Hippocampus memory profile

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

Comparison rationale

This profile classifies accepted records, not target execution. Semantic extraction, saving and editorial repair are afforded by model-facing instructions. Hook delivery and Git persistence are wired. No finding is upgraded to observed or causally supported. The vault is a file-based linked knowledge system: links alone do not establish graph storage, and external agent weights are not accumulated vault state.

Reconciliation supersedes EPI-OBJ-source, EPI-OBJ-pages and EPI-OBJ-hot with their memory-member counterparts. The classifications therefore cite MEM-OBJ-source, MEM-OBJ-pages and MEM-OBJ-hot. Archive records overlap without whole-record identity. The combined epistemic catalog/journal record is not substituted for the separate catalog and journal. Mechanical and editorial report parts retain distinct forms and bases: EPI-OBJ-mechanical-findings and EPI-OBJ-editorial-findings.

Conversation saving supplies a qualifying automatic trace-fed write despite its user trigger. The model selects durable conclusions, and later agents receive pages or their cache derivatives. This supports afforded trace learning and session-log origin, not measured improvement. Learning authority at the update consumer is separate from knowledge authority downstream. Arbitrary source payloads remain unresolved rather than receiving unsupported negative learning values; no accepted evidenced-absence record supplies such a negative. See RT-RTE-conversation-save and MEM-OBJ-pages.

Consolidation follows the requested selection/compression mechanism. It does not certify fidelity. The source word synthesis does not establish new claims absent from inputs; semantic novelty remains unresolved. Approved merging and deletion afford deduplication and forgetting, but mere duplicate-hash rejection, age flags and contradiction callouts do not establish those transformations by themselves. See RT-RTE-source-ingest and RT-RTE-vault-lint.

The automatic cache branch is coarse push: an event and file-existence check supply the whole hot file. Requested catalog and lexical searches are pull selection, not extra push signals. Missing host truncation semantics prevent an exact deployed context-size claim, not classification of the wired selector. See RT-RTE-lifecycle-hooks and RT-RTE-vault-query.

Write agency classifies content admission, not physical persistence. The stop hook's wired staging and commit remain storage findings only; they do not upgrade the afforded automatic content admission to wired. Optional diagnostic report admission is separately afforded: the model writes the report on user request or when findings are numerous, without a prescribed per-content approval step. A user request starts that production rather than approving its generated content. Mechanical findings are wired outputs, but their optional retention and the editorial report remain model-facing instructions. See RT-RTE-lifecycle-hooks, EPI-OBJ-findings, EPI-OBJ-mechanical-findings and EPI-OBJ-editorial-findings.

Source-read log: none. This mapping used only the supplied accepted members and reconciliation; no frozen target files were reread.