Harness² memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The operative memory is not everything the runtime stores. The comparison covers accumulated behavioral artifacts, execution/history evidence received by later invocations, access metadata and recovery records. Unchanged baseline instructions are excluded until admitted as retained exemplars. External model components do not by themselves establish model-weights memory storage or parametric learning. The opaque native-memory branches remain included uncertainties rather than being erased by the inspected filesystem path. See MEM-RTE-evidence-frame, MEM-CLM-compaction-boundary and RT-RTE-resume-reuse.
No supersessions affect these values: reconciliation retains the declarations as distinct referents. EXPERIENCE.md, the reflected payload and INDEX.md overlap without being interchangeable. The evidence-tree container likewise does not substitute for every opaque raw asset or consumed projection. The profile therefore cites both object and route warrants where their scopes differ. See MEM-OBJ-experience, EPI-OBJ-experience, MEM-OBJ-library-index and MEM-OBJ-evidence-tree.
The library reaches improvers as evidence, not automatically as installed solver instructions. At the update consumer, retained execution evidence can have learning authority; at the reference consumer it supplies knowledge; only a subsequently admitted and installed harness grants instruction or executable force to solvers. These paths establish wired trace learning without establishing improved capacity, exercised self-improvement, actual rationale use or lesson truth. The fixed demo is kept separate from ordinary model-authored reflection. See RT-RTE-experience-admission, RT-BAP-harness-consumption and MEM-CLM-demo-experience.
Curation values follow implemented transformations. Clipping reduces historical evidence, edits revise existing harnesses, and prefix deletion forgets entries. Lesson generation requests transferable explanations, but no generated instance establishes whether those explanations merely compress prior claims or add one. The parser can also discard qualifiers. No synthesis finding is derived from that request or defect. See MEM-RTE-evidence-frame, MEM-RTE-library-withdrawal and EPI-RTE-lesson-parse.
Automatic coarse supply and subsequent requested browsing are separate read-back operations. Exact fingerprints identify requested cache returns, not an unsolicited targeted push. Trial grade-availability ranking is independently supported, but it is neither higher-score preference nor an expected-answer oracle for lesson acceptance. Retained scoring records supply report data, not a wired adaptation update. See MEM-OBJ-call-fingerprint, MEM-RTE-evidence-frame, RT-RTE-solver-execution and EPI-RTE-score-report.
Every axis is partial because supported framework findings coexist with included opaque alternatives. Missing external payloads, write control, original inputs and delivery policies prevent complete classification. No bounded absence record licenses a trace-learning no. All positive bases are wired implementation, not observed operation or causal support.
Source-read log: none. This classification uses only the supplied accepted members and reconciliation; no frozen-source ambiguity required an additional read.