Reflexion memory profile

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

Comparison rationale

This profile classifies accepted implementation records, not executed behavior. Every wired finding is conditional on reaching the inspected path and on its external calls returning. No observed or causal basis is supported by the supplied members. All axes remain partial because their inventories include unresolved parts, not because source wiring is weak evidence of the parts it actually exposes.

The central boundary is task-local accumulated experience. MEM-OBJ-alfworld-memory, MEM-OBJ-hotpotqa-reflections and MEM-OBJ-reflections give advice to later model calls. Advice is knowledge authority here, not standing instruction authority. Raw experience has learning authority at the automatic update producer; its derived output has a separate downstream advisory role. Executable persisted programs instead have instruction authority at the actual code consumer in MEM-RTE-result-validation. Success flags and result counts have host-control authority, not the language interpretation granted to plan text.

MEM-OBJ-webshop-memory qualifies for trace learning through the later reflection consumer, despite the absence of plan delivery to the action model in RT-RTE-webshop-trial. This positive does not claim an action-guidance improvement pathway. HotPotQA's retained lossy prior-attempt excerpt also qualifies when automatically written and supplied to a later call; neither novelty nor a model-generated semantic summary is required. Its compression and local forgetting do not imply consolidation of every accumulated diagnosis into a checked theory.

Requested configuration and file acquisition is pull. Automatic indexed-plan, latest-value and solved-row selection is push at its own later operation. Environment-index correspondence is an actual identity selector; a UCS outcome-bit ID merely passed to missing search code is not evidence that identity matching selected the material ultimately delivered. Latest/last-three, strategy, budget and outcome predicates remain coarse signals. Read-back is not inferred from physical serialization alone: MEM-OBJ-hotpotqa-checkpoints establishes a writer, not a saved-agent reload caller.

The supersession of the epistemic program-diagnosis object resolves to MEM-OBJ-reflections. The two standalone audit supersessions resolve to MEM-RTE-result-validation and MEM-RTE-leet-evaluation. Their duplicate evidence does not create additional consumers or increase evidence strength. The LeetCode scaffold remains afforded: its contradictory argument handling and unavailable grader prevent a wired standalone submission finding. Valid epistemic containers for task state, and their narrower memory parts, remain distinct rather than being merged.

The Python scope of RT-RTE-program-attempt and RT-RTE-program-variants is preserved. The Rust generation caller's required-argument omission blocks complete generation-to-successor progression; independently callable Rust validators still consume saved source. Missing UCS scheduling and the evaluated-node versus archived-candidate mismatch prevent attributing a solved archived row to the selected node.

Unresolved initial dataset/example provenance, resumed-content control, serialized library-object internals, historical trace instances and possible provider-accumulated state remain local limitations. Fixed pretrained model weights are not automatically included as use-dependent memory. Their known parametric form cannot fill the unknown provider-memory unit. No explicit manual content-admission witness or bounded whole-memory negative is available, so neither is invented. Requested new plans and diagnoses do not establish synthesize without recorded output semantics.

Source-read log: none. This mapping used only the supplied accepted members, reconciliation and contracts; it neither added evidence nor reopened report verification.