SkillX memory profile

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

Comparison rationale

This profile classifies accepted records, not source execution. All positive findings concern inspected wiring unless explicitly marked otherwise. No record establishes exercised downstream compliance, improved capacity or causal gains. Coverage remains partial where initial content, remaining bundled artifacts, optional transformations or trace origins are unresolved. Those limits do not erase automatic trace-to-instruction writes.

Reference implementations are natural-language material at their model consumer, not symbolic executable tools. Library keys, eligibility fields and checkpoint controls are symbolic at their fixed-code consumers. Dense retrieval encodings are parametric access material; the numerical vectors are not provider weights. The persisted FAISS-HNSW label establishes no additional implemented index substrate. See MEM-OBJ-functional-skill, MEM-OBJ-retrieval-vectors and MEM-OBJ-skill-library.

History has knowledge authority as evidence and learning authority at the durable update consumer. Output plans and reference skills provide contextual advice, rather than standing host instructions. The downstream formatter route is distinct from the update route; neither requires observed improvement for wired classification. See RT-RTE-library-build and RT-BAP-skill-guidance.

Requested service task/step retrieval is wired pull even when code automatically makes subsidiary vector queries. The alternative public PlanRewriteService.get_pseudo_plan invocation is a separate afforded pull finding: its wired internal progression does not establish a shipped caller or later delivery. In contrast, subsequent extraction models do not request the provisional proposal list or prior tool entry: code selects and supplies them. Whole-list availability is coarse; the atomic map's actual tool identity match is identifier selection. Automatically substituted feedback and checkpoint recovery are separate coarse supply operations, not evidence of unsolicited embedding-based push. See RT-RTE-prompt-prepare, MEM-RTE-proposal-context, MEM-RTE-feedback-summary and MEM-OBJ-task-checkpoint.

Consolidation does not require verified semantic faithfulness. Feedback compression and cluster merging have implemented reduction routes; semantic novelty in actual model outputs remains indeterminate, so synthesis is not inferred. Same-name revision does not imply history-preserving invalidation, global predecessor withdrawal or rollback. See MEM-RTE-feedback-summary, RT-RTE-library-build and MEM-OBJ-skill-library.

Reconciliation supersedes EPI-RTE-feedback-summary with MEM-RTE-feedback-summary and EPI-OBJ-plan-content with MEM-OBJ-plan-skill. All classifications use the canonical memory records. Combined raw/summary, skill-content and synthetic-task records are not superseded by their noncoextensive parts. The corrected exporter preflight remains blocked, and the afforded public rewrite helper does not repair the shipped service mismatch or establish persistence. See EPI-RTE-export-validation, RT-RTE-prompt-prepare and MEM-CLM-format-boundary.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no frozen-source ambiguity required an additional read.