Hindsight memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
Every positive finding is source-wired, not an observed run or demonstrated benefit. The ten axes remain partial because accepted records retain additional shipped paths and opaque alternatives. No known database, caller or consumer resolves another route's missing facts. Unknown units are not negative findings.
Parts, derivations and authority
Default SQL content and vector access structures coexist with unresolved configured stores. Dense embeddings receive parametric form without being classified as trained weights. Facts, observations and page prose are distinct from their access IDs, edit addresses and audit state. A readable snippet does not classify an opaque attachment or store payload. See MEM-OBJ-fact-embedding, MEM-OBJ-page-embedding and MEM-OBJ-document-text.
Authored page creation does not mean human admission. The caller may be software, while later model/code refresh admission is automatic. Source-conditioned transformations support compiled lineage without establishing new claims, faithful semantic preservation or truth. Consolidation and revision are supported; the semantic uncertainty prevents inferring synthesize from a route's name. Optional semantic-dedup handling remains incompletely inspected. See MEM-RTE-page-maintenance, MEM-RTE-observation-consolidation and EPI-RTE-page-refresh.
Learning authority belongs to retained experience/evidence at durable-artifact update consumers. Knowledge belongs to selected prose at the later answering consumer. Directive instruction is not semantic enforcement; the prompt's rejection wording is not a witnessed response checker. In contrast, retained source/target identities can supply fixed-rule validation and veto criteria. Bank overrides and refresh windows route operations without establishing improved capacity. See MEM-BAP-reflect-content, MEM-BAP-directive-prompt, MEM-OBJ-refresh-response and RT-BAP-config-consumer.
Chains and coverage
The hook requests recall upstream and independently pushes selected context downstream. Requested recall/page results are pull, not duplicate pushes. Automatic directive loading and maintenance baseline supply are separate operations. Push signals retain their local selector limits: an embedding/reranker witness does not classify every retrieval arm, and a requested ID alone does not establish identifier push. Mirror synchronization is inspected server-to-disk delivery; later agent reading is unresolved. See RT-RTE-claude-hooks, MEM-RTE-recall-selection, MEM-RTE-page-delivery and MEM-RTE-filesystem-mirror.
Transcript-to-fact-to-later-context wiring qualifies as trace learning under this profile's write-route definition. It does not require a measured benefit, learned weights or epistemic acceptance. Conversation and optional tool-call content establish distinct source findings. Hook scheduling does not add event-stream content. Current-request map/reduce history is excluded because the accepted account establishes no durable reload chain for it. See RT-RTE-claude-hooks and RT-RTE-context-synthesis.
Reconciliation and source-read log
Reconciliation supersedes the duplicate epistemic observation-batch and page-evaluation routes with MEM-RTE-observation-consolidation and MEM-RTE-page-evaluation. Those canonical destinations are used here; supersession does not upgrade their evidence or remove the semantic and evaluator-feedback limits. Valid containers and narrower disposition/withdrawal routes remain distinct, not silently replaced. The conditional refresh/reflection model alias does not prove identical weights or accumulated parameter learning.
No target source files were read for this classification. All warrants come from the supplied accepted records and their annotations, resolved through the supplied reconciliation. Missing alternative-store, initialization, integration-consumer and semantic-transformation facts prevent stronger assessments; they are not reopened report-verification findings.