LHTB memory profile

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

Comparison rationale

The two positive trace-fed chains are context handoff to later foreground solving and persisted trial interpretation to later aggregation. Their classification follows automatic derivation, retention and later consumption, not proof of better performance. The update consumer gives prior experience learning authority; the resulting text supplies downstream knowledge and, for the continuation handoff, instructions. These are separate effects: RT-RTE-context-handoff, MEM-OBJ-handoff-context, MEM-RTE-trial-analysis and MEM-BAP-analysis-guidance.

Successful three-call handoff installs replacement history. Short and ultimate fallback prompts instead pass through unwound Chat; the profile does not pretend that every recovery installs the same frame. Summary generation and aggregation support consolidation without establishing faithful preservation. The indeterminate epistemic content relation does not negate implemented compression, nor license synthesis of a new claim: RT-RTE-context-handoff and EPI-RTE-job-aggregation.

Requested file access remains pull, while independent host selection and supply are push. Recovery's saved/planned configuration identity match and sweep task identity have actual selection consequences. A requested pathname, a trial name printed in an aggregate, or an upload duplicate-ID veto alone does not establish identifier-based push. The sweep implementation is retained at its wired layer despite unresolved singular reward-key compatibility: RT-RTE-job-recovery, MEM-RTE-sweep-selection, MEM-RTE-trial-analysis and MEM-RTE-artifact-access.

The Responses chain is a service-object at the exposed API boundary, not a known internal substrate or form. Readable trajectories do not classify it. Analysis SDK context and alternate-agent memory remain separate opaque units: RT-CMP-terminus-model, MEM-CMP-analysis-model and RT-RTE-agent-dispatch.

Reconciliation supersedes the duplicate epistemic analysis-model component with MEM-CMP-analysis-model. That component supports the retained SDK uncertainty, not a positive model-weights memory finding. No other supersession affects these values. The broad memory analysis route and its narrower epistemic assessment/cache/aggregate records remain compatible; retained trial/job files are not equated with the broader object that also includes task-quality checking: MEM-OBJ-analysis, EPI-OBJ-analysis, EPI-RTE-analysis-cache and EPI-RTE-task-check.

All axes are partial because their included inventories have unresolved alternatives, not because wired evidence requires observed execution. Raw checkpoint retention has no established later JSONL reader; live reward control is not a reload of that file. Optional task-check output has no established later consumer. Neither unknown licenses an absent route or a trace-learning negative. Maintenance adoption is separately afforded and its actor is not assumed human: MEM-RTE-checkpoint-retention, EPI-RTE-task-check and RT-RTE-runtime-patching.

Source-read log: none. Classification used only the supplied accepted records and reconciliation; no source inspection, execution, new evidence or record declaration was performed.