OpenRSI memory profile

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

Comparison rationale

All positive findings classify accepted implementation records, not observed operation. The inventory is deliberately partial on every axis: the accepted memory report retains included alternate copies and consumers that have not been traced. A known witness does not resolve them. External numerical state has known SFT update wiring but unresolved initial provenance and historical deployment. The upstream RL optimizer remains distinct from the vendored SFT trainer. See RT-CMP-training-models, RT-RTE-sft-training and RT-RTE-rl-training.

Reconciliation supersedes the epistemic rich-summary object with MEM-OBJ-rich-summary, and its summary-generation route with MEM-RTE-summary-cache. Those are the canonical citations used here. This changes identity resolution, not the strength or semantic warrant of any finding. The targeted display route remains separate.

Memory advice is knowledge at the proposal consumer; its source experience can separately have learning authority at a durable-summary or numerical-update consumer. Neither entails measured capacity gains. Generated metric code instead acts as executable validation criteria; its rejection path supplies enforcement. The profile does not upgrade advisory experience into standing instruction or warranted explanatory knowledge. See RT-BAP-operator-guidance, MEM-RTE-summary-cache and EPI-RTE-package-validation.

Summary method overviews and boards reduce existing experience; the comparative explanation field can contain unknown semantic additions. Consolidation is locally supported, while synthesis of new claims remains unresolved. SQL pruning forgets rows; island eviction only removes active eligibility and can leave journal evidence. Metadata merging revises retained entries. These effects have different scopes, rather than a global forgetting or truth-maintenance guarantee. See MEM-OBJ-experience-card, MEM-OBJ-rl-program and MEM-RTE-sft-search.

Requested checkpoint/cache/dataset reads and automatic later prompt selection are different operations, so both pull and push are supported. Identity-targeted push requires actual relationship matching, not merely a node-ID cache filename. Error-signature matching is lexical selection; numeric rank and recency are coarse selection, not inferred model judgment. Evo shared-parent siblings and RL children-of-current-parent are not interchangeable. See MEM-RTE-targeted-memory and RT-RTE-rl-training.

The trace-source findings concern recorded model-work conversations, tool execution feedback and linked program-change/outcome sequences. Imported task files do not become event streams merely by entering a journal or pipeline. No local negative trace-learning finding is invented: records establish positive routes and unresolved alternatives, not a bounded all-route absence. See MEM-OBJ-alignment-dataset, MEM-OBJ-journal and RT-RTE-gym-build.

No frozen-source files were read for this mapping. Only the supplied members and governing contracts were read; no new records, quotations or evidence were added.