ai-modules memory profile

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

Comparison rationale

The comparison preserves local evidence strength. Model-interpreted procedures supply claimed transformations and admissions; executable hashing, policy matching and benchmark consumption supply wired mechanisms only at those parts. No inspected execution or causal comparison upgrades them to observed or causally supported findings. Coverage remains partial even where a positive file witness exists.

Reconciliation supersedes EPI-OBJ-wiki-pages with MEM-OBJ-wiki-pages. All page-object findings therefore cite the memory declaration; epistemic composition/remediation routes remain distinct consumers/functions, not superseded records. The raw-sidecar and Lint-settings containers remain valid alongside their finer body/metadata and acceptance parts. No other supersession changes a value.

Admission and lineage must not be conflated. RT-RTE-wiki-triage extracts visible-session content, but human selections make page admission manual. The positive automatic trace-learning witnesses are instead MEM-RTE-finding-acceptance and RT-RTE-readiness-repair. Their later checking-policy and task-contract consumers satisfy the write-route definition without establishing improved capacity. Experience supplies learning force at the update consumer; its outputs separately supply validation/enforcement or instruction downstream.

MEM-OBJ-raw-body retains the maintenance instruction conflict: general read-only raw content versus source-drift re-recording. Hub re-ingest supports revision, but neither immutable preservation nor unrestricted maintenance replacement follows. EPI-RTE-wiki-compose leaves semantic novelty indeterminate, so the word synthesis does not establish synthesize. EPI-RTE-wiki-archive supports invalidate only for active reliance/navigation; archived bodies remain searchable rather than revoked.

Requested retrieval is pull even when index/topic selection or fixed matching occurs internally. The inspected pull-only units make push signals locally inapplicable; unresolved remaining consumers prevent a whole-axis inapplicable judgment. Static guidance installation and SPR output intent do not supply accumulated-memory push. MEM-CLM-summary-transfer specifies no retained checkpoint save/reload route.

Initial corpus provenance, image consumption, remaining output-consumer paths, exact automatic lesson trace mixtures and several admission alternatives remain unresolved. These prevent complete-value statistics and system-wide negatives. Opaque inference weights are outside this memory inventory because RT-CMP-host-model establishes inference use, not accumulated provider memory. It is not evidence that learning is absent elsewhere.

Source-read log: none. This member was classified only from the supplied accepted records and reconciliation; no frozen-source ambiguity required direct inspection.