EAL-Bench memory profile

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

Comparison rationale

All axes are partial because the accepted reports preserve material included alternatives whose write, selection or consumer paths were not exhaustively traced. Positive findings concern their named parts, not an exhaustive system value list. Missing primary execution archives prevent observed and causal classifications; they do not erase inspected wiring.

MEM-OBJ-memory-artifact is a valid container, not superseded by its typed and prose parts. The event snapshot and cumulative log remain separate consumer units. The sole object supersession affecting this profile replaces the public annotation object with EPI-OBJ-annotation-state. That record adds concrete file-backed Procurement production, while generic public caller control and persistence remain unresolved; the replacement does not resolve those alternatives.

Form follows consumption. JSON state is symbolic at parsers/reducers and interpreted content at models. No parametric finding is imported from the hosted inference components into the accumulated-memory boundary. Static domain criteria are not classified as retained memory merely because they validate generated material.

The feedback finding uses the profile's write-route definition of trace learning, not the separate improved-capacity meaning. RT-RTE-closed-feedback and its memory annotation establish an automatic recorded-action/outcome → seeded update → retained payload → later request chain. The adapter's synthetic actor and timestamp do not establish original session or environmental-event provenance. Tool-decision feedback supplies the positive source; other mixed inputs remain unresolved.

Consolidation, replacement, withdrawal and diagnostic widening have different warrants. The writer affords compression but does not prove faithful output; deterministic reduction and admission provide wired transformations. EPI-RTE-variant-construction explicitly establishes a new non-entailing permission in controlled widening, so synthesis is retained only for that intentionally untrusted branch. Merely erroneous extraction is not relabelled synthesis.

Pull describes requested saved-state reuse; push describes automatic delivery into later model contexts. Identifier findings name an actual chain/case/pool association selecting delivered retained parts, not a filename or an ID printed in a catalog. Whole-candidate model judgment is distinct from symbolic gold-fidelity comparison. Neither is collapsed into a universal retrieval policy.

No new records, quotations or source evidence were added. No target-source reads were performed for this classification; all warrants are the canonical accepted records and their annotations. The supplied reconciliation was used only to resolve identities and compatible coverage limits.