WikiSkill memory profile

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

Comparison rationale

This comparison classifies accepted records of the documented Raw/Wiki/Skill loop, not an inspected executable system. All positive findings have claimed basis. Source-native descriptions and author-reported examples do not establish wiring, observed activation or causal benefit. Coverage and evidence strength are independent: a documented part can have known classification at claimed strength while the enclosing axis remains partial.

Inventory and unresolved alternatives

The seven filesystem objects are distinct payloads, not seven copies of runtime routes. Retained training records contain both external evidence and model-produced predictions; imported evidence and generated content must not inherit one another's lineage. The incoming and generated parts are identified by EPI-OBJ-imported-evidence and EPI-OBJ-produced-predictions. The profile does not add current task evidence or pretrained model parameters to accumulated memory simply because the runtime consumes them.

Potential provider state, retained validation trajectories and durable restore snapshots remain unresolved in RT-CMP-model-backends, MEM-OBJ-raw and RT-RTE-validation-gate. Those records do not establish an opaque provider memory payload, a replay archive or a durable checkpoint; neither do they warrant excluding those alternatives to obtain complete coverage. Initialization and companion maintenance are separately unresolved. Unknown means no controlled value established within that named branch, not evidenced absence.

Authority and direction

Experience has learning authority at automatic wiki/skill update consumers; the resulting procedures have instruction authority at task inference. Wiki explanations and intervention history supply knowledge and update guidance, not universal enforcement. Index summaries also influence relative relevance of page reads, supporting ranking authority at the proposer without a learned ranking algorithm. These effects are separate from the retained content's write agency (MEM-OBJ-index, RT-RTE-skill-proposal, RT-BAP-skill-injection). The current scalar threshold supplies the harness's gate, but it is ordinary loop state rather than an established accumulated retained criterion object. Its enforcing comparison is not attributed to model-read impact history.

The initial wiki/index/history supplies, stratified trace delivery and full skill injection are push. Requested read_file results are pull, even though a model chooses relevant pages using index summaries. Paths resolve those requests; no accepted record establishes targeted identifier-based push. Coarse outcome categories, availability and budgets remain coarse selectors (RT-RTE-wiki-maintenance, RT-RTE-task-execution). Alternative inference wiki access has an unresolved selector, so its signal is not inapplicable. Separate PURPOSE.md delivery is also unresolved, rather than inferred from its rationale role (MEM-OBJ-purpose, MEM-BAP-wiki-guidance).

Transformations and trace-fed writes

Condensed index/log views support consolidation; explicit patches/refinements support evolution. Skill rollback withdraws active reliance while retaining the failed intervention in history, supporting bounded invalidation rather than wiki-wide retraction. These are declared transformations, not executed curation observations. No-duplicate instructions alone establish no semantic deduplication. Requested root causes and novel remedies cannot establish synthesis when the accepted content-edge assessment remains indeterminate (EPI-RTE-wiki-formulation, EPI-CLM-knowledge-growth, RT-RTE-validation-gate).

The trace-learning positives follow automatic trace/feedback-to-durable-artifact-to-later-consumer chains, including wiki views, procedures and retained intervention feedback. They do not require new propositions, parameter updates or observed improvement. Archiving raw traces alone is not a qualifying update. The original training experience includes linked trajectories and recorded tool activity; fixed benchmark references do not become traces merely because they enter the same update context. Validation feedback has a described rollout origin without an established persisted raw validation archive (MEM-OBJ-raw, EPI-RTE-impact-retention). No complete negative is derived from an unresolved companion or provider branch.

Supersessions and disagreement

Reconciliation supersedes the seven epistemic duplicate object IDs with the corresponding memory IDs: raw traces, wiki patterns, wiki index, evolution log, skill impact, skill content and skill purpose. Every profile object citation uses the canonical memory declaration. Supersession connects identity and keeps both analysts' findings; it does not upgrade evidence or erase unresolved consumers. The epistemic imported-evidence/produced-prediction parts and the diagnostic, formulation, retention, scoring, disposition and integration phases remain distinct records, not superseded runtime containers.

Reconciliation retains one delivery disagreement: RT-RTE-validation-gate describes later proposer/maintainer history contexts, whereas MEM-OBJ-skill-impact and EPI-RTE-impact-retention leave necessarily complete maintainer tracker supply unresolved. The profile retains the positive proposer route and an unresolved maintainer unit; it selects neither side as a complete joint delivery fact. Unified diffs versus full rejected-content wording also cannot establish reconstructable whole files. The log-name ambiguity changes no files finding, while companion delivery/patch/rollback uncertainty prevents a complete authority or consumption account.

Source-read log

No source files were read. Classification used only the supplied accepted runtime, memory, epistemic and workflow members and their reconciliation. No additional evidence, quotations or records were introduced; no report verification was reopened.