auto-harness memory profile

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

Comparison rationale

The memory boundary is the retained experiment loop, not every state holder in the runtime. Local diagnostic files, evaluation controls and candidate revisions have different consumers. Ordinary task messages and BIRD in-memory sessions are working state, not demonstrated durable derived memory. BIRD diagnostic previews do not establish a model-window checkpoint; external model parameters and coding-host compaction therefore add no substrate, form or trace-learning value here. These distinctions follow MEM-OBJ-traces, RT-RTE-bird-session, RT-RTE-tau-task and RT-CMP-coding-model.

The accepted suite inventory separates MEM-OBJ-suite-tasks, MEM-OBJ-suite-threshold and MEM-OBJ-suite-last-results. Promotion changes future selection only after successful checks; diagnostics can persist before rejection; the threshold is carried forward rather than learned. This prevents assigning one authority or admission history to the whole suite. The score recorder similarly admits caller numbers without authenticating a gate result, leaving a separate content-control uncertainty despite implemented physical append; see RT-RTE-iteration-record.

Suite and split membership are automatic trace-fed durable behavior-shaping results. Their source is execution feedback, classified as tool traces, rather than natural-language generalizations. Their downstream authority is routing; feedback has learning authority at the update consumer. This is a memory-route classification, not a claim of improved future capacity or exercised self-improvement. Those conclusions remain independently limited in RT-RTE-candidate-revision. No event-stream or trajectory source is inferred merely from JSON event labels or a field name; MEM-OBJ-traces leaves external payload fidelity unresolved.

Candidate adoption contains two operations. Imports request Python and are pull. Subsequent context construction supplies the adopted candidate's instruction constants without a model request and is coarse push. The memory claim applies only when those constants have changed through use. It does not count unchanged shipped instructions as memory. The import/delivery channel is wired, while experience-responsive revision remains afforded/requested and actual replacement timing is qualified by caching; see RT-BAP-candidate-runtime, RT-OBJ-agent-instructions and RT-RTE-candidate-revision.

Reconciliation supersedes the epistemic learning-log declaration with MEM-OBJ-learnings. All profile citations use that canonical identity. The supersession resolves identity, not a missing reader or explanatory warrant. The valid suite and task-content containers remain alongside their parts; none of those parts supersedes a container. No other supersession changes a value.

Partial assessments preserve specific unknowns: external trace construction, supplied-score/configuration provenance and control, absent concrete semantic note transformations, arbitrary revised-code effects, and learning-log/historical-diagnostic later consumers. A known local path does not erase these alternatives. No observed-run or causal basis is used. No source files were read for this profile; all classifications use the supplied accepted records, so there is no additional source-read log.