Ecdysis memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies the accepted records, not fresh source facts. No source files were read for this mapping. Evidence is implementation wiring unless a finding says otherwise; no inspected execution or causal comparison supports an observed benefit.
The memory boundary is narrower than all runtime storage. Static role prompts and provider parameters are not accumulated memory of the shipped improvement plane. The generic live harness is included even though its internal form and consumer force are unresolved. Its in-process substrate is known; this does not establish external persistence or deep immutability (RT-OBJ-frozen-harness, RT-BAP-retained-harness). Opaque artifact metadata and task/output payloads likewise remain visible in the unresolved units rather than being classified from JSON display alone (MEM-OBJ-skill-artifact, MEM-OBJ-inference-record).
Reconciliation supersedes the epistemic transcript, round-record and evolved-skill declarations with the memory declarations. All classifications cite the canonical MEM-OBJ-transcript, MEM-OBJ-round-record and MEM-OBJ-evolved-skill. This changes citation identity, not a value or evidence strength. The formatted failure display and proposed-change part remain distinct from extracted trajectories and the whole specification (EPI-OBJ-failure-context, EPI-OBJ-proposed-changes).
Read-back direction is operation-relative. A caller's checkpoint-path request and explicit artifact load are pull. Automatic history-tail supply to the next review role is push even within an operator-started workflow. The requested cached return is not counted twice as a second push. Automatic trainer supply of the retained spec/reference has only coarse workflow selection; path names, context/model compatibility and duplicate IDs do not establish a targeted identifier push (RT-RTE-checkpoint-recovery, RT-RTE-failure-review, RT-RTE-artifact-utilities).
Trace learning here means automatic trace-fed retained guidance with a later consumer, not proven improvement. Optional checkpointed review turns/specification provide that wired chain even without an observed run. The score-retained harness update also has a wired evidence/admission/consumer chain, while its real semantics and initial provenance remain unresolved. The synthetic editor's disregard of the spec prevents review-content attribution, not the existence of delivery and reference replacement (MEM-OBJ-checkpoint, RT-BAP-review-editor, RT-RTE-candidate-admission). Separately stored skills neither inherit that provenance nor acquire task-agent instruction force merely from their tip text (MEM-ABS-skill-injection).
Curation maps actual reductions, duplicate suppression and withdrawal of checkpoint reliance. It does not map requested conservative review to proven synthesis, first-ID merging to semantic revision, history-tail omission to forgetting, or candidate rejection to deletion of retained advice. Reasons can survive in transcript/specification without reaching the final harness or skill consumer (MEM-OBJ-transcript, RT-OBJ-review-spec, RT-RTE-artifact-utilities).
Every supported local finding contributes to the derived union without resolving other units. Missing caller-control facts prevent manual admission classification; missing opaque content/consumer facts prevent complete form, authority and transformation coverage. No local negative is promoted to a system-wide no, and no known pull route makes an unresolved push alternative inapplicable.