Eva-brain memory profile

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

Comparison rationale

This profile classifies accepted records, not a fresh source inspection. All positive bases concern implementation wiring or model curation affordances; none asserts observed activation, factual correctness or causal benefit.

The comparison inventory is the retained brain and its later consumers, not the whole runtime. Files and Git history are storage; parsed adjacency is an access structure rebuilt from retained metadata, not a separate accumulated graph memory. Natural-language page content and symbolic matching/dispatch parts remain distinct (MEM-OBJ-page-metadata, MEM-OBJ-log, RT-RTE-brain-settings). Unknown initial provenance does not erase known later import, authoring, formatting or admission. The lineage finding for compiled output is limited to Rust-formatted saved analyses and their constructed reference metadata. Persisted identifiers, frontmatter and links elsewhere remain separately unresolved: deterministic parsing or consumption does not establish their derivation. The recorded page-level import and authoring routes do not resolve every metadata part's origin (MEM-OBJ-page-metadata, EPI-RTE-analysis-format).

Curation has weaker evidence than persistence. The model is asked to digest, revise and merge instead of duplicate; the code makes such edits retainable. These establish affordances, not a verified semantic transformation. Neither a named synthesis product nor the research profile proves that a new claim was created. Answer formatting is independently wired but does not supply a controlled semantic curation operation (MEM-OBJ-knowledge-pages, MEM-OBJ-summaries, EPI-RTE-analysis-format).

Wiki retrieval is pull even when deterministic lexical scoring selects the requested result. The same text can have ranking authority without an inferred-lexical push signal. Dynamic profile/configuration supply on task startup is a different operation: retained settings are selected automatically for dispatch and guidance. This coarse push does not turn a requested page read into identifier-based push (MEM-RTE-mcp-navigation, RT-RTE-source-ingest, RT-RTE-brain-settings). Static prompt construction alone is excluded from this classification.

The automatic operation log has a narrow trace-learning mapping. Its later consumer uses filename occurrence to alter future admission, including after rejection. Thus the retained result is more than an unused activity log. The wired yes is confined to this behavior-shaping state route; it does not contradict the records' refusal to establish semantic learning or improved capacity. Ordinary source-derived knowledge is not automatically called trace learning, and manually reviewed answer retention does not become automatic merely because Rust performs the write (MEM-OBJ-log, RT-RTE-source-ingest, RT-RTE-analysis-save). Arbitrary user documents and agent-summary origin leave the learning/source inventories partial. No bounded absence record supports a local no, and no system-wide negative is inferred.

Reconciliation supersedes the epistemic raw-source declaration with MEM-OBJ-raw and its duplicate navigation declaration with MEM-RTE-mcp-navigation. All findings use those canonical records. The corpus container and its prose/metadata parts remain distinct valid records; formatting remains a part of saving rather than a second admission route. No supersession changes a controlled value or upgrades its basis.

Additional source reads: none. The accepted records supply all warrants; no new quotes, records or evidence were added.