academic-research-skills memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies accepted records, not new source evidence. All ten axes remain partial because their inventories include unresolved host or optional paths. A strong cache or file-writing witness does not resolve another admission mechanism, payload or later consumer. Findings with different bases remain separate even when their controlled value agrees.
The resolver cache is SQLite with symbolic, compiled decisions. Its retained result supplies knowledge to later verification; the original resolver results have learning authority at the automatic durable-update consumer. The trace learning value means this write-and-later-consumer route, not improved accuracy or parameter training. Its source is tool execution feedback, not an inferred session transcript. See MEM-OBJ-verification-cache and MEM-RTE-cache-verification.
Selected session words and tool outcomes also feed durable recovery state. The complete capture-to-recovery chain is prescribed rather than observed; the ledger append helper's wiring does not upgrade that chain. This supports a claimed trace-learning route without treating every raw log as learned guidance. The retained decisions supply directions, while receipts/counters route recovery. See RT-OBJ-run-ledger and RT-RTE-checkpoint-recovery, including their memory annotations. Host summary production remains opaque, not an evidenced negative or a hidden complete-value assertion.
Corpus imports supply data, not standing directions. Imported and compiled lineage coexist; starting an import does not make its automatic per-record admission manual. Explicit reading declarations, author inquiry adoption and revision authorizations instead have local human control. Criteria bindings validate identities and receipts, not substantive application of criterion meaning. See MEM-RTE-corpus-reuse, MEM-RTE-read-attestation, MEM-RTE-inquiry-summary, MEM-OBJ-review-criteria-binding and RT-RTE-revision-review.
Read-back direction is operation-local. Exact-key cache queries and explicit history rendering are pull. Automatic checkpoint inquiry display and session activation restoration are push. Inquiry's summary selector is inspected code, while its insertion into model context is an afforded orchestration route. A requested file identifier is not an identifier-push witness. See MEM-RTE-cache-verification, MEM-RTE-activity-history, MEM-RTE-inquiry-summary and RT-RTE-command-dispatch.
Curation findings follow transformations, not names. Read rescission invalidates reliance while retaining history; cache replacement evolves a decision and TTL expiry removes reuse eligibility. Activity deletion forgets selected history. The prescribed process narrative consolidates existing history. Index updates, idempotence checks and a request for new synthesis do not establish deduplication or novel-claim synthesis. See MEM-RTE-read-attestation, MEM-RTE-cache-verification, MEM-RTE-activity-history, MEM-OBJ-process-record and EPI-RTE-synthesis-content.
Reconciliation supersedes the duplicate epistemic reading-ledger object with MEM-OBJ-human-read-log. This profile cites that canonical object, retaining the distinct writer/resolution routes. No other supersession changes a value. Valid stage/final-output containers do not replace the more specific cache, corpus, reading or screening records.
Activity extraction is retained history but its contract excludes model, gate and checkpoint input. The process record has a human reader without an established automatic lesson-consumption route. Neither establishes a blanket negative trace-learning classification; no bounded absence record is declared for that purpose. Unknown initial-content control, generic cache caller provenance, complete criterion consumers and optional audit/secondary retention remain limitations, not silently excluded alternatives. See MEM-RTE-activity-history, MEM-OBJ-process-record, MEM-OBJ-review-criteria-binding, EPI-RTE-claim-judgment and RT-RTE-cross-verification.
Source-read log: none. Classification uses only supplied records and their annotations; no frozen-source files or incumbent/reference profiles were read.