SkillRL memory profile

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

Comparison rationale

The profile separates retained instruction bodies, symbolic targeting fields, dense access representations and fitted actor state. The same JSON file can support different forms and authorities at different consumers. Its storage is not a repository substrate merely because the initial artifact is shipped in Git. The included inventory follows MEM-OBJ-skill-content, MEM-OBJ-skill-metadata, MEM-OBJ-experience-records, MEM-OBJ-retrieval-index, RT-RTE-sft-production and RT-RTE-policy-update.

Reconciliation supersedes EPI-OBJ-skill-principles with MEM-OBJ-skill-content. All profile findings use that canonical object. Containers and their inspected parts remain distinct: MEM-RTE-trace-abstraction and EPI-RTE-trajectory-cleanup do not replace the complete SFT route. Likewise the afforded withdrawal method is not a wired maintenance workflow. Its forgetting effect supports local decay, not history-preserving invalidation.

Requested manager retrieval is pull at the requesting operation; automatic observation supply is push at the later policy consumer. File loading by SFT stages remains pull. This union does not invent a foreground policy memory tool. The push signals distinguish lexical category selection, coarse slices and embedding ranking; a dynamic prefix is not an exact identity match. See RT-RTE-environment-rollout and MEM-RTE-experience-selection.

Automatic trace-fed writes qualify without evidence of improved capacity. Failure fragments can produce durable instructions and reward-bearing experience can change actor parameters. The learning authority of retained experience at an update consumer is separate from the output's later force. The controlled authority vocabulary does not describe the actor weights' downstream generative force, so that unit remains not-determinable rather than being labeled instruction or learning by analogy. See RT-BAP-policy-parameters.

Known forms and local mechanisms do not authenticate shipped instances or resolve generic API caller decisions. Missing source-control facts prevent manual admission and complete lineage findings. Missing generated instances prevent synthesis from being inferred from requests for new rules. The implemented reduction in the inspected SFT abstraction supports consolidation, without asserting that every generated proposition is semantically preserved. MEM-RTE-experience-save cannot be connected to a wired live learning loop without the missing caller and index refresh. Environment-specific repair semantics and alternate numerical workers remain locally unresolved. These uncertainties remain in units rather than being removed from the inventory.

All positive bases are source-level wiring or exposed affordance. No finding claims observed operation or causal benefit. No source files were read for this classification; it uses only the supplied accepted members and reconciliation.