Agent Workflow Memory memory profile

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

Comparison rationale

This profile classifies accepted records, not new source observations. Reconciliation makes no supersession: broad epistemic containers and narrower memory parts remain valid, so citations retain their original declaring members. Optional retrieval uses its own afforded consumer handoff; the broader wired selection record does not upgrade either that handoff or the broken semantic workflow branch. MEM-RTE-workflow-selection and MEM-RTE-exemplar-ablation preserve the distinct index handoffs.

The operative workflow store is a file, not a workflow graph. Action notation is interpreted as guidance. WebArena's system-message procedures have instruction authority, while the whole file's local factual assertions about orders, dates and website locations also supply knowledge. System-message placement does not erase this reference-context effect. Mind2Web's user-message demonstrations supply knowledge. At the separate update consumer, saved experience and imported offline annotated training examples have learning authority because they feed durable behavior-shaping files. These paths do not establish compliance, improved capacity or causal benefit. See EPI-OBJ-workflow-text, MEM-OBJ-arena-workflows, RT-BAP-workflow-system, RT-BAP-workflow-exemplar, MEM-OBJ-conversation, MEM-OBJ-arena-traces, RT-RTE-mind-admission and MEM-OBJ-mind-workflows.

Eligibility signals supply both a criterion and a veto: validation and enforcement apply to source admission, not to the truth of generated workflows. Retained evaluator reasons have a possible human-review consumer, but induction consumes the Boolean rather than those reasons. Metric records are diagnostics, not the online writer's success gate. See MEM-OBJ-autoeval-signal, MEM-OBJ-autoeval-thoughts and EPI-OBJ-annotation-metrics.

Generated manual-rule examples retain trace-extracted lineage: human approval changes admission control, not the origin of selected query/thought/action content. See RT-RTE-rule-admission and MEM-OBJ-arena-workflows. Known later transformations do not resolve the initial lineage or content-control of shipped exemplars and website files. The vector access route likewise establishes derivation from selector text, not exact embedding-parameter provenance. The final screenshot remains an unresolved consumed part on the form axis rather than being classified through its caption. These limits are carried in MEM-OBJ-exemplar-content, MEM-OBJ-exemplar-specifier, MEM-OBJ-arena-workflows, RT-CMP-openai-embeddings and EPI-RTE-trajectory-check.

Curation separates inspected reduction from requested abstraction. Saved-message projection and concrete rule extraction reduce retained experience; abstract-action deduplication selects representatives. Neural prompts ask for common subroutines, but generated semantic fidelity is unobserved, leaving their actual transformation partial rather than asserting synthesis. Whole-file regeneration does not establish editing an existing entry, decay, promotion, or invalidation that preserves history. See MEM-OBJ-conversation, RT-RTE-rule-admission, RT-RTE-mind-admission, RT-RTE-neural-admission and EPI-OBJ-workflow-text.

Read-back is pull at the requested consumption operations. A configured file path is not an identity-matching push selector, and supplying requested text is not another push. Optional retrieval is requested preprocessing followed by an optional configured file read. Enabled scratchpad history is also a configured context request, not independently unsolicited supply. Thus read-back signal is inapplicable over the complete recorded direction inventory, not absent or embedding-valued simply because retrieval uses embeddings. See RT-RTE-benchmark-action, RT-RTE-browser-action, MEM-RTE-workflow-selection, MEM-RTE-exemplar-ablation and MEM-RTE-episode-memory.

Trace learning here is automatic trace-fed durable artifact production, not proven improvement. Offline annotated trajectories can qualify even though they are imported rather than this agent's own runs. Online prediction sessions qualify without live browsing. WebArena automatic routes qualify when selected traces reach regeneration; the default filename mismatch and possible zero-example write prevent treating every scheduled update as an exercised trace-fed one. Manual rule approval stays outside the automatic route. Scratchpad display is only afforded, ordinary run.py disables it, and its durable retention is not established; neither a system-wide no nor a resolved trace source follows. See RT-RTE-mind-admission, RT-RTE-neural-admission, RT-RTE-rule-admission, WF-RTE-pipeline-handoff and MEM-RTE-episode-memory.

Source-read log: none. Classification used only the supplied accepted records and reconciliation; no source ambiguity required a permitted source read. There are no new records, quotations, observed runs or causal results.