llm-wiki memory profile

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

Comparison rationale

The inventory separates accumulated project content from unchanged shipped instructions and transient helper results. Markdown is a file substrate, not a graph substrate merely because reconstructed wikilinks form a graph. Optional Git initialization does not establish repository-based retained memory. The alternative index stays unresolved rather than inheriting BM25 classifications. See MEM-OBJ-page-metadata, MEM-RTE-metadata-read, RT-RTE-project-init and RT-RTE-qmd-search.

Prescribed semantic writing supplies claimed authorship, consolidation and evolution. Implemented template rendering supplies other-compiled lineage, not semantic synthesis. The epistemic container retains indeterminate semantic correspondence; its name cannot establish a novel claim. Initial provenance, generic caller identity and filing offers remain separate admission gaps. See RT-RTE-source-ingest, RT-RTE-page-create, RT-RTE-wiki-query and EPI-OBJ-wiki-pages.

Retained text can supply knowledge to an agent and ranking features to code without supplying enforcement or learning. Retained association fields can supply structural validation criteria and route pending notifications without authenticating source claims. Changed maintainer guidance supplies instruction authority only at its external consumer. These effects do not establish activation, benefit or a log-fed update. See MEM-BAP-wiki-content, RT-BAP-maintainer-instructions, EPI-RTE-source-coverage and RT-RTE-instruction-sync.

Requested ranked returns and explicit heading matches remain pull. They are not push merely because software selects and delivers text. Changed guidance loading remains unresolved, so read-back signal is not globally inapplicable. Recent-log delivery omits bodies and hyphenated operations; the retained full log still has a separate reload path. See RT-RTE-wiki-query, MEM-RTE-log-read and MEM-CLM-log-parser.

Reconciliation supersedes the epistemic raw, source-page and contradiction objects with the corresponding memory objects, and the epistemic catalog route with the memory metadata route. All classifications use those canonical declarations. The broader epistemic wiki-page container and combined log-inspection route remain valid rather than being split-superseded. No supersession changes a positive controlled value; the separate memory objects preserve local lineage and consumer distinctions.

No frozen-source files were read for this classification. Only the supplied accepted members and contracts were used. No records, quotations or evidence were added. Missing execution evidence limits observation and benefit, not the inspected helper wiring. Missing update/input records prevent a stronger trace-learning or trace-source assessment; an unknown is not converted into a negative.