MemexOS memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This comparison classifies accepted records, not live operation. All positive implementation findings retain the wired basis; none establishes compliance, improved answers or causal benefit. The manual snapshot procedure has only an afforded basis, separate from unattended admission (MEM-RTE-session-distillation).
The sole identity supersession affecting this profile is the epistemic wiki object to the memory wiki object. Accordingly the content classifications cite MEM-OBJ-wiki, not the superseded declaration. Catalog and review containers overlap other objects without identical scope, so they are not counted as independent extra stores merely because two analysts describe them. Evaluation parts keep their separate consumers rather than inheriting a blanket production-admission effect from their container (EPI-OBJ-questions, EPI-OBJ-rank-results, EPI-OBJ-tool-metrics).
The qmd index contributes vector storage and parametric representation, not model-weight training. Its readable snippets do not resolve internal vector schema or per-vector provenance (MEM-OBJ-vectors). Scalar timing state instead has symbolic form and routing authority; its automatic trace-fed write has a later budget consumer, so it qualifies for trace learning without claiming learned neural weights or semantic improvement (MEM-RTE-rerank-budget).
Session distillation is consolidation even when raw messages remain available. General writer compilation is left semantically unresolved: requested synthesis and a general ability to rewrite do not establish a new claim absent from actual inputs (MEM-RTE-session-distillation, EPI-RTE-writer-transform). Catalog synchronization establishes revision of effective metadata, not an independent semantic truth check (EPI-RTE-catalog-sync). Search hit suppression, FTS rebuilds and unused-name checks do not establish retained-memory deduplication.
Read-back is split within chains. Search/get fulfills a caller request and is pull, despite automatic ranking. Ingest, classifier precedent supply and fix/lint select retained inputs for fresh workers and are push. Exact URL/author matching establishes identifier selection only for the precedent branch; availability/status/budget selection stays coarse (RT-RTE-search-get, RT-RTE-compile-publish, MEM-RTE-catalog-reuse, MEM-RTE-usage-repair).
Trace learning means the automatic trace-fed write plus durable behavior-shaping output and later consumer required by this member's contract. It does not upgrade the reports' independent learning-as-improved-capacity conclusion. Session and usage routes yield a supported system existence finding of yes. No system-wide no or complete-value conclusion follows; unresolved media, research, taxonomy and evaluator provenance remain included. Session logs and tool outcomes are original inputs; serialization as events does not establish event-stream origin (MEM-RTE-session-distillation, MEM-RTE-usage-repair).
Retained review verdicts can validate and veto enabled publication while their prose directs a later repair writer. Published page content supplies knowledge to asking agents. Neither path implies uniform enforcement of truth, and the gate-off/no-change and metadata-sync paths remain exceptions (MEM-OBJ-review, EPI-RTE-review-decision, RT-BAP-retrieved-pages, EPI-RTE-catalog-sync).
Every axis is partial because the accepted inventory includes incompletely inspected alternatives. These limits are not negative findings. No evidenced-absence record supports a local trace-learning no. Caller identity alone leaves manual admission unresolved for category, capture and inbox decision APIs. Static initialization does not resolve subsequent human control (MEM-OBJ-taxonomy, MEM-OBJ-raw, RT-RTE-inbox-maintenance).
Source-read log: none. No named record ambiguity required direct source inspection; no new evidence, record, source or report correction is introduced.