Omni-Memory epistemic report

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

Source-and-claim boundary

Question: what licenses reliance on retained project claims, and which consumers give that license consequences? The selected whole-system scope and exclusions are those in Boundary and evidence; source identity and evidence layers are in its Source register. All new inspections below use SRC-1 at the same revision; checkout HEAD matched the pin. Implementation and doctrine/design establish routes, not exercised behavior. No execution evidence or causal comparison was supplied.

Assessed families: acquisition/extraction, noise admission, file-identity checks, trust-tier promotion/withdrawal, relevance feedback, duplication/conflict handling, quarantine/restoration, and delivery. Public/design warrant claims: RT-CLM-verified-context, EPI-CLM-earned-verified, EPI-CLM-hygiene-truth. No single system-wide negative follows from this coverage.

Entry point or operation	Source path (SRC-1)	Coverage and records	Limit / conclusion prevented
Frozen CLI dispatch and options	`omni_memory/cli.py`	Direct read of main/dispatch, hook dispatcher, recall, check, used, gc, doctor ranges; RT-RTE-local-edits	Setup/status/config/gain internals not fully read directly; no exhaustive command-correctness conclusion
Plugin hooks	`hooks/hooks.json`, `omni_memory/cli.py`	Direct event wiring and hook read; RT-RTE-capture-admission, RT-RTE-context-delivery	Actual host delivery not observed; no registered model-tool API appears in this inspected hook/CLI surface
Capture and model build/artifacts	`omni_memory/session_memory.py`, `omni_memory/cleanup.py`; supplied runtime: `omni_memory/llm.py`, `omni_memory/context.py`, `omni_memory/artifacts.py`	Direct capture/prompts/heuristics/batch/filter reads; RT-RTE-build-artifacts, EPI-RTE-candidate-filter, EPI-RTE-filter-admission	No provider outputs tied to routes; semantic fidelity and accepted novel claims not established
Remember/global, edit, merge, reanchor, resolve, forget	`omni_memory/store.py`, `omni_memory/serve.py`, `omni_memory/cli.py`	Direct store and dashboard mutation reads; supplied RT-RTE-local-edits; EPI-RTE-merge-disposition	Caller identity does not authenticate human review; reanchor clears flags without semantic test
Refresh / doctor / source integrity / graduation	`omni_memory/branch.py`, `omni_memory/staleness.py`, `omni_memory/cli.py`	Direct full_refresh, signature gate, recompute/reconcile/graduation/demotion, doctor entry; EPI-RTE-integrity-check, EPI-RTE-tier-graduation, EPI-RTE-tier-withdrawal	Graph propagation helpers and Git adapter internals not fully audited; no graph/source-enumeration completeness proof
Citation feedback and point of use	`omni_memory/cli.py`, `omni_memory/session_memory.py`; supplied `omni_memory/witness.py`	Direct citation parsing, bump_uses and callers; supplied RT-RTE-use-check	No semantic success result consumed; later agent response to warning unknown
Inject / recall / AGENTS / digest / OpenCode	Direct `omni_memory/rank.py`, `omni_memory/store.py`, `omni_memory/cli.py`; supplied `omni_memory/inject.py`, `omni_memory/agentsmd.py`, `omni_memory/install.py`	EPI-RTE-query-ranking, supplied RT-RTE-context-delivery, RT-BAP-agent-context	Different projections do not preserve identical caveats; full digest/graph/dashboard display transformations not directly inspected here
Branch merge / duplication / conflict	`omni_memory/store.py`, supplied `omni_memory/branch.py`	Direct overlap/reconcile/resolve/history reads; EPI-RTE-merge-reconcile, EPI-RTE-merge-disposition	Text overlap is not logical equivalence; no complete contradiction detector
gc / check / restore / purge	`omni_memory/eviction.py`, `omni_memory/store.py`, `omni_memory/cli.py`	Direct full sweep and score, store quarantine/restore/purgeable; EPI-RTE-eviction-check, EPI-RTE-quarantine-disposition	Heuristic withdrawal not epistemic refutation
Team / snapshot	Supplied `omni_memory/team.py`, `omni_memory/store.py`, `omni_memory/cli.py`	RT-RTE-shared-import; direct clamp_evidence	Import trust choice and Git author labels do not authenticate source truth
Health/system maps, savings, collector, browser, tests	`omni_memory/healthmap.py`, `omni_memory/systemmap.py`, `omni_memory/savings.py`, `omni_memory/collector.py`, `omni_memory/static/index.html`, `tests/`	Available but not deeply inspected in this member; boundary/runtime cover existence and roles	No system-complete negative about evaluators, map fidelity, savings or diagnostic correctness; test definitions would not supply executed outcomes

Missing run-linked transcripts, DB transitions and genuine context deliveries prevent conclusions that these routes operated. Missing downstream agent decisions prevent activation claims. Missing controlled interventions prevent benefit/component attribution. Provider internals and exact weights remain inaccessible in this boundary: see separately RT-CMP-agent-cli, RT-CMP-gemini, RT-CMP-openai, RT-CMP-anthropic. No local parameter-update route is established by those adapters; no inference about remote weight updates follows.

Epistemic-object inventory

Fields owned by records say see rather than recopy their descriptions.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-candidates	Project propositions or task/policy text; scope per item	see EPI-OBJ-candidates	SRC-1 `omni_memory/session_memory.py`	A kind/evidence label cannot individuate a true proposition
EPI-OBJ-memory	Retained textual project assertions; also todos and policies	see EPI-OBJ-memory	SRC-1 `omni_memory/store.py`	No supplied live DB or accepted candidate instance
EPI-OBJ-evidence	File identity/change claims; tiers and counters are symbolic metadata	see EPI-OBJ-evidence	SRC-1 `omni_memory/staleness.py`	Source identity does not establish text follows from sources
EPI-OBJ-artifacts	Generated repository explanations and maps	see EPI-OBJ-artifacts	SRC-1 `omni_memory/artifacts.py` via RT-RTE-build-artifacts	Content-specific relation not determinable without input/output pairs
RT-OBJ-inject-cache	None; delivery control	see RT-OBJ-inject-cache	SRC-1 `omni_memory/cli.py`	Delivery frequency is not endorsement
RT-CMP-agent-cli, RT-CMP-gemini, RT-CMP-openai, RT-CMP-anthropic	Distributed model state not inspectable as localized claims	see component records	SRC-1 `omni_memory/llm.py`	Hidden processing, exact version and parameter mutation not determinable

Authority-route ledger

All observed-run results are uninspected. Status below is architectural. Each function remains separate even where a shared route implements more than one function. Generic route details remain owned by the cited declarations.

Route ID: RT-RTE-capture-admission Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-candidates Content/update relation: truth-apt transformation: indeterminate Transition or check target: Raw transcript → model candidate text Evaluator/condition and domain: Model extracts under EXTRACTION_PROMPT; optional fallback heuristic has a separate transformation below Activation and timing: Capture/precompact/session-end or explicit capture Possible or observed result: Candidate array; no observed operation supplied. Implemented force: candidate generation only Epistemic authority and scope: JSON import acquires caller content; original warrant unknown; heuristic preserves selected wording but drops role context Operational authority: Proposals proceed to filtering, not truth acceptance Behavioral-authority path: model/heuristic → remember_many; permissive candidate supply; same invocation Evidence source ID and local anchor: SRC-1 omni_memory/session_memory.py; see RT-RTE-capture-admission. Claim IDs: none Mismatch marker: none Gap/limit: Model reshaping/entailed derivation/ampliative conjecture not decidable from instructions alone; need item-level inputs and outputs

Route ID: RT-RTE-capture-admission Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-candidates Content/update relation: truth-apt transformation: acquisition/import Transition or check target: Caller JSON → Python candidate list Evaluator/condition and domain: JSON parser plus accepted envelopes; no truth evaluator Activation and timing: Explicit capture or other JSON ingestion before remember_many Possible or observed result: Imported assertions; implementation only Implemented force: Candidate supply; admission separate Epistemic authority and scope: Original source warrant unknown; import does not produce a newly accepted claim Operational authority: Supplies items to batch filter Behavioral-authority path: caller → parser → remember_many; permissive; current invocation then retained rows if admitted Evidence source ID and local anchor: SRC-1 omni_memory/session_memory.py; see RT-RTE-capture-admission. Claim IDs: none Mismatch marker: none Gap/limit: Syntax and shape are not truth checks

Route ID: RT-RTE-capture-admission Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-candidates Content/update relation: truth-apt transformation: non-ampliative reshaping Transition or check target: Transcript lines → selected de-framed line text and category Evaluator/condition and domain: Deterministic regex / length rules; not semantic validation Activation and timing: Fallback heuristic or document scan Possible or observed result: Up to 40 candidate lines; implementation deduction below, not target execution Implemented force: Candidate generation and selection Epistemic authority and scope: Selected wording retained, role framing removed; speaker attribution is degraded Operational authority: Selected candidates proceed to stricter doc/heuristic filtering Behavioral-authority path: _heuristic_extract → cleanup → remember_many; permissive; same call Evidence source ID and local anchor: SRC-1 omni_memory/session_memory.py; see RT-RTE-capture-admission. Claim IDs: none Mismatch marker: none Gap/limit: No new semantic proposition derived by the regex; dropping role context does not establish warranted project assertion

Concrete source deduction: the single input line assistant: GET /orders returns orders. is stripped to GET /orders returns orders., passes the 12–240 character window, matches the first HTTP-method/path pattern and becomes an endpoint candidate with no file anchors. This is the extractor's result before cleanup, not a claim that it would be accepted or true. An assistant's assertion becomes de-framed project-memory text; preserving words does not preserve attribution.

line = _ROLE_PREFIX.sub("", line.strip("-* \t")).strip() --- omni_memory/session_memory.py

if len(line) < 12 or len(line) > 240: continue for pat, kind in _PATTERNS: if re.search(pat, line, re.I): out.append({"kind": kind, "text": line, "files": _real_files(root, line) if root else []}) break return out[:40]

--- omni_memory/session_memory.py

Route ID: RT-RTE-build-artifacts Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-artifacts Content/update relation: truth-apt transformation: indeterminate Transition or check target: Repository context → model facts/docs Evaluator/condition and domain: Prompted model; artifact length check only Activation and timing: Explicit build, with independent AI/docs generation switches Possible or observed result: Facts/docs; no observed operation supplied. Implemented force: candidate generation, length admission for docs Epistemic authority and scope: No semantic warrant established Operational authority: Stores facts or writes docs for later readers Behavioral-authority path: build model → store/artifact writer → context/dashboard; permissive; later reads Evidence source ID and local anchor: SRC-1 omni_memory/artifacts.py; see RT-RTE-build-artifacts. Claim IDs: none Mismatch marker: none Gap/limit: Truncated context and opaque outputs; not enough to call every result a novel conjecture

Route ID: EPI-RTE-candidate-filter Route function: check/evidence production Architectural status: implemented Object/candidate ID: EPI-OBJ-candidates Content/update relation: no content change Transition or check target: Candidate storage-worthiness, not propositional truth Evaluator/condition and domain: Python length, structure, chatter and anchor rules; doc/heuristic stricter Activation and timing: remember_many before storage Possible or observed result: keep/drop Boolean; no observed operation supplied. Implemented force: enforcing at batch gate Epistemic authority and scope: Only conformity to heuristic shape rules Operational authority: Blocked candidates do not enter batch storage Behavioral-authority path: remember_many → filter_items → kept list; enforcing; admission invocation Evidence source ID and local anchor: SRC-1 omni_memory/cleanup.py; see EPI-RTE-candidate-filter. Claim IDs: none Mismatch marker: none Gap/limit: Manual remember and dashboard bypass this batch gate

Route ID: EPI-RTE-filter-admission Route function: disposition/acceptance Architectural status: implemented Object/candidate ID: EPI-OBJ-candidates Content/update relation: no content change Transition or check target: Whether a batch candidate proceeds to remember Evaluator/condition and domain: filter result and source-sensitive ceiling; no semantic oracle Activation and timing: After candidate check Possible or observed result: Dropped count or admitted row; no observed operation supplied. Implemented force: enforcing admission Epistemic authority and scope: Storage eligibility, not evidence-consuming acceptance of meaning Operational authority: Admitted text becomes eligible memory Behavioral-authority path: capture/build → remember_many → SQLite; permissive; later sessions Evidence source ID and local anchor: SRC-1 omni_memory/session_memory.py; see EPI-RTE-filter-admission. Claim IDs: none Mismatch marker: none Gap/limit: Admission criterion is not truth acceptance

Route ID: EPI-RTE-staleness-check Route function: check/evidence production Architectural status: implemented Object/candidate ID: EPI-OBJ-evidence Content/update relation: no content change Transition or check target: Whether anchored files or mapped symbols changed since the memory's commit Evaluator/condition and domain: Git diff plus available code-graph symbol/dependent mapping; file fallback Activation and timing: recompute during check or refresh orchestration Possible or observed result: is_stale Boolean and touched anchors; implementation only Implemented force: Supplies separate flag update Epistemic authority and scope: Possible source-relevant change, not factual falsity Operational authority: Selects stale/clear metadata update Behavioral-authority path: recompute predicates → set_stale; enforcing locally; maintenance call Evidence source ID and local anchor: SRC-1 omni_memory/staleness.py; see EPI-RTE-staleness-check. Claim IDs: none Mismatch marker: none Gap/limit: Graph helpers not fully audited; unknown/missing commit/files may skip evaluation

Route ID: EPI-RTE-staleness-flag Route function: disposition/acceptance Architectural status: implemented Object/candidate ID: EPI-OBJ-evidence Content/update relation: non-truth-apt policy/content update: staleness metadata Transition or check target: Mark or clear possible staleness without rewriting memory Evaluator/condition and domain: is_stale and prior flag; Python decides Activation and timing: After recompute predicate Possible or observed result: SQLite stale flag, timestamp and touched files; implementation only Implemented force: Persisted warning/retention input Epistemic authority and scope: Alert to recheck, not acceptance/rejection of sentence truth Operational authority: Warnings on marked reads; uncited stale age contributes to eviction Behavioral-authority path: SQLite stale → rendering and eviction; advisory to host, retention scoring; subsequent invocations Evidence source ID and local anchor: SRC-1 omni_memory/staleness.py; see EPI-RTE-staleness-flag. Claim IDs: none Mismatch marker: none Gap/limit: Not every projection reproduces warnings; drift demotion is separate

Route ID: EPI-RTE-integrity-check Route function: check/evidence production Architectural status: implemented Object/candidate ID: EPI-OBJ-evidence Content/update relation: no content change Transition or check target: Identity/existence of recorded source-file bytes Evaluator/condition and domain: Git digest comparison; worst recorded anchor wins Activation and timing: reconcile / graduation during full_refresh; doctor also reconciles Possible or observed result: FRESH / DRIFTED / ORPHANED; legacy uncheckable; no observed operation supplied. Implemented force: measurement; disposition separate Epistemic authority and scope: Identity of compared sources only, conditional on declared anchor map Operational authority: Supplies demotion and promotion predicates; source-loss stale flag input Behavioral-authority path: reconcile/graduate → verdict → metadata consumers; enforcing locally; refresh horizon Evidence source ID and local anchor: SRC-1 omni_memory/staleness.py; see EPI-RTE-integrity-check. Claim IDs: EPI-CLM-earned-verified Mismatch marker: warrant narrower than claim Gap/limit: Not semantic entailment; incomplete source maps remain possible

Route ID: EPI-RTE-tier-graduation Route function: disposition/acceptance Architectural status: implemented Object/candidate ID: EPI-OBJ-evidence Content/update relation: non-truth-apt policy/content update: evidence tier Transition or check target: Promote active stated memory with positive uses and matching nonempty blob map Evaluator/condition and domain: Code decides on citation counter plus FRESH; no outcome-success evaluator Activation and timing: full_refresh after reconcile; skip on unchanged refresh signature possible Possible or observed result: evidence becomes verified; no observed operation supplied. Implemented force: ranking/retention metadata write Epistemic authority and scope: Evidence of citation and unchanged mapped bytes, not confirmed proposition Operational authority: Eligible query ranking boost and eviction shield Behavioral-authority path: SQLite evidence → rank.score_one / eviction.evict_score; ranking; until changed Evidence source ID and local anchor: SRC-1 omni_memory/staleness.py; see EPI-RTE-tier-graduation. Claim IDs: EPI-CLM-earned-verified Mismatch marker: citation is not outcome success Gap/limit: AI capture/build/doc ceiling inferred prevents direct graduation from those initial tiers

Route ID: EPI-RTE-tier-withdrawal Route function: disposition/acceptance Architectural status: implemented Object/candidate ID: EPI-OBJ-evidence Content/update relation: non-truth-apt policy/content update: evidence tier and stale flag Transition or check target: Revoke verified on mapped-source drift or loss Evaluator/condition and domain: reconcile digest verdict; legacy handling differs Activation and timing: Refresh or doctor reconcile Possible or observed result: verified → stated; orphan flagged stale; drift counted; no observed operation supplied. Implemented force: enforcing metadata update Epistemic authority and scope: Withdraws old source-identity assurance, not disproof Operational authority: Removes tier boost/shield on later score calculation Behavioral-authority path: reconcile → SQLite → rank/eviction; enforcing locally; later reads Evidence source ID and local anchor: SRC-1 omni_memory/staleness.py; see EPI-RTE-tier-withdrawal. Claim IDs: EPI-CLM-earned-verified Mismatch marker: none Gap/limit: Content remains; source drift is not necessarily factual error

Route ID: EPI-RTE-query-ranking Route function: operational admission/selection/consumption Architectural status: implemented Object/candidate ID: EPI-OBJ-memory Content/update relation: no content change Transition or check target: Which scoped active memories surface for a query Evaluator/condition and domain: Code lexical/file/graph scoring with uses and evidence multipliers Activation and timing: Store query path; ordinary empty-query browse differs Possible or observed result: Sorted matching rows; zero-match omitted; no observed operation supplied. Implemented force: ranking Epistemic authority and scope: Relevance signal only; not truth validation Operational authority: Changes memory order/delivery to model Behavioral-authority path: Store.memories → rank → inject/recall; ranking; each query Evidence source ID and local anchor: SRC-1 omni_memory/rank.py; see EPI-RTE-query-ranking. Claim IDs: EPI-CLM-earned-verified Mismatch marker: none Gap/limit: No-query/no-context path sorts by updated; boost is not universal

Route ID: EPI-RTE-merge-reconcile Route function: check/evidence production Architectural status: implemented Object/candidate ID: EPI-OBJ-memory Content/update relation: no content change Transition or check target: Text overlap / shared-symbol claim-kind collision Evaluator/condition and domain: Token-set Jaccard; anchor-relaxed threshold; kind/symbol rules Activation and timing: Add or merged-branch reconciliation Possible or observed result: Similarity or conflict classification; no observed operation supplied. Implemented force: heuristic result; disposition separate Epistemic authority and scope: Text overlap and co-anchoring only Operational authority: Feeds supersession / conflict recording Behavioral-authority path: Store._similar → duplicate/reconcile callers; enforcing local predicates; mutation invocation Evidence source ID and local anchor: SRC-1 omni_memory/store.py; see EPI-RTE-merge-reconcile. Claim IDs: none Mismatch marker: none Gap/limit: Negation can preserve high token overlap; distinct claims can share a symbol

Route ID: EPI-RTE-merge-disposition Route function: disposition/acceptance Architectural status: implemented Object/candidate ID: EPI-OBJ-memory Content/update relation: no content change Transition or check target: Retire overlap losers; mark collisions; caller resolves keep/both Evaluator/condition and domain: Code applies citation/tier/tie rules; caller decides explicit resolution Activation and timing: Add, branch merge, CLI/dashboard resolve or merge Possible or observed result: Superseded/active rows and conflict flags; dashboard merge unions files; no observed operation supplied. Implemented force: enforcing retrieval membership Epistemic authority and scope: No checked logical contradiction or truth winner Operational authority: Changes active retrieval set; conflict warns instead of blocking Behavioral-authority path: Store statuses/conflicts → active-row selectors and marker; enforcing membership/advisory conflict; later sessions Evidence source ID and local anchor: SRC-1 omni_memory/store.py; see EPI-RTE-merge-disposition. Claim IDs: none Mismatch marker: none Gap/limit: Not post-acceptance lifecycle integration; source-file union does not prove support

Route ID: EPI-RTE-eviction-check Route function: check/evidence production Architectural status: implemented Object/candidate ID: EPI-OBJ-memory Content/update relation: no content change Transition or check target: Retention eligibility, not falsehood Evaluator/condition and domain: Branch state, staleness, age, kind, confidence, citations, tier; protected exemption; retro-noise filter Activation and timing: gc and check; dry_run computes without mutation Possible or observed result: Score threshold / noise / purgeable lists; no observed operation supplied. Implemented force: selection recommendation consumed by sweep Epistemic authority and scope: Heuristic retention priority only Operational authority: Selects withdrawal candidates Behavioral-authority path: evict_score/is_noise → sweep → quarantine; enforcing unless dry run; maintenance invocation Evidence source ID and local anchor: SRC-1 omni_memory/eviction.py; see EPI-RTE-eviction-check. Claim IDs: EPI-CLM-hygiene-truth Mismatch marker: use is not demonstrated truth Gap/limit: Uses shield and verified tier can protect false but cited text

Route ID: EPI-RTE-quarantine-disposition Route function: disposition/acceptance Architectural status: implemented Object/candidate ID: EPI-OBJ-memory Content/update relation: no content change Transition or check target: Active retrieval membership / deletion Evaluator/condition and domain: sweep result; explicit purge; caller restore/forget Activation and timing: check/gc maintenance or explicit edits Possible or observed result: abandoned rows, optional deletion, restore active with noise_exempt; no observed operation supplied. Implemented force: enforcing membership Epistemic authority and scope: No semantic rejection or acceptance established Operational authority: Quarantined excluded from default active retrieval; restored reeligible Behavioral-authority path: SQLite status → Store.memories → context; enforcing; cross-session Evidence source ID and local anchor: SRC-1 omni_memory/eviction.py; see EPI-RTE-quarantine-disposition. Claim IDs: EPI-CLM-hygiene-truth Mismatch marker: none Gap/limit: Restore waives retro-noise, not score eviction; purge removes history rows

Route ID: RT-RTE-capture-admission Route function: retention Architectural status: implemented Object/candidate ID: EPI-OBJ-memory Content/update relation: no content change Transition or check target: Admitted rows → SQLite Evaluator/condition and domain: Store writes after caller-specific gates Activation and timing: Following admission Possible or observed result: Persisted memory plus anchors; no observed operation supplied. Implemented force: durable retention Epistemic authority and scope: Retention does not accept a proposition Operational authority: Makes active rows reachable in later calls Behavioral-authority path: SQLite → Store.memories → context; permissive; cross-session Evidence source ID and local anchor: SRC-1 omni_memory/store.py; see RT-RTE-capture-admission. Claim IDs: none Mismatch marker: none Gap/limit: No post-acceptance lifecycle integration established merely by storage

Route ID: RT-RTE-shared-import Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-memory Content/update relation: truth-apt transformation: acquisition/import Transition or check target: External shard/snapshot → local memory Evaluator/condition and domain: Shape/ID/noise validation and tier clamping; explicit trust exception Activation and timing: Sync/start/snapshot import Possible or observed result: Local rows; no observed operation supplied. Implemented force: admission and retention Epistemic authority and scope: External source warrant unknown; tier ceilings reduce label authority but do not validate meaning Operational authority: Makes imported content eligible for delivery Behavioral-authority path: JSON → importer → SQLite → context; permissive; later sessions Evidence source ID and local anchor: SRC-1 omni_memory/store.py; see RT-RTE-shared-import. Claim IDs: none Mismatch marker: none Gap/limit: Automatic import does not establish human review; later graduation can promote stated imported rows

Route ID: RT-RTE-use-check Route function: check/evidence production Architectural status: implemented Object/candidate ID: EPI-OBJ-evidence Content/update relation: no content change Transition or check target: Pinned mapped sources versus current file digests Evaluator/condition and domain: Witness comparison; bound>0 needed for valid coverage Activation and timing: used command or retrospective capture citation scan Possible or observed result: Clean/changed/missing-source diagnostics; no observed operation supplied. Implemented force: advisory to host; durable warning events possible Epistemic authority and scope: Source movement only, not claim truth or action success Operational authority: Warns revalidate/re-source; no action lock Behavioral-authority path: witness → stdout/stderr/events → host; advisory; current use/later session Evidence source ID and local anchor: SRC-1 omni_memory/witness.py; see RT-RTE-use-check. Claim IDs: RT-CLM-verified-context Mismatch marker: verification is source-bounded Gap/limit: See route limits, including observation-bound versus pin-time binding

Route ID: RT-RTE-context-delivery Route function: operational admission/selection/consumption Architectural status: implemented Object/candidate ID: EPI-OBJ-memory Content/update relation: truth-apt transformation: non-ampliative reshaping Transition or check target: Selected rows → narrower rendered context Evaluator/condition and domain: Scope/ranking/budget/clipping rules; AGENTS projection differs Activation and timing: Start/pull/auto prompt/OpenCode compaction Possible or observed result: Instructions and selected textual claims; no observed operation supplied. Implemented force: advisory content; software selects projection Epistemic authority and scope: No semantic license beyond underlying sources; clipping and caveat omission can degrade warrant visibility Operational authority: Host may rely on text without another enforced check Behavioral-authority path: See RT-BAP-agent-context Evidence source ID and local anchor: SRC-1 omni_memory/inject.py; see RT-RTE-context-delivery. Claim IDs: RT-CLM-verified-context Mismatch marker: verified heading is not per-row proof Gap/limit: Reshaping is projection only; does not certify preservation of complete propositions or causal chains

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
RT-CLM-verified-context	see claim	SRC-1 `omni_memory/skills/omni-memory/SKILL.md`; doctrine/design	see claim	RT-RTE-context-delivery, RT-RTE-use-check	none supplied	Host compliance external; no intervention	Instructions and source warnings wired	Not semantic verification or enforced host obedience
EPI-CLM-earned-verified	see claim	SRC-1 `omni_memory/staleness.py`; doctrine/design embedded in code	see claim	EPI-RTE-integrity-check, EPI-RTE-tier-graduation, EPI-RTE-tier-withdrawal, EPI-RTE-query-ranking	none supplied	Citation and identity can hold for false text	Narrow file-identity/citation qualification with consequential metadata	“held up” and “outcome + reality” exceed predicate
EPI-CLM-hygiene-truth	see claim	SRC-1 `omni_memory/eviction.py`; doctrine/design	see claim	EPI-RTE-eviction-check, EPI-RTE-quarantine-disposition	none supplied	No action outcome reference or component comparison	Reversible heuristic retrieval withdrawal	Citation is not demonstrable truth; quarantine is not refutation

Bounded conclusion

Omni-Memory acquires caller and shared assertions with unknown original warrant, retains them with code anchors, and supplies selected projections to later agents. Batch filtering is a storage-worthiness gate; direct manual/dashboard and explicit trusted import follow different admission rules. Model extraction/build requests project claims, but neither those instructions nor a storage write establish an evidence-consuming acceptance decision for an ampliative proposition. Semantic transformation remains indeterminate without content-specific pairs; JSON acquisition and code-controlled selection can be identified independently. See RT-RTE-capture-admission, RT-RTE-shared-import, RT-RTE-build-artifacts, EPI-RTE-filter-admission.

The strongest inspected deterministic check concerns source-file identity. Its target is not the truth of the memory sentence. Tier promotion consumes a nonempty matching blob map and a positive citation count for active stated rows; withdrawal consumes drift/loss. The implementation can therefore rank and protect content without demonstrating its truth. AI-session/AI-build/doc inputs are capped at inferred here; that is a real countercase to a claim that all captured content can immediately graduate by citation. Explicit/manual and trusted import routes are countercases to a universal earned-tier rule. See EPI-RTE-integrity-check, EPI-RTE-tier-graduation, EPI-RTE-tier-withdrawal, RT-RTE-local-edits.

Citations change uses and later ranking/retention even without semantic success. Reconciliation heuristics alter active membership and conflict flags; callers can resolve, merge, restore or reanchor without a truth test. Quarantine and purge are retention decisions, not falsification. These routes are operational adaptation of memory selection, not evidence of improved future capacity. See EPI-RTE-query-ranking, EPI-RTE-merge-disposition, EPI-RTE-quarantine-disposition.

No inspected route here establishes post-acceptance lifecycle integration of a newly warranted claim. That bounded finding is not a system-complete absence: map/diagnostic/browser internals and provider processing have the explicit coverage limits above. Context delivery and point-of-use warnings have advisory force through RT-BAP-agent-context; metadata/status decisions enforce local ranking and membership but not external agent obedience. No demonstrated benefit, learning, exercised self-improvement or method-directed theory criticism follows from these source-only findings. Static guidance persists while claims/counters change; independent learning, reflection, autonomy and theory-builder conclusions must not be collapsed into one verdict.

Shared records

Components

none declared in this member

Operative objects

EPI-OBJ-candidates

Label: Extracted memory candidates

SRC-1 omni_memory/session_memory.py. Natural-language text plus symbolic kind/files/symbols/evidence fields in transient Python JSON-compatible arrays. Inputs are caller JSON, model output or heuristic-selected text; source warrant unknown. No supplied runtime operative-object counterpart: RT-RTE-capture-admission and RT-RTE-build-artifacts describe routes, not this candidate object. Not all items are truth-apt: todos are prospective tasks. Project fact/flow/decision assertions are truth-apt over the project/revision they describe.

array; each item: {{"kind": one of {_KINDS_LINE}, "text": one concise sentence, "files": [paths], "symbols": [names], "evidence": "stated"|"inferred"}}.

--- omni_memory/session_memory.py

EPI-OBJ-memory

Label: Retained memory text

SRC-1 omni_memory/store.py, omni_memory/session_memory.py. Natural-language memory.text in SQLite, indexed by symbolic IDs/status/branch/anchors. Distinct from RT-OBJ-inject-cache (delivery digest/counter) and from metadata part below. Closest route containers: RT-RTE-capture-admission, RT-RTE-local-edits, RT-RTE-shared-import; no object counterpart supplied. In-memory text can be replaced by caller updates; retained text reaches later retrieval/projections without mandatory truth acceptance.

m = Memory(text=text.strip(), kind=kind if kind in KINDS else "fact", branch=branch, files=files or [], symbols=symbols or [], commit_range=commit, source=source, evidence=evidence, blob_shas=shas, observed=observed, unbound=unbound, author=gitmeta.git_user(root))

--- omni_memory/session_memory.py

EPI-OBJ-evidence

Label: Evidence and source-integrity metadata

Part of: EPI-OBJ-memory

The retained memory record contains this symbolic operative part; the parent above denotes the retained memory text with its row, not merely an isolated string. SRC-1 omni_memory/store.py, omni_memory/staleness.py. SQLite evidence tier, blob map, observation flags, citation counts and staleness flags are consumed by fixed code rules separately from natural-language text. Witness and read-ledger tables support the source-check paths in RT-RTE-use-check. No supplied object counterpart; not a duplicate of the injection cache.

evidence: str = "stated" --- omni_memory/store.py

EPI-OBJ-artifacts

Label: Generated API map and linkup

SRC-1 omni_memory/artifacts.py; evidence/passage supplied by RT-RTE-build-artifacts. Natural-language Markdown documents with diagram/code notation, stored as .omni-memory/api-map.md and linkup.md. Human/dashboard readers interpret explanations; no formal proof checker established. Distinct from retained SQLite rows: generated files have independent length-based replacement and no inspected version rollback. No supplied object counterpart.

Routes

EPI-RTE-candidate-filter

Label: Candidate noise check

Part of: RT-RTE-capture-admission

SRC-1 omni_memory/cleanup.py. remember_many → cleanup.filter_items → Boolean keep/drop; Python owns check, candidates remain transient.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Kept list and dropped count.
Later read-back: inapplicable — transient check alone retains nothing
Delegated visibility: inapplicable — no delegation in deterministic filter
Selection predicate: All batch items; structural/length/chatter checks for all, concrete-anchor check additionally for doc/heuristic.
Invalidation or expiry: inapplicable — check result exists for current call only
Activation or effect: Controls which candidates pass to batch admission.
Evidence limits: This is storage-worthiness, not evidence about truth.

if t.endswith("?"): # questions aren't durable facts return True --- omni_memory/cleanup.py

if source not in _STRICT_SOURCES: return False # trust AI/manual/session prose --- omni_memory/cleanup.py

EPI-RTE-filter-admission

Label: Filtered candidate disposition

Part of: RT-RTE-capture-admission

SRC-1 omni_memory/session_memory.py. Kept candidates → remember → add_memory; code owns batch admission and evidence ceiling.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Accepted/dropped counts or failure.
Later read-back: Admitted rows reach later retrieval through RT-RTE-context-delivery
Delegated visibility: Optional model proposes upstream; filter and writer are local.
Selection predicate: Filter pass then source-tier clamp; inferred for ai-build/ai-session/doc ceilings, stated for other non-human labels.
Invalidation or expiry: Later edit/forget/maintenance can change or retire rows.
Activation or effect: Permits storage; does not establish changed host behavior.
Evidence limits: Direct remember/dashboard admission differs; no semantic oracle.

Decision roles and admission: code applies the predicates above; caller/model proposes upstream text, and caller can later edit/forget/restore where applicable. No authenticated human per-item veto is mandatory. Open-ended memory operation, not a bounded experiment/curriculum. No semantic answer oracle; hash equality is a reference only for file identity. Static extraction/usage guidance shapes proposals, not a learned method revision. What persists is memory text and metadata, not validated criticism or provider reasoning. Rejection/recovery limits are stated above.

items, dropped = cleanup.filter_items(items, source=source) n = 0 for it in items: --- omni_memory/session_memory.py

ceiling = "inferred" if source in ("ai-build", "ai-session", "doc") else "stated" return ev if _EVIDENCE_RANK[ev] <= _EVIDENCE_RANK[ceiling] else ceiling

--- omni_memory/store.py

EPI-RTE-staleness-check

Label: Changed anchor check

SRC-1 omni_memory/staleness.py. Endpoints: commit/file/symbol anchors and code graph → changed-file/affected-symbol intersections → is_stale. Code owns evaluation. No supplied counterpart after comparison with RT-RTE-use-check: this checks commit-relative source changes rather than retrieval pins.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Changed-anchor predicate and touched anchors, internally.
Later read-back: inapplicable — the check itself is transient; separate flag update persists results
Delegated visibility: inapplicable — deterministic local source check
Selection predicate: Active memories with files and commit anchor; graph-resolved symbol IDs when available, file-level fallback otherwise.
Invalidation or expiry: inapplicable — predicate recomputed against current source state
Activation or effect: Supplies staleness flag decisions, not semantic disproof.
Evidence limits: Full graph propagation and Git error behavior not audited here; code changes need not invalidate historical decisions.

is_stale = bool(mem_ids & affected_cache[anchor]) touched = sorted(mem_ids & affected_cache[anchor])

--- omni_memory/staleness.py

hit = files & changed_files is_stale = bool(hit) touched = sorted(hit) --- omni_memory/staleness.py

EPI-RTE-staleness-flag

Label: Staleness flag update

SRC-1 omni_memory/staleness.py. Endpoints: changed-anchor verdict → SQLite stale metadata → later context/eviction. Code owns flag setting/clearing; no supplied counterpart after comparison with RT-RTE-use-check, which emits retrieval-time warnings instead.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: checked/stale/cleared counters.
Later read-back: Persisted stale flag and timestamp feed warnings and uncited stale-age eviction scoring.
Delegated visibility: External host receives warnings only where the selected projection emits them.
Selection predicate: is_stale sets flag; previously stale but now unaffected clears it.
Invalidation or expiry: Next recompute can clear flags; explicit reanchor also clears them.
Activation or effect: Warning authority and retention priority change; text itself stays unchanged.
Evidence limits: No observed response to warnings; not semantic acceptance or rejection.

Decision roles: code evaluates and changes flags; caller launches check or refresh, not per-item approval. No answer oracle for sentence truth; open-ended maintenance. Caller may reanchor/edit/forget later; static guidance says reverify stale claims, but no completed revalidation is established.

if is_stale: store.set_stale(r["id"], True, None if was_stale else now, touched) stale += 1 elif was_stale: store.set_stale(r["id"], False, None, []) cleared += 1

--- omni_memory/staleness.py

EPI-RTE-integrity-check

Label: Source integrity check

SRC-1 omni_memory/staleness.py. Recorded blob map → current Git hashes → verdict; reconcile/graduate own checks. Compared with RT-RTE-use-check: same class of hash comparison but capture-anchored integrity instead of retrieval-time witness; no duplicate route.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Counters/verdicts; doctor prints coverage.
Later read-back: Stored flags/tiers can affect subsequent retrieval; verdict itself recomputed.
Delegated visibility: inapplicable — local Git/SQLite check, no model delegation
Selection predicate: Active rows with declared anchors; nonempty recorded map checked file by file.
Invalidation or expiry: Recomputed on refresh/doctor; no-source legacy rows uncheckable or name-resolved.
Activation or effect: Feeds separate promotion/demotion; no semantic inference.
Evidence limits: Scope is only mapped files; completeness and fidelity of chosen anchors unproved.

for path, sha in shas.items(): cur = current.get(path) if current is not None else gitmeta.blob_sha(root, path) if not cur: return "ORPHANED" # object missing at HEAD → deleted if cur != sha: worst = "DRIFTED" return worst --- omni_memory/staleness.py

EPI-RTE-tier-graduation

Label: Verified-tier promotion

SRC-1 omni_memory/staleness.py. Active stated rows + positive uses + nonempty FRESH blob map → SQLite verified; full_refresh owns transition. No supplied counterpart: related to but distinct from RT-RTE-use-check.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Promotion count.
Later read-back: Tier consumed by query rank and eviction; no-query browse can bypass tier weighting.
Delegated visibility: inapplicable — local transition, no delegation
Selection predicate: status active, evidence stated, uses>0, shas nonempty, all compared hashes match.
Invalidation or expiry: reconcile can demote on drift/loss; callers can edit tiers.
Activation or effect: Grants rank/eviction authority to same text, not semantic acceptance.
Evidence limits: Refresh signature can defer check; initial inferred AI/doc rows do not satisfy stated predicate.

"SELECT id, blob_shas, source FROM memory WHERE status='active' " "AND evidence='stated' AND COALESCE(uses,0) > 0" --- omni_memory/staleness.py

if _integrity_verdict(root, shas, current) == "FRESH": store.db.execute("UPDATE memory SET evidence='verified' WHERE id=?", (r["id"],)) n += 1

--- omni_memory/staleness.py

EPI-RTE-tier-withdrawal

Label: Verified-tier withdrawal

SRC-1 omni_memory/staleness.py. Reconcile verdict → _demote / stale flag → SQLite. Distinct from RT-RTE-use-check warnings: this mutates persisted qualification.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Reconcile counters.
Later read-back: Changed tier/flag consumed by later ranking, maintenance and projections.
Delegated visibility: inapplicable — no delegation
Selection predicate: Verified rows with nonempty recorded blob map and drift/loss; orphan stale handling broader.
Invalidation or expiry: Later matching stated+cited row may graduate again; caller can intervene.
Activation or effect: Removes metadata authority without rewriting text.
Evidence limits: Drift does not disprove a claim; AGENTS projection need not surface per-row flags.

elif verdict == "DRIFTED": drifted += 1 # counted; recompute owns staleness _demote(store, r) --- omni_memory/staleness.py

if row["evidence"] != "verified": return store.db.execute( "UPDATE memory SET evidence='stated' WHERE id=? AND evidence='verified'", (row["id"],))

--- omni_memory/staleness.py

EPI-RTE-query-ranking

Label: Citation and tier ranking

Part of: RT-RTE-context-delivery

SRC-1 omni_memory/rank.py. Scoped rows → rank.score_one → sorted hits → Store.memories; Python owns relevance decision.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Ranked matching list.
Later read-back: Uses/evidence persist and influence later query invocations; ranking results need not persist.
Delegated visibility: External host sees only downstream selected projection, not all rows.
Selection predicate: Positive lexical/file/graph score; uses/log feedback and verified ×1.2 or inferred ×0.8; empty query handled separately.
Invalidation or expiry: Recomputed each query as rows/config/state change.
Activation or effect: Ranking changes which text can reach host; actual compliance unobserved.
Evidence limits: Score is relevance/priority; no reference answer or correctness assessment.

score *= 1.0 + _USE_W * math.log(1 + (m.get("uses") or 0)) # citation feedback --- omni_memory/rank.py

score *= {"verified": 1.2, "inferred": 0.8}.get(m.get("evidence"), 1.0) --- omni_memory/rank.py

EPI-RTE-merge-reconcile

Label: Duplicate and conflict check

SRC-1 omni_memory/store.py. New/merged row and peers → text overlap/symbol-kind check → duplicate/conflict result. Distinct from supplied RT-RTE-local-edits caller route and RT-RTE-capture-admission ingestion route; this shared check serves both and branch movement.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Similarity/conflict classification internal to callers.
Later read-back: inapplicable — check itself does not retain outcome; disposition does
Delegated visibility: inapplicable — code comparison, no delegated model
Selection predicate: Same branch/kind active peers for new-row dedup; incoming branch rows against base on merge.
Invalidation or expiry: inapplicable — predicates evaluated against current peer texts
Activation or effect: Supplies membership mutation and conflict-recording decisions.
Evidence limits: No logical equivalence or contradiction checker; heuristics admit false positives.

jacc = len(ta & tb) / len(ta | tb) return jacc >= thresh --- omni_memory/store.py

if shared_sym and m["kind"] in _CONFLICT_KINDS: anchor = ", ".join(sorted(m_syms & r_syms)) self.record_conflict(m["id"], r["id"], anchor) return "conflict"

--- omni_memory/store.py

EPI-RTE-merge-disposition

Label: Supersession and conflict disposition

SRC-1 omni_memory/store.py. Similarity/collision result or caller decision → superseded rows/conflicts/merged files; store owns mutations. Overlaps supplied RT-RTE-local-edits and RT-RTE-capture-admission but distinct shared submechanism, not wholly part of one.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Mutation counts or status; resolve reports result.
Later read-back: Statuses constrain active reads, conflicts mark ranked injection; history follows supersedes chain.
Delegated visibility: inapplicable — no model call; local caller may itself be an agent
Selection predicate: Duplicate thresholds; tier/uses guards on relaxed insertion; merge reconciliation prefers citation/tier strength; caller keep/both.
Invalidation or expiry: Caller can later edit or retain both; no automatic semantic rollback established.
Activation or effect: Enforces membership; conflict flag advisory. Dashboard merge unions files then forgets drops without synthesizing text.
Evidence limits: Protected insertion peers exempt, not a universal proof of protection across every route.

if relaxed and not strict: if r["uses"] > 0 or _EVIDENCE_RANK.get(r["evidence"], 1) > m_rank: continue self.db.execute("UPDATE memory SET status='superseded',updated=? WHERE id=?", (time.time(), r["id"])) m.supersedes_id = r["id"]

--- omni_memory/store.py

self.update_memory(keep_id, files=files) n = 0 for did in drop_ids: if did != keep_id and self.forget(did): n += 1 return n --- omni_memory/store.py

EPI-RTE-eviction-check

Label: Hygiene score and noise check

SRC-1 omni_memory/eviction.py. Active row metadata / branch state → evict_score and retro-noise → sweep lists; code owns evaluation. Distinct from RT-RTE-local-edits umbrella maintenance callers.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Candidate/noise/purgeable lists.
Later read-back: inapplicable — score computation itself retains nothing
Delegated visibility: inapplicable — no delegation
Selection predicate: Threshold meta default 0.5; protected returns 0; source/tier/uses/kind/staleness/branch factors; noise requires uncited nonprotected nonexempt.
Invalidation or expiry: inapplicable — ephemeral recomputation on maintenance
Activation or effect: Supplies quarantine/deletion decisions; dry run does not apply them.
Evidence limits: Uses is not outcome success; confidence is stored input, not measured truth probability.

if evidence == "inferred": score += 0.15 elif evidence == "verified": score -= 0.30 --- omni_memory/eviction.py

if uses > 0: # citation shield — proven-useful memory resists eviction score -= min(0.5, 0.15 * math.log(1 + uses)) --- omni_memory/eviction.py

EPI-RTE-quarantine-disposition

Label: Quarantine purge and restoration

SRC-1 omni_memory/eviction.py. sweep/caller → quarantine/forget/restore/delete → SQLite status or deletion; code/caller own changes. Distinct from broad RT-RTE-local-edits route; material maintenance subroute with separate evaluator.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Counts/IDs; restore can return failure.
Later read-back: Abandoned excluded by default active selector; restored rows reappear; deleted rows unavailable.
Delegated visibility: inapplicable — no delegation
Selection predicate: Non-dry-run candidates/noise; purge explicit and grace-aged uncited quarantine; caller chooses restoration.
Invalidation or expiry: Restore marks noise_exempt; purge irreversible at this API.
Activation or effect: Changes future retrieval membership.
Evidence limits: No outcome-based semantic acceptance/rejection; ordinary false text may survive.

if not dry_run: for c in candidates: store.quarantine(c["id"], f"{c['reason']} (score {c['score']})") for c in noise: store.quarantine(c["id"], "extraction noise (conversational/thread)")

--- omni_memory/eviction.py

"UPDATE memory SET status='active', quarantined_at=NULL, " "quarantine_reason=NULL, noise_exempt=1 WHERE id=? AND status='abandoned'", --- omni_memory/store.py

Claims

EPI-CLM-earned-verified

Label: Earned verified tier

SRC-1 omni_memory/staleness.py. Doctrine/design in graduation comments claims citation plus re-fetchable unchanged anchors warrants verified and means a cited memory held up. Closest supplied RT-CLM-verified-context concerns enforced usage; this is the distinct warrant attributed to tier qualification. Implementation supports predicates and their consequences, not outcome success.

blob sha still matches HEAD) AND the agent has cited it (uses>0). Content can never self-declare verified; it must be earned by outcome + reality. --- omni_memory/staleness.py

locally. And uses > 0 means you cited it and it held up. Excluding shared memory --- omni_memory/staleness.py

EPI-CLM-hygiene-truth

Label: Citation shield as truth

SRC-1 omni_memory/eviction.py. Doctrine/design interprets continued citation as demonstrated truth and calls quarantine false/dead-memory hygiene. Distinct from RT-CLM-verified-context (use instructions) and EPI-CLM-earned-verified (tier promotion): this claim licenses retention protection. The score consumes citations, not tests of the cited assertion.

The citation shield means a memory the agent keeps using never gets evicted, even on a dead branch — because it's demonstrably still true. --- omni_memory/eviction.py

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member

Keys	Action
`?`	Open this help
`n`	Next page
`p`	Previous page
`s`	Search