Meta-n memory profile

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

Comparison rationale

The profile classifies accepted records, not new source evidence. All positive findings are wired implementation witnesses; no finding claims observed operation, faithful compaction, actual helper adoption, causal benefit or improved capacity. Partial coverage is not weak wiring: every axis preserves the known routes and the included unresolved alternatives.

The revision's reconciliation supersedes the epistemic Python/bash objects and prior-summary object with their memory counterparts, and the epistemic summary replacement route with memory's compaction route. This profile cites the canonical memory declarations. The valid runtime helper container and emitted-context part remain distinct. Supersession does not resolve the epistemic uncertainty about summary fidelity or the semantics of unsupplied generated claims.

The learning force of retained experience at Ω is separate from instruction/knowledge force at a later solver. Similarly, stored scores rank candidates, checkpoint state routes continuation, and persisted bars validate and veto best-index admission. Executable availability alone does not establish a permit/refuse criterion or every generated program's authority. Verification stamps retain their conditional runner/JSON warrant, rather than being upgraded into authenticated execution evidence.

Requested archive views are pull. Automatic supplies to subsequent solving invocations are push: chain guidance and continuation summaries use coarse availability/position/budget selection, while native frozen dispatch actually matches a task identifier to retained source. A requested file's identifier alone supplies no identifier-push finding. The Ω trace-sampling predicate therefore is not listed as another push merely because code returns a formatted view.

The summary's retention horizon is later invocations within one task, whereas intervention chains and maps survive trials and disk resume. Both support automatic trace-fed writes into behavior-shaping artifacts. Manual seed supply classifies the operator's explicit content-import decision, not human authorship, per-function approval, or a negative about its unknown upstream learning history. Tool execution feedback and recorded solving turns supply the supported trace origins; telemetry pointers do not reveal an opaque transcript's source or form.

Consolidation follows compression/selection of retained content, not the route name. Same-name overrides and repairs evolve a new candidate's organization; cleared verification status invalidates current reliance while history remains. Indexed winner replacement promotes selection salience. No actual generated claim establishes synthesis, and budget eviction from a request is not durable decay. The unresolved diagnostic, SDK and cache routes remain included instead of being dropped to obtain complete coverage.

Source-read log: none. No target source was read for this classification; all warrants come from supplied accepted records and their annotations, with duplicate identities resolved through the supplied reconciliation.