CQ memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies the supplied accepted records, not live operation. Wiring establishes capacity; neither host adherence nor benefit is observed. Static installed guidance is excluded because it is shipped configuration rather than memory accumulated through use. External embedding-provider parameters are excluded; the retained numerical outputs remain included. MEM-OBJ-knowledge-vectors and RT-RTE-host-installation delimit those surfaces.
The insight supersession resolves EPI-OBJ-insight to MEM-OBJ-insight. All classifications use that canonical identity. The combined EPI-OBJ-evidence, MEM-OBJ-selection-metadata and MEM-OBJ-flags remain overlapping containers, not synonyms. Optional prose criticism is supported locally, but its missing independent declaration prevents a complete part-addressed inventory. The profile preserves that limitation in form and authority rather than either inventing a record or dropping the prose finding. Store replacement and EPI-RTE-store-deletion likewise remain different scopes, with an afforded edit capability and wired physical deletion.
Advice is knowledge at the host, while the same text can be ranking input at the FTS/relevance consumer. Imperative wording does not make task advice standing instruction. Review status enforces query eligibility and validates a further review request; approval is not a tier promotion or evidence of truth. Confirmation raises salience and supported flags downweight it. Server/Python duplicate requests that fail construction establish no retained objection, deduplication or changed score. RT-RTE-knowledge-query, RT-RTE-remote-review, RT-RTE-knowledge-feedback.
Requested retrieval is pull even when code ranks or delivers its results. The Cursor stop hook supplies an identity-selected failure without a request for that memory. Resolver cache selection is another automatic supply operation, internal to an explicitly requested remote call: address identity selects retained NodeInfo for the HTTP client. This does not make the returned knowledge query another push. RT-RTE-failure-reminder, MEM-RTE-discovery-resolution.
The Cursor chain qualifies as trace learning in the comparison contract: a tool-failure event automatically produces bounded retained context for a later invocation. It need not generate a diagnosis or prove improvement. Its source content is tool feedback and its origin is a runtime callback event. The requested interactive session-to-lesson path has a distinct human content decision, and headless proposal/feedback records do not expose an actual trace-fed input chain. They therefore cannot be upgraded by the Cursor witness. Reflection's requested session/tool sources are not counted as positive sources of a qualifying automatic update. RT-OBJ-failure-record, RT-RTE-failure-reminder, RT-RTE-session-reflection, RT-RTE-knowledge-proposal.
No accepted bounded-absence record licenses a system-wide trace-learning negative or an exhaustive absence of additional curation. Index rebuilding, vector regeneration and identity collisions alone are not curation operations. Transfer imports existing prose and forgets the local copy, but does not consolidate feedback history or preserve global identity. Initial store/cache/vector provenance, arbitrary extension meaning, full alternative-store mutation parity and custom-store internals prevent stronger inventory assessments. RT-RTE-local-drain, MEM-RTE-maintenance-migration, MEM-OBJ-extensions, MEM-OBJ-knowledge-vectors.
Source-read log: none. Classification used only the supplied members and contracts; no source ambiguity required a direct frozen-source read.