Citadel memory profile

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

Comparison rationale

This profile classifies accepted records, not the whole runtime's storage. All ten inventories remain partial: known mechanisms coexist with named provider, capture, connector, repair or consumer uncertainty. A configuration branch's wired status does not establish deployed configuration or observed operation. No source files were read for this profile; supplied records contain the supporting passages.

The reconciled node-note object is MEM-OBJ-node-note, not the superseded epistemic object. The optional direct-learning route is MEM-RTE-source-enrichment, not the superseded epistemic enrichment route. Its generated output remains EPI-OBJ-enriched-chunk. Neither supersession enables enrichment on HTTP ingest. The extraction part EPI-RTE-note-distill remains distinct from its parent capture route and qualifies note fidelity without defeating consolidation or the retained trace-fed pathway.

Trace-learning yes applies to automatic production and admission of selected experience that later consumers can receive, not to training weights or demonstrated improvement. Its original inputs support session-logs and tool-traces. Already-supplied shared text and opaque Cognee successors are separate unresolved units. There is no evidenced-absence record supporting a local no; missing improvement evidence is not such a warrant. See RT-RTE-session-capture, MEM-RTE-share-session and MEM-OBJ-search-projections.

The task hook has a request-driven service step and an automatic host-context supply step. Lexical/vector relevance selects retained text; scope and caps are separate coarse controls. Explicit document expansion is pull, not identifier-based push. Maintenance's automatic supply of recent labels/tags also has a coarse selector. A requested identifier or a model's relevance verdict alone does not establish an additional push signal. See RT-RTE-task-search, MEM-RTE-document-drilldown and RT-RTE-index-curation.

SQLite source content, JSON vault history and provider projections are different objects. Tombstones invalidate current-search reliance while leaving history; the vault may retain a revision before search admission rejects it. Accordingly, enforcement and validation are bounded to actual selectors and gates, not universal erasure or semantic acceptance. See MEM-OBJ-source-revisions, MEM-RTE-vault-sync, MEM-BAP-current-head and EPI-RTE-automation-receipt.

Known automatic admissions do not resolve generic caller identity or queued approval into manual control. The approval instruction is policy; workflow details retain software clients as alternatives even where browser confirmation is offered. Optimization notes are additive outputs, not proven edits to the original source or proven synthesis of new claims. Provider improve remains opaque rather than receiving the strongest basis of the explicit note-write branch. See RT-RTE-promotion-review, RT-CLM-explicit-approval, MEM-OBJ-optimization-note and MEM-OBJ-search-projections.