OKF Harness memory profile

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

Comparison rationale

The compared memory is retained workspace material, not the enclosing agent's whole storage or prompt. Local Markdown and JSONL map to files; Git checkpoint objects map to repo. Parsed graph edges do not imply a graph storage engine. Opaque source bytes prevent a complete form classification without making their known filesystem substrate opaque. See MEM-OBJ-source-bytes, MEM-OBJ-backlinks and MEM-OBJ-checkpoint-tree.

Requested retrieval is pull even when code chooses excerpts automatically. Ingest metadata planning has a wired pull finding only for its executable metadata consumption and recommendation return. Source/wiki reading assigned to the agent by ingest and reconciliation guidance has a separate claimed pull finding; neither metadata planning nor pair-validation code establishes implementation of that agent behavior. See RT-RTE-source-ingest and RT-RTE-source-reconcile. Identifiers on a requested read and lexical matching in an evidence request therefore establish no push signal. The unresolved semantic consumer of ledger notes is kept local rather than converted into either an absent route or a positive authority. See RT-RTE-evidence-answer, RT-RTE-direct-retrieval and MEM-OBJ-reconciliation-notes.

Lexically scored retained text has ranking authority as well as knowledge authority at the answering agent. Source/citation fields also participate in planner-local enforcement. Exact ledger pairs instead route bookkeeping: clearing currency is not validating revised prose or unsealing evidence. See MEM-OBJ-wiki-metadata, RT-CLM-evidence-boundary and MEM-CLM-currency-evidence. No learning authority is inferred from retaining a rationale without an update consumer.

Agent-authored semantic changes retain their claimed evidence layer; executable registration and retrieval remain wired. In particular, an instruction to synthesize wiki prose does not decide whether a claim was consolidated or newly created. Requested revision/withdrawal supports claimed evolution/invalidation, not demonstrated semantic correctness. The explicit human yes for conversation write-back establishes manual admission, independently of model authorship and later knowledge use. It is not truth certification or automatic trace learning. See RT-RTE-source-ingest, RT-RTE-source-reconcile, RT-RTE-evidence-answer and EPI-OBJ-wiki-page.

Partial axes retain their unresolved included alternatives: initial provenance, arbitrary source encodings, generic caller admission and note consumption. No bounded absence record supports a controlled trace-learning no finding; non-qualifying local mechanisms do not make all trace-source units inapplicable. External host compaction and model state are outside this memory boundary, not negative examples. No observed or causal finding is available from these records.

Reconciliation supersedes the epistemic workspace-check and graph-report routes with MEM-RTE-workspace-check and MEM-RTE-graph-report. All profile citations use those canonical identities. The supersessions preserve predicate-limited checks and parsed-connectivity derivation; they add no semantic acceptance, synthesis or trace-learning finding. Mixed source/page/ledger containers remain valid; this profile uses their narrower operative parts where consumers differ.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no frozen-source ambiguity required a source read.