GMemory memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The profile distinguishes content from access structures and follows the accepted records, not a reconstructed execution. Task documents support both readable examples and update inputs; the task-query graph is a different graph from recorded interaction graphs. Rule text supplies advice at solving consumers and learning input at update consumers, while associations rank candidates and scalar scores govern survival (MEM-OBJ-mas-message, MEM-OBJ-task-layer, MEM-OBJ-insights-metadata, RT-BAP-retained-guidance). No retained static system prompt is classified as learned instruction.
The two duplicate epistemic rule objects resolve to MEM-OBJ-insights-rule and MEM-OBJ-insights-metadata. Their supersessions affect rule form, storage, lineage and authority citations, not the underlying values. Key steps and failure explanations remain distinct consumed parts within task payloads. The stale attribution of effective reward pruning in memory's correction request does not warrant a curation value: the current RT-RTE-task-retention and MEM-RTE-state-roundtrip both establish that attempted deletion leaves backing states unchanged. The attribution discrepancy does not prevent classification of their agreed mechanism.
Model extraction and merging afford consolidation, while semantic novelty and preservation remain unresolved. EDIT supplies wired evolution and score decrement/pruning supplies decay. Requested causal generality does not prove synthesis, and rejection of a proposed duplicate does not itself deduplicate retained entries (RT-RTE-insight-maintenance, MEM-RTE-cluster-merge, EPI-RTE-rule-disposition). The same distinction prevents upgrading requested consolidation to observed model behavior.
Task-query fulfillment is pull, including its requested delivery into prompts. Count-triggered maintenance instead supplies automatically selected retained inputs to new update invocations; temporal edges likewise supply prior node output without a content request from that node. These push paths explain the identifier, embedding and coarse signals without relabeling every relevance selector as push (RT-RTE-memory-delivery, RT-RTE-insight-maintenance, MEM-OBJ-node-memory). A task name supplies an identifier signal only where membership actually selects delivered material.
Trace learning is an architectural write-route classification here, not a claim of improved capacity. Automatic trace-derived task extras and reusable rules persist for later consumers. Their originating multi-step task trajectories also contain executed environment action requests and results. Recorded agent prompts are not sufficient to add session-log extraction, and describing environmental feedback as a stream would not alone establish an additional source value (RT-RTE-task-retention, MEM-OBJ-mas-message, RT-RTE-insight-maintenance). MacNet's in-process forwarding remains distinct from these durable transformed artifacts.
Initial-store origin and approval, vector provenance and opaque provider retention remain explicit unresolved units rather than inferred manual authorship, training or absence. Opaque model state is not positively classified as use-accumulated model weights merely because the backend is parametric (RT-CMP-gpt-chat, RT-CMP-embedding-func). These limits prevent complete-value claims but do not weaken inspected automatic writes. No operation, improvement or causal benefit is inferred from wiring.
Source-read log: none. This classification used only the supplied members and contracts; it added no source evidence and reopened no report verification.