From Agent Memory to Portable Skills memory profile

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

Comparison rationale

All findings retain claimed strength: the accepted records describe an article-level design, not inspected implementation or observed use. Coverage assessments therefore describe classification within named scopes, not strength of proof. The Neo4j package is not assigned files merely because its export is called SKILL.md. The structural graph and readable procedure have different consumption paths; arbitrary tool payloads stay unresolved. See MEM-OBJ-tool-call, RT-OBJ-procedure-graph and RT-OBJ-skill-package.

The learning authority of recorded experience belongs at the distillation update consumer; instructional authority belongs to the later host. Automatic production of a durable instruction artifact supports trace learning even though a human separately approves initial publication. Neither that classification nor the presence of source references establishes semantic warrant, exercised capacity improvement or parameter learning. See RT-RTE-skill-distillation, RT-RTE-review-delivery, RT-BAP-skill-consumption and RT-CMP-llm.

Consolidation is the described reduction of recurring retained activity. Deterministic topology and model authorship are separate derivation contributions, not proof of entailment or a novel claim. Source-survivor repair evolves grounding and supersedes a version; contradiction flagging alone does not demonstrate revocation. The meaning of generated claims remains unresolved, as does whether installed consumers receive a revised version. See EPI-OBJ-claims, RT-RTE-step-repair and RT-RTE-drift-check.

Requested evidence selection and client fetching support pull. The document does not settle every host-loading or governance trigger, so signal coverage is not-determinable rather than inapplicable. An evidence-link identity on a requested read is not a push signal. See RT-RTE-review-delivery and RT-RTE-skill-distillation.

Reconciliation establishes no supersessions: scoped-input and package containers remain valid beside their parts. Consequently no citation or value was redirected. No source files were read for this classification; all warrants come from the supplied accepted records and their annotations. Model storage and parameter lifecycle remain outside the through-use memory inventory rather than being inferred from model inference.