OpenResearch memory profile

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

Comparison rationale

The comparison covers accumulated research and session material, not every byte the runtime stores. SQLite transcript rows, native session payloads and bootstrap prose are distinct. A readable transcript does not resolve a native checkpoint's form, provenance, selector or update semantics. Known local facts therefore coexist with unresolved alternatives rather than licensing complete axis inventories; see MEM-OBJ-native-session, MEM-OBJ-chat-messages and MEM-OBJ-bootstrap-context. The accepted inventory also leaves literature-cache, remote and platform lifecycle details untraced. These are coverage limits, not evidenced absences.

Fallback compaction is the strongest recorded trace-learning witness. The automatic chain retains a summary of bounded recorded session work and prior seed, then supplies it to a later invocation. User initiation does not make admission of the generated summary manual. This is consolidation and trace learning under the comparison contract, without observed improvement or reliable preservation of reasons. Selected-path ACP/side bootstrap is another retained continuation representation of recorded activity; ordinary transient recovery text alone is not a separate retained write. See MEM-RTE-chat-compaction, MEM-RTE-branch-context and MEM-BAP-bootstrap-injection. Earlier seeds and native checkpoints remain separate uncertainty units. Learning authority at the summary-update consumer and knowledge authority at the later model consumer are different paths.

Requested log/file/description returns are pull. Ordinary goal/seed/shell-context supply and subscribed wake-up are push operations within longer chains. Run identity on a requested log is not a push signal. Subscription identity and ancestry matching do select delivered retained parts; availability, readiness and budget gates remain coarse selectors. Slash expansion fulfills a caller-selected instruction request, while native skill discovery/loading stays unresolved. See MEM-RTE-turn-context, MEM-RTE-log-retrieval, RT-RTE-run-wakeup and RT-RTE-skill-upload.

Deletion/pruning supports only local forgetting; it does not retract exported or already delivered content. Description/skill replacement supports evolution without implying synthesis. Archive flags affect visibility rather than evidential reliance. Visibility and changed comparison context do not establish relative priority, so the archive/parent authority unit emits no controlled finding. The records do not support ranking at that consumer, despite the epistemic ledger's wording. Winner promotion is retained only at its claimed instructional basis, not upgraded by wired branch/run primitives. See MEM-RTE-artifact-access, EPI-RTE-description-store, EPI-RTE-tree-organize and EPI-RTE-winner-promotion. Static renderer validation does not establish that retained memory itself supplies a validation criterion; see EPI-RTE-chart-render.

Reconciliation resolves the duplicate epistemic run-log and research-output objects to the memory run-log and project-artifact declarations. It also resolves epistemic/workflow log-preview routes to the memory log-retrieval declaration, and epistemic artifact retention to memory artifact access. All profile citations use those canonical declarations. Description, tree organization and chart production remain distinct, as do turn composition and its enclosing turn route. These identity resolutions change no evidence basis.

No source files were read for this profile. All classifications use the supplied accepted records and their retained passages. Wired findings establish implemented routes, not exercised behavior, semantic correctness, model compliance or causal benefit. No bounded absence record supports trace-learning no, nor complete absence of any curation or authority value.