Meta-Harness memory profile

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

Comparison rationale

The profile classifies accepted records, not a new source inspection. No source files were read for this mapping. Evidence is implementation or shipped procedure: no finding claims observed activation, measured benefit or causal support.

The example-list duplicate resolves to MEM-OBJ-examples; the superseded epistemic declaration is not used as a separate witness. No other supersession changes a value. The valid evolution-summary container is represented through MEM-OBJ-summary-results and MEM-OBJ-summary-rationale. Software-derived counters/results do not confer derived explanatory warrant on copied model hypotheses. The broader EPI-OBJ-val-results adds numerical result-file coverage, not another rationale lineage.

Storage and form are separate. A file checkpoint does not make an opaque payload symbolic merely because its name ends in JSON. Few-shot Q/A strings supply knowledge through interpretation, while list fields, resume counters and scores have fixed-rule consumers. NoMemory and unknown generated state prevent treating successful reload as universal behavioral activation. MEM-OBJ-memory and MEM-RTE-classification-memory retain those alternatives.

Pull and push are operations within chains, not labels on entire workflows. A requested checkpoint or digest replay is pull; automatic demonstration selection into the solver request is push. Its input hash seeds sampling, rather than matching a retained example identity. KIRA context pressure triggers automatic continuation delivery, but the inherited summarizer's precise selection remains unknown. The visible fallback retains a screen suffix and instruction alongside surviving messages, not necessarily an entirely replaced context. See MEM-RTE-classification-memory, MEM-RTE-completion-replay and the memory annotation on RT-RTE-kira-execution.

Trace learning is not an improvement verdict. KIRA's retained continuation qualifies even within one task. Reference experience-to-code/report use remains afforded procedure; pilot feedback-to-edit wiring is not observed evidence-responsive revision. Ordinary pilot tool-message accumulation is not separately counted as derived memory. Historical scoring uses retained eligibility/reward fields, not the final candidate bytes, and does not restore best code. See RT-RTE-classification-search, RT-RTE-terminal-search, RT-RTE-pilot-search and MEM-OBJ-evaluation-history.

Fresh-start deletion is not classified as invalidation. The memory annotations on RT-RTE-classification-search and RT-RTE-terminal-search establish selective cleanup and residual artifacts, not retention of the history of the particular withdrawn contents. That missing relation prevents a controlled curation finding for reset; it does not erase the independently supported consolidation, evolution and promotion findings.

All ten axes remain partial because the accepted inventory contains generated-candidate alternatives, unspecified checkpoint initialization, opaque inherited/provider state or selection, or unproduced model guidance. Those are local coverage limits, not evidence of absent memory or negative learning. A strong local witness never upgrades a merely afforded branch. No unsupported global absence, complete-value union or empirical improvement follows.