OKF memory profile

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

Comparison rationale

This profile classifies accepted records, not source files or an executed OKF session. Implementation findings use wired basis; requested agent semantics retain claimed basis. The external computation consumer is an afforded role, while its specified attestation/refusal rules remain claimed. No observed or causally supported basis is available. Evidence limits do not erase the known file-based paths.

The retained inventory is not the whole runtime's storage. Parsed folders/corpora are access structures over retained files; the optional backend remains opaque despite its SQLite-like package name. Git history appears only as the requested journal-reconstruction input. Fixed shipped skill and hook machinery do not become acquired memory merely because installation copies them. See MEM-OBJ-journal, RT-RTE-mcp-retrieve and RT-RTE-governance-install.

Canonical citations reflect all three reconciliation supersessions: concept prose uses MEM-OBJ-concept-body, provenance uses MEM-OBJ-concept-metadata, and collision feedback uses MEM-RTE-reconcile-search. The corresponding epistemic declarations are not cited as separate objects/routes. The computation contract remains separate from the broader reference-payload inventory, because those referents are not coextensive. Its symbolic fields cannot classify an opaque executable or numerical payload. See EPI-OBJ-computation and MEM-OBJ-references.

Knowledge, ranking, routing and validation concern distinct consumption paths. Searchable retained text can influence ranking without a learned retrieval policy. Trust/status filters route optional result membership; human-review metadata and board links also become criteria at the pairing checks. The board cap can veto promotion through a mechanical writer. These are retained-data effects, not a claim that every concept is an enforced instruction or semantically validated. See EPI-RTE-trust-tier, EPI-RTE-read-pairing, MEM-OBJ-board and RT-RTE-pro-writes.

Ordinary model admission without per-content operator approval is the automatic write-agency branch, although its operation remains unobserved. Human verification and refine approval are separately claimed manual decisions. A generic caller at Writer or the Snapshot append does not resolve that caller's control. Mechanical derivation and physical writing are not human approval; computing a Snapshot does not write it. See RT-RTE-skill-work, RT-RTE-refine-proposal, EPI-RTE-semantic-review and MEM-RTE-snapshot-context.

Curation findings retain their requested layer. Correction evolves existing content; deprecation invalidates reliance while keeping context. Distillation, closure extraction and proposal terminology alone cannot decide consolidation versus synthesis for a particular retained-input transformation. No new claim is inferred from requested cross-cutting explanation, and faithfulness is not established by structural conformance. Mechanical promotion moves board work, not necessarily a memory item to a higher salience tier. See EPI-RTE-semantic-disposition, MEM-OBJ-concept-body, RT-RTE-pro-writes and WF-RTE-maintenance-closeout.

Read-back separates request, selection and delivery. CLI/MCP query or id requests are pull; exact-id delivery does not become identifier push. SessionStart and eligible Write feedback supply retained material without requesting those parts. Recency/category selection is coarse; automatic filename-term collision search is inferred-lexical. Host attachment may supply model context without a model fetch decision, but the accepted records do not reveal the host's selector. Its attachment request to OKF remains pull at that interface. See RT-RTE-mcp-retrieve, MEM-RTE-snapshot-context, MEM-RTE-friction-feedback and MEM-RTE-reconcile-search.

Trace learning cannot be reduced to persistence, a Learning label or improved-capacity claims. Caller-appended counters and automatically logged raw intents have later consumers, but the records do not establish an automatic trace-fed write of a durable behavior-shaping derivative. Nor do they declare an evidenced absence spanning each route. Thus neither a local controlled no nor a system-wide no is warranted; source classification remains unresolved rather than inapplicable. The telemetry's event origin does not by itself make event-streams a qualifying learning source. See MEM-RTE-snapshot-context, MEM-RTE-friction-feedback, MEM-OBJ-friction and RT-RTE-refine-proposal.

The supported system unions are derived only from findings; opaque alternatives remain in their units. In particular, the known fallback does not settle optional-backend storage/form/lineage, a reference inventory does not settle payload semantics, and claimed semantic operations do not acquire wired strength from executable storage. These unresolved facts prevent complete-value coverage, not existence of the supported branches.

Source-read log: none. No source ambiguity required a permitted frozen-source reread; all classifications use the supplied accepted records and their annotations.