SwarmWorld memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The accepted records establish implementation wiring, not observed episodes or benefit. All positive findings therefore use wired. The ten axes remain partial because the reports leave offline transformation and downstream consumer internals uninspected; a whole-system frozen source boundary does not establish complete memory classification. See MEM-OBJ-event-trace and EPI-RTE-proposal-evaluation. These unresolved alternatives are retained on every axis rather than silently excluded.
Reconciliation supersedes EPI-OBJ-empirical-knowledge with MEM-OBJ-empirical-knowledge. The profile uses that canonical object. Other containers and intersecting routes remain distinct: private collections, research-state payload, registry instructions, executable programs and skill indices have different consumers. No other supersession changes a value.
Retained prose is knowledge at model-context consumption; its delivery is not proof of compliance. Numeric feedback is ranking input downstream and an automatic trace-fed behavior-shaping update. Resource/provenance and known-parent data supply both validation criteria and enforced vetoes. Installed instructions have executable force at the VM, separate from their descriptive use in model context. See MEM-BAP-evidence-admission, MEM-RTE-retrieval-feedback and RT-BAP-program-execution.
Explicit human content-bearing actions support manual admission locally; starting an episode does not convert software sensing, model-state admission or feedback into manual writes. The ordinary software/model routes remain automatic. See RT-RTE-direct-control and MEM-RTE-social-retention.
The continuation-state route qualifies as trace learning because automatically generated state persists and reaches a later macroturn, not because a hypothesis field proves novelty or improvement. State-update semantics remain indeterminate for curation. Empirical summaries consolidate; eviction forgets; feedback revises access metadata. Reversible prompt factoring deduplicates strings without certifying semantic synthesis. Request fulfillment withdraws pending status without deleting ledger history. See MEM-OBJ-research-state, MEM-OBJ-event-trace and MEM-OBJ-message-records.
Scheduled context is push even though internal code calls retrieval. Requested record resolution, teaching and trace replay are pull operations. Identifier push requires actual focus/recipient/reply matching; identifiers on requested content alone do not qualify. Trimming changes delivered context, not authoritative store retention. See MEM-RTE-context-selection and MEM-RTE-social-retention.
Trace sources follow original content. Simulator activity supports event streams; linked recipe/sample/test experience supports trajectories on its enabled route. A fresh provider message pair does not alone establish retained conversational session logs, nor do simulator actions alone establish tool traces. Unknown original origins of prior prose prevent a complete research-state source union. Replay restores committed research records, not all adaptive selector evolution; recorded digest agreement would not establish complete adaptive continuation. See RT-RTE-recorded-replay, MEM-OBJ-research-mission and MEM-RTE-retrieval-feedback.
Source-read log: none. Classification used only supplied accepted records and their annotations; no source ambiguity required an additional frozen-source read.