Prime Agent memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies the supplied accepted records, not the source checkout. No source files were read for this job. Evidence is implementation or doctrine/design; no finding claims observed operation, causal support, fidelity or benefit. The retained inventory remains partial because the memory account includes incompletely traced controls and recovery alternatives as well as unknown initial content and opaque checkpoint/provider state. See RT-RTE-daemon-session, RT-OBJ-session-context, MEM-OBJ-harness-entry, MEM-OBJ-kernel-state and RT-CMP-provider-model.
Reconciliation establishes no supersessions. The epistemic factual/policy groupings overlap the finer memory objects without adding independent objects or stronger evidence. Session summary, payload and manifest remain separate parts of valid runtime containers. This profile therefore cites each actual part and consumer rather than counting overlapping groupings as additional memory. See EPI-OBJ-harness-memory, EPI-OBJ-harness-policy, MEM-OBJ-compaction-summary, MEM-OBJ-kernel-state and MEM-OBJ-kernel-manifest.
The unresolved resource-admission disagreement does not justify upgrading the whole installation route. Its findings retain the runtime's afforded basis; separately recorded current-session body expansion does not prove complete authoring, discovery, installation and reload wiring. Human resource authoring supports manual admission only on the explicit human-content path. Model-generated edits and refinement remain automatic even when a human starts the task or enables factory. Generic helper callers and unknown initial stores are not silently made manual or automatic. See RT-RTE-skill-install, RT-RTE-python-effects and RT-RTE-harness-refine.
Compaction and optional abandoned-path summarization qualify as trace learning: recorded activity feeds automatic retained summaries with later model consumers. Consolidation describes the implemented reduction, not verified semantic equivalence. The epistemic assessment of summary meaning remains indeterminate; it neither removes the wired summary path nor establishes synthesis. Refiner edits likewise support evolution and trace learning without warranting novel claims, successful criticism or improved capacity. See RT-RTE-context-compact, MEM-OBJ-compaction-summary, MEM-RTE-branch-navigation and RT-RTE-harness-refine.
Session-logs and tool-traces classify original recorded content, not serialization or the route's trajectory label. Mixed derived/custom inputs remain unresolved. Namespace capture preserves values, whereas manifest/notice reports names; neither the display nor a model's ability to create values resolves the opaque payload's original content or trace-fed production. No absence record supports a blanket trace-learning no. See RT-RTE-harness-refine, MEM-OBJ-compaction-summary, MEM-OBJ-kernel-state and MEM-RTE-kernel-revival.
Read-back separates requests from unsolicited selection within a chain. Requested harness search, stored-machine lookup and child collection are pull; automatic digest, template resolution, planner history and reconstruction are push. Lexical ranking supplies both ranking authority and a lexical push signal only on the automatic digest path. An actual subagent ID/title match selects full child templates; names on catalogs, checkpoints or requested objects alone do not establish identifier push. See MEM-RTE-digest-delivery, MEM-RTE-harness-inspection, MEM-OBJ-subagent-entry, MEM-BAP-subagent-template and RT-RTE-child-spawn.
The audit retains reasons and snapshots while automatic planner/digest readers receive projections. Those projections support consolidation and knowledge at their actual consumers, not complete reason read-back or validation of expected outcomes. Fingerprint freshness controls delivery, not endorsement; query drift alone does not force reselection. Delete/rollback changes future loads without recalling delivered copies. See MEM-OBJ-refinement-result, MEM-RTE-digest-delivery and RT-BAP-harness-digest.
Machine guards supply code-level validation/enforcement and routing, separately from model-interpreted task instructions. Static built-in machine prompts and ordinary current-run verdict/port state are not counted as use-derived memory merely because they influence progression. Completed retained child results do have explicit later collection; builder status remains a preview and does not establish full report delivery. See RT-OBJ-factory-machine, RT-BAP-machine-guards, EPI-RTE-port-capture and WF-RTE-builder-report.