JAZ memory profile

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

Comparison rationale

The comparison concerns retained use-derived material, not every object in the runtime. The local witnesses do not settle arbitrary live Python payloads, provider state or remote trace alternatives. All ten axes therefore retain partial inventory coverage rather than converting implementation witnesses into complete-value claims. Unresolved remote trace parts are carried as coverage limits: the accepted records do not supply a canonical remote-payload declaration from which to classify them.

The console supplies both an automatic transcript/history path and a later requested object-access path. The main-history display cannot classify the form or provenance of its opaque live values. The helper instead receives rendered transcripts, including rejected proposals and permission state; these are context, not validated knowledge. See MEM-OBJ-conversation-history, MEM-OBJ-helper-history and RT-RTE-console-turn.

Replay is not semantic retrieval or model training. A requested trace load is followed by automatic sequential saved-response selection. Saved source supplies executable directions; saved comparison expectations supply validation criteria; the divergence latch routes continuation and vetoes further saved service. Queue exhaustion is not forgetting the trace. See RT-RTE-trajectory-replay, EPI-RTE-replay-comparison and MEM-OBJ-atif-trajectory.

Generated workflow code is a trace-fed durable behavior-shaping artifact with an afforded later Python consumer. Its positive trace-learning finding does not require new claims, autonomous installation, observed execution or improved capacity. The training sample route likewise retains an afforded update consumer, not a wired end-to-end trainer. No accepted evidenced-absence record supports a local trace-learning negative. The unresolved update inventory prevents complete classification independently of these positive witnesses. See RT-RTE-workflow-emission, MEM-RTE-rollout-export and MEM-RTE-weight-update.

The reconciliation supersedes the epistemic rollout object and export-route identities with the memory declarations. All classification citations use the canonical MEM-OBJ-rollout and MEM-RTE-rollout-export. Their structural eligibility checks do not certify token provenance, reward, training success or accepted knowledge. Neither supersession connects export to the independent weight-update transport.

Readline retention caps support local decay; settings/credential edits support revision; replay divergence supports withdrawal of reliance while preserving the trace. These do not turn routine event projection into semantic consolidation or synthesis. Current-run observation abbreviation and advisory context warnings are outside the retained-memory curation findings. See MEM-OBJ-readline-history, MEM-OBJ-credentials, RT-BAP-settings-load and RT-RTE-context-control.

Source-read log: no source files were read. Classification uses only the supplied accepted members and their records; no evidence or records were added. Implementation and documented affordances establish neither observed activation nor causal benefit.