Cortexes memory profile

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

Comparison rationale

The comparison separates canonical content, retrieval representations, retained control state and the short-lived continuation channel. The inventory is grounded in MEM-OBJ-raw, MEM-OBJ-notes-projects, MEM-OBJ-vectorstore, MEM-OBJ-distill-plan and MEM-OBJ-takeoff. Known coverage describes the recorded mechanisms, not observed adherence, semantic fidelity or deployed benefit.

Reconciliation supersedes the epistemic Raw and Notes object declarations with their memory counterparts and the epistemic index-withdrawal route with MEM-RTE-index-delete. This profile uses those canonical destinations. Summary generation, prefix checking/removal and evaluation subfunctions remain distinct parts rather than duplicate whole-route replacements. Index withdrawal changes access, not truth reliance; the early return can leave BM25 present. No invalidate value is inferred from it.

Manual admission is local to the explicitly prescribed human content decision. Starting takeoff does not approve its generated body. Consequently the continuation chain supports a claimed trace-learning yes, while human-confirmed distillation does not establish an automatic write. The baton has both contextual facts and standing next-step guidance. Its learning role at production and instruction/knowledge roles at the later consumer are separate (RT-RTE-takeoff-handoff, RT-BAP-baton-context). Unknown initial page provenance prevents turning every derived summary into a fully classified trace-learning route. No evidenced-absence record supplies a local no.

Requested retrieval can rank by lexical, embedding and model judgments without those becoming push signals. The pushed material is only startup topic metadata and repo-selected baton previews; full-page and full-baton loads remain requested returns (RT-RTE-session-start, RT-RTE-vault-search). A display summary does not license claims about the fidelity of the full payload.

Consolidation denotes implemented or prescribed reduction of retained content, not proven semantic faithfulness. Novel synthesis and factual acceptance cannot be inferred from fluent model output or conversational revision. Evaluation gold has criteria authority at the metric consumer, not factual authority over page contents or a wired production updater (EPI-RTE-eval-score).

Source-read log: none. This mapping used only the supplied accepted records and reconciliation; no source ambiguity required a direct source read.