linkage licenses tracing the cited text only.	Batch publication permits later search/read and possible hook/writeback retrieval; failed validation/generation blocks publication	Future search/read caller or hook host, via operation result or additionalContext; actual uptake uninspected	SRC-1 src/consolidate.rs; see EPI-RTE-1	none	none	No expected factual answer, independent reviewer, or evidence-consuming truth criterion was found. No observed batch or downstream use.
RTE-4	check/evidence production	implemented	EPI-OBJ-1	no content change	Whether the proposed evidence text occurs in the cited event, and whether candidate fields meet local constraints	Program checks text occurrence/order and format/domain constraints over the selected event and fact schema	During LLM proposal validation; only on the configured consolidation path	A proposal can pass or fail local checks; no observed result	Passing allows candidate to proceed operationally; it is not accepted as true	Supports evidence-text traceability to that event plus schema admissibility only	Pass permits batch consideration/publication; failure omits/quarantines candidate	Subsequent store consumers only if batch publishes; no use observed	SRC-1 src/consolidate.rs; quote in EPI-RTE-1	none	none	It does not establish that the quoted text supports the proposition, that the source is true, or that the candidate is useful.
RTE-4	disposition/acceptance	implemented	EPI-OBJ-1	no content change	Candidate admission to published fact tree	Code applies extraction disposition and domain validation; Git CAS admits the batch against current tree. No truth oracle or human per-fact approval is described.	At batch publication after candidate checks	Accepted/quarantined disposition is possible in code; no persisted observed candidate in this evidence set	Operational admission to durable memory; no epistemic acceptance	No criterion for truth-based acceptance identified; persistence is not acceptance	Permits candidate retention and later retrieval if admitted	Later search/read/hook consumers, conditional on eligibility and selection	SRC-1 src/consolidate.rs, src/repo.rs; RTE-4	CLM-2 concerns publication only	none	A disposition or successful publication cannot establish factual warrant; operation was not observed.
RTE-5	content transformation	implemented	caller assertion → EPI-OBJ-2	truth-apt transformation: acquisition/import (caller-supplied proposition; warrant unknown) for remember; truth-apt transformation: indeterminate for correction because the relationship between old and replacement propositions is not specified by validation	Caller-supplied statement and target fact/entity	Caller proposes; code validates request structure, target and domain constraints; no evaluator of truth	Explicit MCP or CLI change request	A validated fact change can be published; no observed request	Durable admission, not epistemic acceptance	Caller is the source; truth and support remain unknown	Permits a new fact or replacement/supersession to be published; stale base can block publication	Calling MCP/CLI client sees receipt; later consumers must retrieve or receive writeback/hook context	SRC-1 src/change.rs, src/validate.rs; RTE-5	CLM-2 covers normal publication protocol	none	Reason text and caller identity do not amount to an independent truth check. No request/result was observed.
RTE-5	disposition/acceptance	implemented	EPI-OBJ-2	no content change	Admission of caller's requested remember/correct change	Structural/domain validation and Git CAS, not truth evaluation	During explicit request	Publish or conflict/error possible; no result observed	Operational admission only	No fact-level epistemic authority conferred	Valid request can alter retained fact status/content; conflict blocks stale write	Caller gets operation result; any later behavioral path is conditional and uninspected	SRC-1 src/change.rs, src/validate.rs, src/repo.rs; RTE-5	CLM-2	none	Successful validation does not establish truth; no independent approver or answer oracle.
RTE-5	retention	implemented	EPI-OBJ-2 / OBJ-2	no content change for forget; erase removes/redacts stored representations and source history	Retrieval eligibility or requested erasure target	Explicit caller target; code controls retrieval suppression or repository/journal rewriting	Explicit forget/erase request	Suppression or erasure is possible; no operation observed	Changes availability/retention, not truth status	No claim about truth follows from forgetting, retraction, or erasure	Forget blocks ordinary later retrieval; erase removes matched material within its rewrite boundary	Future store readers; no later use observed	SRC-1 src/change.rs, src/erasure.rs, src/fact.rs; RTE-5	none	none	Forget preserves ordinary history while suppressing retrieval; erase is a separate history rewrite. Neither is lifecycle integration.
RTE-2 / RTE-3	operational admission/selection/consumption	implemented	OBJ-2 → EPI-OBJ-3	no content change	Query-to-fact relevance for returned ordering	Lexical match and optional JEV model score relevance to the query; JEV's stated question is answer relevance, not factual accuracy	On explicit search; automatic prompt hook uses facts-only lexical matching without reranking	Ordered result may be returned; no live response observed	Ranking/selection force over which facts are presented	No truth warrant; at most a query relevance ordering	Affects which facts are returned/prominent; does not publish or correct memory	MCP/CLI caller or Claude hook receives results; agent uptake uninspected	SRC-1 src/jev.rs; RTE-1–RTE-3; quote in EPI-RTE-2	none	none	Provider score correctness and effect on host behavior were not observed.
RTE-1 / RTE-2 / RTE-3	operational admission/selection/consumption	implemented	OBJ-2 → OBJ-3	no content change	Fact eligibility and selection by status, expiry, scope, prompt/query	Code filters retracted, superseded, expired and (by default) disputed facts and applies route-specific lexical/project selection	Hook prompt/start or explicit read/search; conditional on host setup and request	Eligible text can be emitted; no host session observed	Supplied context or tool result, with host authority treatment uninspected	Applicability and freshness filters do not endorse propositions	Permits contextual availability; does not force model reliance	Claude additionalContext or MCP/CLI result; actual consumption uninspected	SRC-1 src/fact.rs, src/hook.rs, src/ops.rs; RTE-1–RTE-3; quote in EPI-RTE-3	none	none	No evidence that a host loaded or acted on a returned fact.
RTE-6	truth-apt transformation	implemented	OBJ-2 → EPI-OBJ-4	truth-apt transformation: non-ampliative reshaping insofar as it renders selected stored statements into a managed block; no changed proposition is established	Selected active fact statements for a chosen project/file	Code filters eligibility and writes formatted natural-language content; operator chooses invocation and destination	Explicit operator command	File content can be written; no project-file output or host read observed	Makes selected claims available in a host instruction channel	Preserves the source fact's unverified status; rendering does not add warrant	Enables a later host to load the generated block	Host reading selected instruction file; force/horizon depends on host and is uninspected	SRC-1 src/cli.rs; RTE-6; BAP-3	none	none	Whether formatting preserves every proposition exactly and whether the host loads/follows it were not observed.
RTE-7	behavior/policy adaptation	implemented	no candidate truth-apt output	non-truth-apt policy/content update: host configuration registers hooks, MCP, or skill access	Selected host integration settings	Setup code edits configuration according to operator-selected target and existing state	Explicit mem setup or removal	Capability/configuration change is possible; host invocation not observed	Operationally makes access paths available; no direct host behavior is enforced	None	Changes which memory operations the host can invoke	Selected host runtime; actual hook/tool invocation and effect uninspected	SRC-1 src/setup.rs; RTE-7	none	none	Configuration capability does not establish retrieval, use, or behavioral effect.

Per-object lifecycle disposition

candidate object ID	relevant route IDs	transformation	discovery lifecycle / applicable warrant	missing evidence/limit
OBJ-1	RTE-4	truth-apt transformation: acquisition/import of source content, with source warrant unknown	Discovery lifecycle not applicable to imported source statements. The event text can be retained and traced to its source; no independent verification of its truth is established.	No event instance or source-side verification was supplied.
EPI-OBJ-1	RTE-4	transformation: indeterminate; possible non-ampliative reshaping/entailed derivation or ampliative conjecture. The passage check and source-event association preserve a trace link, not semantic entailment.	Candidate instance state: no instance observed. No lifecycle phases are upgraded from code alone; no ampliative lifecycle record is warranted because ampliation itself is not established.	Need an observed candidate and source event plus a semantic comparison establishing preservation/entailment or a non-entailed proposition; then candidate-linked evidence of any later phases.
EPI-OBJ-2	RTE-5	Remember is caller-content acquisition/import with warrant unknown; correction is indeterminate as to whether replacement entails, contradicts, or otherwise revises the old proposition	Lifecycle not applicable to caller-supplied content as imported. The change protocol does not accept the proposition epistemically.	No observed request; no independent truth evidence or relation criterion for corrections.
EPI-OBJ-3	RTE-2, RTE-3	No content change	No lifecycle record for EPI-OBJ-3: no candidate truth-apt output for this object; its relevance score has operational selection effect only.	No live score or observed downstream selection effect.
EPI-OBJ-4	RTE-6	Non-ampliative reshaping/rendering, contingent on proposition-preserving output; no new warranted content established	Lifecycle not applicable absent a new truth-apt proposition.	No generated file was supplied to compare against its selected facts; host consumption not observed.
OBJ-3	RTE-1, RTE-2, RTE-6	No independent candidate truth-apt output; carries or presents existing fact statements	No lifecycle record for OBJ-3: no new candidate truth-apt output for this object; relevant direct-adaptation or update routes: RTE-6.	Host interpretation and behavioral effect are outside observed evidence.

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
CLM-1	Hook errors are swallowed and hook returns successfully without context/sync output	SRC-1 src/hook.rs; implementation claim represented by CLM-1	Runtime account records the same bounded hook guarantee	RTE-1	None; no host operation observed	None; no comparison of host response	Fail-open behavior is a wired hook-level protocol; it says nothing about fact truth, successful context delivery, or model consumption	Actual host handling of exit/output uninspected
CLM-2	Normal memory writes publish validated compare-and-swap Git revisions	SRC-1 src/repo.rs, src/change.rs; implementation claim represented by CLM-2	Runtime account records bounded protocol purpose	RTE-4, RTE-5	None; no concurrent write or publication observed	None; no causal comparison	The protocol can reject malformed/stale writes and atomically publish admitted trees; it does not validate proposition truth	Erase, external edits, and embedding callers are outside the covered normal-write claim; factual warrant remains unestablished

No separate knowledge-production or truth-warrant claim was found within the inspected runtime/design claims. This does not establish absence outside the recorded inspection boundary.

Bounded conclusion

Instinctual Memory imports source text, proposes durable fact statements from journal inputs, accepts caller-supplied statements into operational storage, filters and ranks stored facts for retrieval, and can render selected facts into host instruction files. The source-event quote check establishes traceability of an evidence string to an event, while the schema and domain checks establish admissibility constraints. These checks do not establish that the source is true or that it supports the candidate proposition (EPI-RTE-1). No inspected route provides an evidence-consuming truth-acceptance criterion, so no produced claim is established as accepted knowledge. Consolidated fact transformation remains indeterminate between paraphrase/entailed content and ampliative conjecture; no candidate-linked execution evidence supports lifecycle phase findings.

Explicit remember/correct routes acquire a caller's proposition or replacement into the store without independent truth review. Search reranking assesses query relevance and affects presentation order, not truth warrant (EPI-RTE-2). Status and expiry filters affect retrieval applicability without endorsement (EPI-RTE-3). Writeback and hook/tool outputs can make stored claims available to a host; whether a host uses them or changes behavior is uninspected. Setup changes access capability, not truth status or enforced behavior. There is no observed acceptance, post-acceptance integration, or causal evidence that any route improves a consumer's factual performance.

Shared records

Operative objects

EPI-OBJ-1 — Consolidation-proposed fact statements

Identity: propositions proposed by rules or the configured LLM from journal events and eligible file content, with candidate evidence references, then represented as structured facts in entity files. Representational form: natural-language statements plus structured Markdown/YAML fields. Storage substrate: the memory.git tree, as a subset of runtime OBJ-2. Possible duplicate of the consolidated fact content in OBJ-2; this record isolates its producer and candidate evidence relationship. Evidence: SRC-1 src/consolidate.rs, src/fact.rs; see RTE-4.

EPI-OBJ-2 — Caller-supplied fact statements

Identity: fact propositions supplied by a user or agent in explicit remember/correct requests. Representational form: natural-language statement plus structured fact fields. Storage substrate: the memory.git tree, as a subset of runtime OBJ-2. Possible duplicate of caller-originated facts in OBJ-2; this record isolates their direct acquisition path. Evidence: SRC-1 src/change.rs, src/fact.rs; see RTE-5.

EPI-OBJ-3 — Query relevance scores and ordering

Identity: ephemeral relevance decisions/order over retrieved fact statements for one search query. Representational form: structured score/decision data. Storage substrate: transient operation result; no memory-tree publication. It is not truth-apt content. Evidence: SRC-1 src/jev.rs; see RTE-2 and RTE-3.

EPI-OBJ-4 — Writeback fact rendering

Identity: selected fact statements rendered in a generated managed block of a host instruction file. Representational form: natural-language text. Storage substrate: operator-selected project/host file, outside the memory revision. Possible duplicate of writeback material included in runtime OBJ-3; this record isolates its content-preserving transformation and host audience. Evidence: SRC-1 src/cli.rs; see RTE-6 and BAP-3.

Routes

EPI-RTE-1 — Consolidation evidence trace check

Source-native identity: check_llm_fact checks proposed fact fields and requires evidence text to occur in the cited journal event; successful proposals flow to consolidation disposition/publication. This is a distinct check function within runtime RTE-4, not an independent factuality evaluator. Target: proposed fact statement and its cited event, for source-text traceability and schema/domain admissibility. Evaluator: programmatic string/field checks; no expected factual answer. Possible duplicate of the evidence-check portion of RTE-4; reconcile by keeping RTE-4 as the complete route and this record as the epistemic check characterization. Evidence: SRC-1 src/consolidate.rs.

/// Accept a model-proposed fact only if its evidence is really in the event. fn check_llm_fact( fact: &LlmFact, event: &JournalEvent, entities: &BTreeMap, ) -> std::result::Result { let evidence = fact.evidence.trim(); let statement = fact.statement.trim(); if !fact.lasting { return Err("not lasting (true for one session only)"); } if evidence.chars().count() < 8 { return Err("evidence too short"); } if !(8..=600).contains(&statement.chars().count()) { return Err("statement empty or over 600 characters"); } if !quote_found(&event.content, evidence) { return Err("evidence not found in the cited event"); } --- src/consolidate.rs:908-927 @ 6acb13dc35765bf5ccfc87e445dd09c480f1c28a

EPI-RTE-2 — JEV relevance ordering

Source-native identity: JEV sends each candidate statement with the query and requests a relevance score based on whether the text answers that query; its output affects ranking. Target: relevance of a retrieved fact statement to a given search query. Evaluator: configured external JEV model/service; no factuality criterion. Possible duplicate of the reranking operation in RTE-2/RTE-3; reconciliation can merge this characterization into those runtime routes. Evidence: SRC-1 src/jev.rs.

"How well does this text answer the search query? Query: {}", clip(query, 400) ), "criteria": [ "Unrelated to the query", "Shares a topic but does not answer the query", "Partly answers the query", "Directly answers the query" ] --- src/jev.rs:143-151 @ 6acb13dc35765bf5ccfc87e445dd09c480f1c28a

EPI-RTE-3 — Fact retrieval eligibility

Source-native identity: Fact::is_eligible filters facts by status and temporal validity for current-context retrieval. Target: current retrieval applicability of a stored fact. Evaluator: deterministic status/time conditions; not a truth assessor. Possible duplicate of retrieval filtering described in RTE-1–RTE-3; retain those routes as canonical runtime records. Evidence: SRC-1 src/fact.rs.

/// True if the fact is eligible for current-context retrieval at now. /// /// Excludes superseded, retracted, expired, and (optionally) disputed facts. pub fn is_eligible(&self, now: DateTime, include_disputed: bool) -> bool { match self.status { FactStatus::Active => {} FactStatus::Superseded | FactStatus::Retracted | FactStatus::Expired => return false, FactStatus::Disputed if !include_disputed => return false, FactStatus::Disputed => {} } --- src/fact.rs:63-72 @ 6acb13dc35765bf5ccfc87e445dd09c480f1c28a

Claims

none declared in this member

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member