llm-wiki memory profile

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

Comparison rationale

This profile classifies the accepted records, not a running vault. The central maintenance and query routes are model-consumed doctrine. Their findings remain claimed; executable setup supports only its own wired findings. The external L1 consumer role is afforded. No finding claims observed use or benefit.

The memory inventory separates current pages, historical wording, live and archived routing, raw read traces, health metadata and L1 guidance. The filesystem is the primary substrate. Optional Git commits add a revision and recovery layer, not atomic admission. Model interpretation makes structured Markdown fields natural-language material on those consumer paths. The initializer's fixed processing consumes current setup inputs to produce retained configuration; it does not establish a symbolic later-consumption path for that retained configuration. The unsupported symbolic finding is withdrawn, not retained under partial coverage. The exact external L1 payload remains unresolved. See MEM-OBJ-config, MEM-OBJ-access-log and RT-RTE-wiki-setup.

Retained page descriptions rank later candidates even without learned scoring weights. Requested semantic selection remains pull, so it does not produce an inferred-judgment push signal. The external host's described automatic index loading supports coarse push at an afforded layer; unknown body/description selection remains separate. Neither a page identifier nor an index pointer proves targeted push. See MEM-OBJ-hub-index, RT-RTE-wiki-query and MEM-BAP-l1-loading.

Compression of existing material supports consolidation, not automatically synthesis of a new claim. Query answer semantics remain indeterminate without actual inputs and outputs. State replacement supports evolution and withdrawal of current reliance with retained history. Archive demotion lowers salience; it does not deny access. These distinctions follow MEM-OBJ-page-history, RT-RTE-source-ingest, RT-RTE-index-prune and RT-RTE-wiki-query.

Automatic raw logging and human-approved index revision are different admission mechanisms. The specified access-feedback chain therefore does not establish an automatic trace-learning write. The README's contrary automatic re-promotion claim remains a separate claimed agency alternative, not upgraded to wiring or silently treated as the confirmed skill route. Its original trace input is not specified as a qualifying learning chain in the records. Initial L1 provenance and admission are also unresolved. These limits and the lack of canonical bounded-absence records prevent controlled trace-learning negatives and trace-source inapplicability. See MEM-OBJ-access-log, RT-RTE-index-prune, EPI-CLM-automatic-repromotion and MEM-OBJ-l1-rules.

The reconciliation supersedes EPI-OBJ-access-log with MEM-OBJ-access-log; only the canonical latter record is used here. Its other object containers remain valid, but this profile uses the finer memory parts where their consumers or coverage differ. No other supersession changes a classification. Missing initial L1 control, exact host payloads and actual semantic outputs remain limitations, not invitations to reopen report verification.

No source files were read for this profile. All classifications use the supplied accepted records and their annotations; no evidence or analyst record is added.