AREX-Skill memory profile

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

Comparison rationale

The profile classifies accepted records, not the entire application's storage. Installed or revised guidance qualifies when later discovery or body reads consume it. Static bundled procedures alone do not establish accumulated memory. The filesystem claim follows MEM-OBJ-repo-skills, not the fact that the source was acquired from Git. Provider weights remain an external component, not an inferred memory substrate.

The ten inventories remain partial. Known container storage does not determine opaque payload form, lineage, admission or consumer force. In particular, MEM-OBJ-custom-entry distinguishes arbitrary state/details from model-facing custom-message content. MEM-OBJ-workflow-journal distinguishes the full interpreter value from readable report previews. Remaining specialized maintenance and script/evaluator paths have no complete accepted-record inventory; no value is supplied merely from a filename or a coverage-table mention.

Default continuation summaries establish automatic trace-fed retention and later consumption through EPI-RTE-summary-generation, RT-RTE-context-compaction and MEM-RTE-branch-navigation. Their learning force at the update consumer and knowledge force at the later model are separate. This does not contradict an indeterminate epistemic content relation: intended compression establishes the curation mechanism, not fidelity, new warranted claims or improved future capacity. No synthesis value follows from possible summarization error.

Diagnostic truncation in MEM-OBJ-agent-history reduces retained content and preferentially forgets older material, but its established consumer is display. Raw log retention and symbolic result replay likewise do not automatically inherit the summary route's trace-learning value. There are no accepted evidenced-absence records warranting a local trace-learning negative. The supported wired positive survives in a derived union without turning unresolved alternatives into known units.

Read-back separates automatic context supply from requested retrieval. RT-RTE-skill-delivery supports catalog push and explicit body pull. Router-guided model judgment belongs to demand-driven reads, not inferred-judgment push. Session projection actually matches retained tree identities; catalog names and requested file paths alone do not establish that signal. MEM-RTE-workflow-context pushes a fixed report projection at completion, whereas MEM-RTE-workflow-resume returns a requested cached value. Neither is an embedding retrieval mechanism.

Admission also has distinct layers. The software transactions in RT-RTE-repo-import and RT-RTE-library-update admit content computationally. The explicit approval alternatives in WF-RTE-repo-verification support manual agency only at the policy layer. Starting a workflow or supplying an overwrite flag is not authenticated per-content human approval. Initial authorship and extension replacement control remain unresolved.

Reconciliation supersedes the memory member's repository verification/revision route with WF-RTE-repo-verification. All affected curation, agency and feedback classifications use that canonical record. EPI-RTE-refine-guidance additionally covers paper refinement and is not superseded. The graph and summary objects with overlapping extents remain valid separate records; narrower text witnesses do not resolve arbitrary details or all candidate lifecycle stages.

Source-read log: none. This classification used only the supplied accepted members and reconciliation; no frozen source ambiguity required an additional read. All warrants and evidence layers remain in the cited records. No observed operation or causal benefit is inferred.