SAGE memory profile

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

Comparison rationale

This profile classifies accepted records, not new source facts. No frozen-source reads were needed or made. Every axis remains partial because the reports explicitly retain unresolved material alternatives; a strong local witness does not complete their inventory. No operation or causal benefit is claimed from source wiring.

Reconciliation supersedes the epistemic content, reflection, correction-replacement and lifecycle-disposition declarations with MEM-OBJ-memory-record, MEM-RTE-reflection-storage, MEM-RTE-correction-cycle and MEM-RTE-lifecycle-change. Those canonical records carry the affected content, lineage, revision and reliance findings. Reflection's initial fact typing and confidence assignment do not establish promotion of an already retained entry; that unit therefore retains no controlled curation finding. Gate evidence, model verdict and human review remain distinct valid parts; MEM-OBJ-review-decision supplies manual admission and conditional validation/enforcement, not learning or model-facing instructions.

SQL relationship tables are not a separate graph substrate merely because the tool calls them a knowledge graph. Dense retained vectors are a vector access substrate and parametric form; symbolic feedback accumulators are not model weights. The daemon PostgreSQL path does not establish full consumer parity. MEM-OBJ-memory-relationships, MEM-OBJ-memory-indexes, MEM-OBJ-validator-stats.

Trace learning and useful learning differ. MEM-RTE-verdict-feedback closes automatic experience-to-statistic-to-decision wiring. RT-RTE-bridge-distillation closes outcome-fed durable advice and later context routes while leaving original trace-source categories and actual summary semantics unresolved. Its requests for distillation and novel insights establish neither consolidation nor synthesis as a curation operation, even with a claimed basis. The bridge curation unit therefore remains not-determinable with no positive findings; this does not withdraw its separately established trace-fed write and later-context route. Capture retains selected trace text, not an automatically generated recommendation or continuation summary; the missing bounded-absence declaration prevents a controlled local no without erasing the qualifying positive routes. MEM-OBJ-capture-unit.

Read-back distinguishes an explicit retrieval request from automatic downstream model-input assembly. Topic embedding and lexical recall are not automatically push signals. The hook's exact thread match is identifier-based push, whereas domain/recency prefetch is coarse. Bridge and benchmark prompt assembly have bounded positive selection evidence but do not inherit complete selectors from unrelated recall paths. RT-RTE-targeted-recall, RT-RTE-hook-context, MEM-RTE-evaluation-readback.

WF-RTE-local-exchange resolves SQLite retention, prose/symbolic form, automatic protocol admission, enforced session ownership and requested later delivery for one coordination branch. It does not complete the task-board, legacy provider, pipeline-journal or federation inventory left open by MEM-OBJ-memory-record and MEM-RTE-sync-copy. These limits prevent complete-value statistics, not existence findings.