Agent-R memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The retained tree has two operative parts. Text supplies context to the revision verifier; numerical and structural fields rank paths and constrain eligibility. Their shared JSON substrate does not make their forms or authorities identical. The inventory also includes exported content, flags, diagnostics, converted output, evaluation results and optional progress storage. Storage and form are known for these recorded objects; this does not establish later semantic use of every stored field. See MEM-OBJ-extended-node, MEM-OBJ-node-statistics, MEM-OBJ-loss-flags, MEM-OBJ-revise-feedback, MEM-OBJ-final-result and MEM-OBJ-progress-file.
Revision construction both consolidates existing path content and synthesizes an inserted self-error admission. The latter is not entailed by copied observations or outcomes. Prefix equality suppression supplies the separate dedup finding. These classifications concern implemented transformations below the defective numeric thresholds, not successful nonempty CLI exports. Good-only and no-critique modes retain their different construction paths without repairing the thresholds. See RT-RTE-revision-export and EPI-RTE-conversation-splice.
Experience has learning authority at the durable training-artifact construction consumer, independently of knowledge authority at the verifier. Trace learning here describes automatic trace-fed artifact production, not demonstrated improved capacity. The dataset construction is wired but conditional; external optimizer consumption and later actor adoption are claimed/unresolved. Conversion is only afforded. These distinct bases cannot be combined into observed training or a closed improvement loop. Static corrective wording inside an example is not standing instruction authority for a deployed agent. See MEM-OBJ-revise-log, RT-CLM-iterative-training and RT-ABS-training-machinery.
The original trace source is multi-step environment experience. Chat-shaped serialization does not add session-logs, and action/observation pairs inside linked search paths do not establish a separate tool-traces or event-streams input. The imported initial instruction and fabricated wait exchange remain non-trace contributions to the derived conversation. Original environment and checkpoint provenance, supplied-prefix origin and actual optimizer inputs remain unresolved. See MEM-OBJ-extended-node, MEM-OBJ-revise-log and MEM-OBJ-final-result.
Requested batch tree consumption is pull. Selection within that request does not become unsolicited push merely because code selects leaves or delivers text to the requested verifier. Filename identifiers therefore establish no positive push signal. Auxiliary helpers and external training do not establish their own retained-source delivery chains, so read-back signal is locally inapplicable on the known pull operations but not determinable across the full inventory. See RT-RTE-tree-collection, MEM-RTE-dataset-rewrite and MEM-CMP-context-helpers.
Reconciliation supersedes the duplicate epistemic rewrite route with MEM-RTE-dataset-rewrite. This profile cites that canonical route while retaining the distinct converted output object, EPI-OBJ-conversation. The supersession changes identity resolution, not an evidence basis. Dropped feedback and loss flags prevent assuming that every prospective training format retains rationale or mask semantics. The remaining mixed containers are not superseded by their operative parts.
Source-read log: none. Classification used the supplied accepted records and reconciliation only; no new source evidence or declarations were introduced.