Zikkaron memory profile

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

Comparison rationale

This profile classifies accepted records, not fresh source inspection. All positive findings are implementation wiring or explicitly bounded library affordances. No observed operation, successful host delivery, improvement or causal benefit is asserted. No source files were read for this mapping.

SQLite is the physical retained store; graph and vector name its operative access structures. RAM buffers, codebooks and reconstructed numerical patterns are included because they accumulate or change through use. Dense encodings map to parametric form without being confused with pretrained neural weights. Scalar confidence/heat and fixed-code relations are symbolic. See MEM-OBJ-selection-state, MEM-RTE-hierarchy-read, RT-CMP-hdc-encoder, RT-CMP-hopfield-memory and RT-CMP-cognitive-map.

Caller composition and import, trace extraction and subsequent compilation can coexist. None establishes human admission control. The known transformations do not settle initial belief population, exact embedding provenance or whether every supplied statement originated in activity traces. MEM-OBJ-derived-beliefs, MEM-OBJ-index-enrichment and MEM-RTE-sync-import retain those limits. Afforded branches are not upgraded by neighboring wired ones.

Update-consumer learning authority and downstream knowledge/ranking authority are separate. Action summaries, query-driven revisions, access/navigation updates and rating-derived controls meet the architectural trace-learning definition without demonstrating better performance. Compaction-triggered checkpoint changes form a narrow event-fed route; they are not captured conversation history. Mixed-origin remember content prevents complete trace-source coverage. See RT-RTE-trace-consolidation, MEM-RTE-recall-revision, EPI-RTE-feedback-update and RT-RTE-replay-context.

Curation labels follow transformations. Extraction/compression remains consolidation even when omitted qualifications damage fidelity. Actual synthetic assertions support synthesis; an archive action does not by itself support invalidation. Stale flags withdraw freshness reliance only on consumers that respect them. Protection criteria and their vetoes operate at particular mutation consumers, not as universal authorization. See MEM-RTE-compression-write, EPI-RTE-synthetic-facts, EPI-RTE-file-validation and MEM-RTE-withdrawal-path.

Requested service/project/hierarchy/story/evaluator returns are pull. Prompt/session/compact hooks, prospective matching and optional related-memory additions are independent unsolicited supply. Identifier matching is attributed only to the project-targeted portion, not globally selected checkpoints/anchors. Semantic service rerankers do not establish inferred-judgment push. Prospective grammar and reinjection details missing from canonical records remain local unknowns. See RT-RTE-hook-context, MEM-RTE-trigger-delivery and RT-RTE-remember-recall.

Reconciliation supersedes the epistemic curation, recall-rewrite, narrative, trigger and withdrawal duplicates with the memory records cited here. Their non-ampliative/indeterminate and authority limits still apply; no supersession upgrades evidence. Archived compression remains a narrower part of the two-branch compression account, and consistency checks remain distinct from promotion. These identities change citation representatives, not the controlled-value unions.

Benchmark preparation independently qualifies as trace learning: software reformulates recorded conversational turns into retained rows without per-memory human approval, and retrieval later supplies selected contents to per-question generation. This supports wired session-log origin, independently of judgment-to-production feedback. The workflow account and annotation specify these stages on RT-RTE-evaluation-mode. Operator initiation is not per-memory approval. Recorded input supplies learning authority at the update consumer; the resulting context supplies knowledge at generation.

The accepted records leave full graph/domain/community semantics, conflict resolution, some resource routes and external/library callers unresolved. Those limits are retained in axis/unit assessments rather than suppressed by strong positive witnesses. Benchmark judging supplies measurement/reporting, not an established automatic production-feedback route: see RT-RTE-evaluation-mode and EPI-RTE-answer-judgment. Complete-value statistics and system-wide negative trace-learning conclusions are not licensed by this partial profile.