deja-vu memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

The comparison inventories memory accumulated or revised through use, not all runtime files. Local binary/gob indexes and dense sidecar files establish filesystem storage. Dense encoding separately establishes parametric form; it does not establish a vector database or mutable model weights. The provider's exact state and retention remain uninspected (RT-CMP-embed-client, RT-RTE-embedding-records). Every axis remains partial because material alternatives or initial provenance/control remain unresolved in the accepted accounts. This is coverage uncertainty, not a downgrade of wired local witnesses.

Canonical fix objects are MEM-OBJ-fixraw and MEM-OBJ-fixes, replacing the epistemic duplicate declarations. Canonical convention selection is MEM-RTE-convention-delivery, replacing the epistemic duplicate route. These supersessions affect storage, form, lineage, authority, curation and read-back citations. They preserve the epistemic consumer-specific checks and first-sentence scope-loss finding; they do not supersede the index container or the distinct check/disposition/retention routes. The exact session-fact capacity is not used: the reconciliation retains a precision limit while command/outcome distinctions remain usable (MEM-OBJ-sessionfacts).

Trace learning here means automatic trace-fed production of durable behavior-shaping artifacts and a later consumer. Fixed fix mining and continuation extraction meet that architectural test without model training, a new explanatory claim or observed improvement. Raw recorded traces alone do not. Retained raw sightings have learning authority at the rejudgment consumer, while served remedies have knowledge and ranking force; flags can additionally validate/veto the plan-remedy branch. A requested lookup may return labelled unconfirmed candidates, so the plan gate must not become a system-wide never-served assertion (MEM-RTE-trace-mining, EPI-RTE-fix-disposition, EPI-RTE-fix-selection, EPI-RTE-fix-retention).

Extracting standing facts or carry from retained sessions is consolidation even when extraction loses a qualification. It is not synthesis merely because fidelity is uncertain. The separate compaction outcome-label edge does add an assertion: the accepted epistemic recovery overlay identifies a passed label inferred from no detected friction, compatible with an undetected failure. That local synthesis finding is neither a truth warrant nor a blanket classification of the entire packet (RT-RTE-compaction-recovery, MEM-OBJ-compaction-data). Lifecycle changes revise current state and withdraw convention reliance without erasing history; accepted convention delivery raises salience, while forgetting and retention caps have distinct local decay effects (MEM-OBJ-notes, MEM-RTE-forget-cleanup).

Standing-rule adoption and copying have different agency and evidence. The skill prescribes human selection before append, so manual agency and semantic consolidation on that path are claimed. Software distribution is wired and automatic. The intended later instruction-loader consumer is afforded, not observed; structural copy/status checks cannot prove loading or obedience. Direct edits and initial rule content remain unresolved. The workflow annotation sharpens those limits without declaring another route (RT-RTE-rules-sync, RT-BAP-standing-rules). Imperative conventions and carried constraints retain their requested instruction force even when framed as untrusted history; effective priority/compliance is not asserted (RT-BAP-recall-reference).

Explicit queries, URI/prefix reads and requested restoration are pull. Event-driven startup, prompt/tool/plan/spawn and continuation supply are push. Identifier push requires a real project or session/workspace match selecting delivered parts; an identifier on requested history alone is insufficient. Coarse recency/budget selection is kept apart from lexical targeting. Dense similarity on requested recall does not establish embedding-driven push. Rule loader inclusion is a separate coarse push affordance. A requested update-code read of old sightings/carry is not another model-facing push (RT-RTE-requested-recall, RT-RTE-auto-delivery, RT-RTE-spawn-context, MEM-RTE-session-recovery).

Session turns and recorded tool requests/results are the supported original trace inputs. Triggering hooks do not turn their payload into event-stream learning sources; error/remedy sequence mining is not sufficient evidence of linked state-action-outcome trajectories. Opaque provider inputs, specialized queued/checkpoint forms and incompletely traced sidecar/usage consumers remain explicit unresolved units, not absence or inapplicability. No accepted bounded absence record supports a local trace-learning no. No execution or causal support is claimed.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no target source was read and no new evidence or records were added.