Curated Thoughts memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The profile classifies accepted records, not fresh source evidence. All positive findings use implementation wiring, not observed operation or causal benefit. No source files were read for this mapping. Reconciliation supplied no supersessions: broader runtime routes and narrower epistemic parts remain valid at their own scopes.
The retained-content inventory is wider than the fact prose alone. MEM-OBJ-source-material, MEM-OBJ-wiki-content, MEM-OBJ-embedding-blobs and MEM-OBJ-verification-ledgers establish separate prose, numerical and symbolic parts. Graph-shaped SQLite rows do not establish graph storage, and embedding blobs do not establish a separate vector store. Provider parameters and shipped static prompts are not substituted for use-accumulated memory.
Authority follows the actual consumer. Fact prose is knowledge at the caller and librarian; vectors and scored prose are ranking inputs. Ledger keys and lifecycle metadata have symbolic eligibility and conformance effects. The proposal watermark and queued IDs route later work. MEM-BAP-retained-fact, MEM-RTE-verification-state, MEM-RTE-wisdom-match and RT-RTE-librarian-synthesis support these distinctions. “Guidance” does not by itself turn a recalled fact into a standing instruction. No learning authority is inferred without the qualifying update consumer.
The proposal filter and live-anchor commit check are successive mechanisms, not interchangeable truth checks. EPI-RTE-evidence-filter can keep a nonempty proposal after unknown references disappear; EPI-RTE-anchor-check can still veto insertion. Manual review and folder-driven automatic approval remain distinct admission units in RT-RTE-proposal-admission. Neither generic CLI identity nor workflow initiation proves a per-content human decision.
Curation findings describe implemented transformations. Normalized-body duplicate rejection in EPI-RTE-fact-retention is deduplication; direct replacement is evolution; soft withdrawal is invalidation; delayed physical forgetting is decay; human-tier confirmation is promotion. The model's semantic relation to its inputs remains indeterminate in EPI-OBJ-proposal-items. Thus the name “synthesis” does not establish new-claim synthesis, and requested summaries alone do not establish implemented consolidation. EPI-RTE-supersession-pending also prevents treating a retained correction request as an applied revision.
External requests are pull. Internal candidate supply on document processing and unsolicited desktop review prompts are push, each with its own selector. MEM-RTE-verification-prompt establishes an actual queued-ID match to current retained content, unlike an identifier merely printed in a requested result. Candidate name/vector targeting and coarse caps remain separate findings on RT-RTE-librarian-synthesis. The opaque external-engine route in RT-RTE-wiki-maintenance is not made known by these local witnesses.
Trace learning remains not determinable, not absent. RT-CLM-agent-workflow requests post-task findings, but MEM-OBJ-source-material explicitly leaves their experience, copied-material or model origin unresolved. The recorded queue event in MEM-OBJ-unverified-match is not itself a durable derived lesson. No accepted bounded absence record permits a local “no,” and no established original trace input permits a supported “yes” or trace-source value. Provider and external-engine opacity remain local limitations rather than negative learning conclusions.
The unresolved import, engine, identity/draft, replay and replica paths named in MEM-OBJ-wiki-content and RT-RTE-wiki-maintenance prevent complete inventories. Supported findings survive those gaps, but no axis with classified findings claims complete system coverage. No retained reasons, confidence stamp or successful return is treated as proof of factual correctness, actual later use or improved capacity.