llm-wiki memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison keeps filesystem bodies separate from access metadata and from opaque referenced payloads. A Markdown dataset manifest does not classify the dataset it names, and a readable session summary does not classify external host compaction state. These limits follow MEM-OBJ-datasets, MEM-OBJ-projects, MEM-OBJ-session-digest and RT-RTE-adapter-execute. Every axis remains partial because its included inventory retains unresolved alternatives; no unsupported absence or system-wide trace-learning no is inferred.
Evidence layers remain local. Model-operated compilation, guide use, semantic review, lesson extraction and user-choice refresh are claimed protocols. Hook state/digest/feedback transformations and context delivery, structural controls and guarded fixture delivery have wired mechanisms. Neither establishes observed operation, useful learning or causal benefit. See RT-RTE-compile-wiki, MEM-RTE-extract-lessons, MEM-RTE-feedback-candidates, RT-RTE-session-bridge and RT-RTE-query-evaluation.
Automatic trace-fed production of durable continuation context meets the profile's trace-learning definition even though the digest is a metadata template rather than semantic transcript compression. The actual automatic delivery is the state-derived summary and digest pointer, not the whole digest. User-turn feedback is session-log input despite its hook envelope; lifecycle callbacks are event-stream input; recorded tool-event descriptors are tool-trace input without claiming access to tool-result bodies. Manual-summary provenance remains unresolved. Update-consumer learning authority and downstream knowledge authority are separate paths. See MEM-OBJ-session-state, MEM-OBJ-session-digest, MEM-RTE-feedback-candidates and RT-BAP-session-context.
Requested query, rehydrate and fixture delivery are pull. The independent hook route is push: exact cwd matching selects eligible states, while digest availability, recency and count limit supply coarse filtering. No inference from a requested file identifier to identifier-based push is made. See RT-RTE-query-wiki, RT-RTE-session-bridge and RT-RTE-query-evaluation.
Curation values classify transformations rather than names. Bounded operational windows forget old entries locally. Feedback-candidate promotion and digest-body promotion separately raise selected operational content into the topic raw-note plane. Refresh flags lower confidence and insert a warning; the accepted record does not establish withdrawal of current reliance while retaining history, so flagging supplies no invalidate finding. Index rebuilding and collision rejection alone add no curation finding. Generalizing lessons and synthesizing articles do not establish a novel claim absent from inputs, so no synthesize finding is asserted from those labels: actual semantic relation remains unresolved. See MEM-OBJ-session-state, MEM-RTE-feedback-candidates, MEM-OBJ-session-digest, RT-RTE-session-bridge, MEM-RTE-refresh-sources, MEM-RTE-extract-lessons and EPI-OBJ-wiki-articles.
Reconciliation supersedes the epistemic lesson-extraction and source-refresh route identities with MEM-RTE-extract-lessons and MEM-RTE-refresh-sources. Citations here use those canonical routes. The surviving refresh delegation disagreement prevents settled context-boundary coverage but does not undo the shared explicit user-choice admission mechanism. Overlapping raw/article/feedback object containers remain valid distinct records; they are not additional independent witnesses. Refresh overwrite, conversation authorship and entry-specific query repair policies prevent universal immutability, raw-only reconstruction or whole-system read-only claims.
No frozen-source files were read for this classification. All warrants are the supplied accepted records and their annotations; no new evidence or records are declared.