WikiSkill memory profile

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

Comparison rationale

All positive findings classify accepted implementation or doctrine records, not an observed run. Files contain interpreted procedural text and separately consumed symbolic metadata. Model weights belong to external inference components, not to the retained memory updated by WikiSkill. Whole-runtime permissions and packaged static role guidance therefore do not become memory-axis values.

The reconciliation supersedes the epistemic product-pattern declaration with MEM-OBJ-patterns. This profile cites that canonical object throughout. The product journal remains a distinct control container, and research Wiki, impact views and paper SKILL/PURPOSE mappings remain separate operative parts. Their different consumers prevent flattening them into a single instruction object.

Trace learning is a wired update path, not evidence of improved capacity. Selected experience can update Wiki, Wiki can shape a procedural proposal, and retained or candidate instructions can reach a later invocation. The study's Maintainer-to-Proposer bridge qualifies without inventing a second final-skill deployment. MEM-OBJ-research-evidence, MEM-OBJ-paper-skills and RT-RTE-isolated-study retain those horizons. Opaque product output and optional trace origin remain unresolved rather than receiving a universal trace-source value.

Semantic consolidation requested of a model stays claimed. Deterministic reduced trace views and extracted summary views support a separate wired consolidation finding; replacement and local patches support evolution. No actual input/output semantic comparison establishes synthesis. Skill-gate rejection preserves Wiki rather than invalidating its claims. Deployment restoration changes checked bytes, not knowledge history. These distinctions follow MEM-OBJ-research-evidence, EPI-RTE-wiki-index, EPI-RTE-paper-contract and RT-RTE-skill-install.

Each axis retains an uninspected unit for additional callable OfficeQA/Spreadsheet controllers, including promotion and candidate-file resume paths named in the accepted epistemic coverage. The broad engine/study records do not resolve those alternatives. That inventory limit prevents complete system classification even where every named inspected object uses files. Generic callers similarly prevent a manual write-agency finding, and an opaque output cannot inherit the representation or authority of its displayed result row. Uninspected host discovery does not acquire a push selector from the installed filename.

The paper/study trace-source unit retains tool-traces and trajectories but is partial: MEM-OBJ-research-evidence names session call/output fallback items without establishing recorded conversational turns. No session-logs finding follows from the container name; the fallback's source content remains unresolved locally.

Export, installation and restoration are filesystem operations, not an independently wired pull read-back route. RT-RTE-skill-install affords future host consumption without establishing its request, selector or delivery into a later invocation. That direction unit remains uninspected with no positive finding. The separate staged-file model pull remains wired on its own records.

No source files were read for this profile. All classifications use the supplied accepted records and annotations; no records, quotations or new evidence were added.