Omni-Memory memory profile

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

Comparison rationale

Inventory and evidence strength

The comparison separates retained claim prose from access metadata, code topology, file projections and control state. It does not classify the external completion models as retained memory merely because they produce text. The graph's SQLite tables establish SQLite storage; graph organization alone is not a second storage backend. JSON columns likewise establish symbolic encoding, not a key-value service. MEM-OBJ-memory-text, MEM-OBJ-memory-state, MEM-OBJ-code-graph.

Every axis is partial because the accepted inventory retains unresolved alternatives. The records identify incomplete map/retained-flow consumer coverage, unknown upstream provenance and caller agency, and opaque model transformations. These limits differ by axis; the units name the prevented conclusions rather than treating every known branch as opaque. Parser fidelity alone would not erase the known graph substrate or form. The unresolved operative inventory, rather than a demand for perfect parser accuracy, prevents complete enumeration. MEM-OBJ-code-graph, RT-RTE-local-edits, MEM-OBJ-generated-documents.

All findings are implementation wiring or documented affordances. None is observed or causally supported. Classifying automatic admission does not authenticate a human caller, and classifying learning does not establish improved capacity. The source-native manual label is deliberately not mapped to manual write agency. RT-RTE-capture-admission, RT-RTE-local-edits, RT-RTE-shared-import.

Consumer-specific forces and selection

A retained sentence can be knowledge at the foreground model and a symbolic ranking input at the lexical scorer. Row status and scope enforce active-read membership; source/tier/peer state supplies local admissibility criteria. Those local gates do not enforce the host's coding actions or certify sentence truth. Citation-fed updates have learning force at the update path and ranking force at the later selector. EPI-RTE-query-ranking, MEM-OBJ-memory-state, EPI-RTE-merge-disposition, MEM-RTE-memory-withdrawal, RT-RTE-use-check.

Managed context's instruction finding concerns retained guidance delivered through the documented standing-context path, not the shipped skill or unchanging wrapper by itself. Generated API/linkup explanations have only an afforded interpretive consumer, kept separate from wired claim delivery. MEM-OBJ-agents-context, MEM-OBJ-generated-documents, RT-RTE-context-delivery.

Pull describes explicit requests for stored content. Session seeds, automatic prompt hooks and compaction-context additions also push selected memory. Coarse event/recency selection is separate from prompt-term lexical ranking and actual file/symbol/co-anchor identity matches. An ID printed beside a claim, a reference link to MEMORY.md, or a requested document filename alone establishes no identifier-based push. Internal command fulfillment is not an additional push at every step of that chain. RT-RTE-context-delivery, MEM-OBJ-event-log, MEM-RTE-recall-search, MEM-OBJ-code-graph.

Transformations and trace-fed writes

Heuristic selection of standing claims from session text warrants consolidation. Opaque model output does not warrant synthesis merely because it is generated; the accepted account cannot decide each output's semantic relation from the prompt. Duplicate retirement, edits, tier graduation/withdrawal and quarantine retain their own operation warrants. Hash freshness and citation do not establish successful outcomes. RT-RTE-capture-admission, MEM-RTE-transcript-window, RT-RTE-build-artifacts, MEM-RTE-anchor-maintenance, EPI-RTE-tier-graduation, EPI-RTE-tier-withdrawal, MEM-RTE-memory-withdrawal.

Automatic session extraction is a qualifying trace-fed write: conversation text produces durable claims that later retrieval supplies as context. Citation feedback changes durable priority/retention state, while activity summaries can orient a later session. These support local yes findings without training weights, new propositions or observed benefit. Session text remains session-logs even when a hook processes it; discrete capture/merge/source-use events support event-streams on their own route. Read-ledger hashes are not full tool-result content, and no linked state/action/outcome trajectory is established. MEM-RTE-transcript-window, RT-RTE-capture-admission, MEM-OBJ-event-log, MEM-RTE-anchor-maintenance, MEM-OBJ-observation-state.

Unresolved explicit-content provenance and additional flow/map consumers remain in both trace inventories. No accepted evidenced-absence record supports a local no. The profile therefore neither invents a negative unit nor makes unresolved sources inapplicable. The known yes witnesses do not establish complete learning/source coverage. RT-RTE-local-edits, MEM-OBJ-code-graph, MEM-OBJ-savings.

Reconciliation and source-read log

Reconciliation supersedes the workflow window-handoff identity with MEM-RTE-transcript-window; its model-prefix/cursor mismatch limits complete capture but does not remove the wired extraction route. It also supersedes the epistemic generated-artifact identity with MEM-OBJ-generated-documents; all document classifications use that canonical memory record. No other overlap is treated as replacement: the broader row container, metadata parts and narrower epistemic maintenance routes keep their own scopes.

No frozen-source files were read for this classification. All warrants come from the supplied accepted records and their annotations; no records, quotes or new evidence were added.