OS-Copilot memory profile

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

Comparison rationale

This revision-2 mapping uses accepted implementation records, not observed runs. All positive implementation findings are wired unless their local basis says afforded. The afforded manual path concerns a human supplying or editing a selected tool file; it does not upgrade generic manager calls to manual admission. Provider inference does not establish provider learning or retained memory. Its possible retained alternatives remain uninspected on every axis. The page service has request-time transformations but unresolved collection lifecycle and no established later-memory consumer; it remains an unresolved alternative rather than a classified persistent summary store.

Canonical identities and coverage

The reconciliation supersedes the epistemic description object with MEM-OBJ-tool-description, the epistemic course-text object with MEM-OBJ-course, and the memory page-context route with EPI-RTE-page-content. The mapping uses those canonical records. The web-model container remains valid; its parts do not imply separate retained stores. Generated/current candidate code is broader than MEM-OBJ-tool-code; only its retained part is classified here. These supersessions change citation identity, not evidence strength or certainty.

The known inventory spans payloads, access structures, write mechanisms and later consumers. It retains separately unresolved initialization, direct manager writes, service lifecycle and provider retention. Consequently no axis claims a complete system-wide value inventory. The ordinary graph, transient critique and example histories have no established accumulated memory route in RT-OBJ-tool-graph and RT-RTE-example-loops. Static prompts are shipped methods rather than learned content in RT-OBJ-friday-prompts.

GAIA rows belong to the inventory because EPI-RTE-gaia-comparison establishes a later accounting/resume consumer. Their narrow exact-match validation and resume-control forces do not establish agent knowledge or a trace-learning artifact. The record supplies no canonical bounded absence warrant for a controlled trace-learning no. This separates retained control/accounting state from display-only logs without asserting complete absence. No dataset quality or postprocessor semantics are inferred.

Local mappings and unions

Tool descriptions have both knowledge authority at planning and ranking authority through their embeddings. Name metadata routes selected payloads. Retained source remains symbolic code used as examples, not automatically executed retrieved functions. See MEM-OBJ-tool-vectordb and RT-BAP-tool-context. Neither a retained instruction consumer nor learning authority follows merely from the tool admission route's trace-learning membership: raw execution feedback is not itself retained as an update-consumer input by this store.

The tool chain pulls requested relevant content, even though runtime transitions initiate requests automatically. Its semantic selector is not a push-signal finding. In contrast, curriculum design receives an automatically selected prior-course window. That push uses dictionary order/count, not an identifier match, recency timestamps or embeddings. Filename-based course initialization is a separate pull operation. Prior-course context supplies knowledge to curriculum design, while lesson dispatch supplies instructions to the agent. See MEM-OBJ-course and RT-BAP-course-tasks.

Docstring selection is consolidation; overwrite/key replacement is evolution. Deletion supports bounded forgetting, not historical invalidation or complete revocation: live dictionary copies can survive and later be reserialized. Index rebuilding and existing-name checks are not deduplication. Requested nonduplication/progression does not prove those curation operations happened. See MEM-OBJ-tool-description, RT-RTE-manual-maintenance and MEM-OBJ-course.

The automatic tool route qualifies as trace learning because execution feedback controls durable admission and later generation receives retained artifacts. The admitted payload need not contain the raw trace. It is still model-authored code, with a separately compiled description and vector, not necessarily trace-extracted code. The repair-score/original-code defect limits faithfulness and reuse warrant without removing the wired route. Tool traces are the established qualifying source; no session-log, event-stream or linked trajectory source is inferred. See RT-RTE-tool-admission and MEM-OBJ-tool-code. A derived union retains this positive yes without manufacturing a local accounting-only no from a route that declares no bounded absence. The positive does not mean observed improvement, and missing optional/demo and direct-caller provenance prevents a complete negative elsewhere.

Source-read log

No source files were read for this classification. The supplied pinned members, types, reconciliation and records contract were sufficient. No new records, evidence passages or source observations are introduced.