RSIAgent memory profile

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

Comparison rationale

All positive bases concern inspected wiring, explicit capacity or requested model behavior. No finding is observed or causally supported. Coverage remains partial on every axis: the accepted inventory includes opaque payloads, partially inspected recovery variants and inaccessible provider internals. Local known units do not close those alternatives. Provider weights are runtime components, not demonstrated use-updated memory; possible use-accumulated provider state is retained as an unresolved unit instead of attributing local history properties to it.

The accepted reconciliation supersedes the epistemic bank object with MEM-OBJ-memory and the epistemic WORK LOG object with MEM-OBJ-work-log. Classifications cite these canonical objects. The narrower EPI-RTE-worklog-fold remains distinct from MEM-RTE-worklog-fold, which also covers restart synthesis and strategy review. No other supersession changes a value.

The bank's arbitrary bytes do not establish one uniform form or authority. Requested prose has an afforded natural-language advice path through RT-BAP-actor-method; scripts/data and opaque scratch remain unresolved. Image schema/identity metadata is symbolic, while pixels have no established value among the controlled forms. Copied journal payloads retain bank uncertainty; manifest fields are separately symbolic inputs to host validation, enforcement and routing (MEM-OBJ-memory-journal, MEM-OBJ-durable-images, MEM-OBJ-verifier-scratch).

Learning authority belongs to retained experience at the update consumer; downstream bank/log/diagnosis products supply knowledge as interpreted context. Neither implies improved future capacity. The same-Actor bank and summary chains qualify for trace learning because automatic trace-fed production, retention and later consumption are wired. Curriculum note-specific derivation remains afforded and partial. Verifier raw replay, snapshots and image preservation alone do not establish an additional extracted behavior-shaping learning product; opaque scratch prevents a negative. No bounded absence records justify a trace-learning no or an inapplicable trace-source axis (RT-RTE-learning-admission, MEM-RTE-worklog-fold, MEM-OBJ-curriculum-notes, MEM-RTE-verifier-replay).

Terminal bank and diagnosis production consumes linked Actor decisions/actions/observations and terminal Verifier outcomes as multi-step experience. Its trace sources therefore include trajectories alongside session logs and tool traces; these values can describe overlapping content. No separately serialized trajectory object or additional learning-unit schema is required. This classification follows the accepted terminal-learning route and its memory annotation, not merely the presence of a tool history. It does not resolve note-specific inputs, other recovery variants or provider-side state (RT-RTE-learning-admission, EPI-RTE-memory-proposal, MEM-OBJ-learning-diagnosis).

Curation bases remain separate. WORK LOG folding implements consolidation even if model output loses a negation; indeterminate semantic relation does not establish synthesis. Bank criticism/reconciliation is requested, revisions and history-preserving withdrawal are afforded, and no generated corpus establishes completed semantic repair. Structural replay forgets actionless live-context pairs without deleting raw artifacts; attachment identity deduplication is not relevance ranking (EPI-RTE-worklog-fold, EPI-RTE-memory-proposal, MEM-RTE-verifier-replay, MEM-OBJ-durable-images).

Bank delivery has both push and pull: host supplies the configured tree/inventory; the Actor requests selected file bodies. Filename requests are not identifier-based push, and foreground relevance judgments are not automatic inferred-judgment retrieval. Summary generation's semantic selection occurs at writing; availability of its retained result selects later delivery. Checkpoint hashes validate whole lineage bundles rather than selecting relevant evidence by identity match. These coarse push routes do not establish reliance or successful delivery on all task paths (RT-RTE-task-execution, RT-RTE-lineage-recovery, WF-RTE-phase-transition).

Missing initial-bank authorship/approval, generated bank/note/scratch instances, variant recovery bodies and provider internals prevent stronger inventory and value completeness. Static charters do not become authored instruction memory merely because later calls receive them. Official score archives and cohort reporting remain measurement/reporting rather than an evidenced adaptive memory loop here (MEM-OBJ-memory, MEM-OBJ-verifier-scratch, EPI-RTE-score-report).

Source-read log: none. This profile uses only the supplied accepted members and reconciliation; it adds no records, quotes or source evidence.