LLM-Wiki-v3 memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The inventory separates retained content from access derivatives and process reports. SQLite, graph and vector findings describe overlapping concrete storage organizations, not three copies of the source prose. Dense page representations are parametric form without being trained model weights (MEM-OBJ-vectors). Optional Git history remains an affordance, not observed preservation or rollback (RT-RTE-page-review).
Admission is classified at its decision mechanism. Human approval in interactive review supports manual admission. Model staging, checked seed execution and bootstrap writes support automatic admission despite operator initiation. Generic callable acceptance and unknown active-dataset initialization retain separate unresolved units (RT-RTE-source-ingest, RT-RTE-page-review, RT-RTE-curated-seed, MEM-OBJ-eval-queries).
Curation separates an extraction request from demonstrated semantics. Selective source-faithful reduction is requested, but no generated output establishes semantic fidelity or new-claim synthesis. Same-name replacement and current-loader exclusion have inspected implementations. Neither first-ID suppression nor rebuilding an index alone warrants semantic deduplication or consolidation (EPI-RTE-page-extraction, EPI-RTE-proposal-selection, RT-RTE-page-review, MEM-OBJ-page-routing).
Read-back includes both pull and push operations. Bootstrap's operator request for generated queries is upstream of code's automatic enumeration and delivery of retained bodies to fresh model invocations. The model does not request those bodies. That input delivery is push, with coarse all-page selection from registered namespace files, including superseded pages; it is not an identifier match or an answering-agent integration. The requested generated-query return is not another push. Query lexical and embedding selectors instead fulfill retrieval requests and are not push signals. Query returns identifiers and reduced source provenance, not bodies or an answering prompt. The separate baseline evaluator reads current corpus text without persisted indexes (RT-RTE-index-query, RT-RTE-query-bootstrap, MEM-OBJ-page-body, RT-RTE-eval-report).
No trace-learning negative is manufactured from missing evidence. The accepted members declare no evidenced-absence records. Audit/report retention does not supply an automatic trace-fed behavioral update, while document extraction and page-derived questions do not establish trace origins. Thus both trace axes retain no controlled finding, rather than asserting system-wide no or inapplicability (MEM-OBJ-audit-log, MEM-OBJ-eval-results, MEM-OBJ-eval-queries).
The supplied reconciliation supersedes the epistemic page-body and query-label objects with the memory declarations. This profile cites their canonical destinations: MEM-OBJ-page-body and MEM-OBJ-eval-queries. The valid combined raw/metadata containers are not treated as replacements for their distinct parts. No supersession upgrades a basis or resolves unknown dataset initialization, raw-byte replay, encoder provenance or actual adoption.
Source-read log: none. Classification uses the supplied accepted records only; no target source or incumbent profile was read. Wired findings establish implemented mechanisms, not observed activation or benefit.