AgentFly memory profile

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

Comparison rationale

This comparison classifies local retained objects and their actual consumers, not all runtime storage. The accepted records establish both case banks, result/recovery records, training pairs, numerical checkpoint outputs and tool-local files. External provider retention is outside the memory boundary; opaque local binaries remain inside it. The missing checkpoint prevents inspecting exact content and deployment, not identifying the numerical form or the inspected update/load mechanisms (MEM-OBJ-checkpoint). No source files were read for this profile; all warrants come from supplied accepted records.

The reconciliation supersedes the epistemic training-entry declaration with MEM-OBJ-training-data, which supplies this profile's pair-artifact citations. The combined EPI-OBJ-case-entry remains a valid container; the two driver-specific bank records supply separate selector and schema findings. No supersession strengthens evidence or removes provenance uncertainty.

Cases give contextual advice through user-role examples. Static system prompts explain how to interpret those examples but are not accumulated instruction memory. The retained labels partition examples; retained text and numerical parameters determine ranking. Training pairs have learning authority at the optimizer, whereas trained parameters have ranking authority downstream. Result keys route and enforce restart skipping; rationale retention does not establish a rationale reader (RT-BAP-case-guidance, MEM-OBJ-query-record, RT-RTE-retriever-training). No validation authority is inferred from a current-run judge or trainer metric that is not itself retained and later consumed.

Requested query retrieval is pull, including its fulfilled model delivery. Startup/reload is a distinct automatic supply operation: available configured banks, checkpoint parameters and restart keys are loaded without selecting individual relevant items. Its coarse signal must not be conflated with the embeddings/classifier scores used on the later pull request. Requested filenames and configured checkpoint paths alone do not establish identifier-targeted push. Unspecified helpers and unread logs supply no additional resolved routes (MEM-RTE-embedding-examples, MEM-RTE-classifier-examples, MEM-RTE-retrieval-display).

The cases reduce completed run material to latest-plan/outcome examples; that is consolidation even though rich results remain. Checkpoint optimization and workspace replacement revise retained state. Append, reload, parsing filters and transient prompt formatting do not by themselves add deduplication, synthesis, decay or promotion. Formatting can omit evidence fields without introducing a new claim (EPI-RTE-case-format). Unknown command/converter transformations stay unresolved rather than receiving unsupported negatives.

Both case guidance and outcome-fed numerical training meet the profile's trace-learning test at wired strength. Starting training manually is not a per-update content decision. Linked benchmark question, task plan and outcome provide reduced trajectory experience; full tool histories and judge reasons are retained only in results and do not thereby become update inputs. Prior case genealogy remains unresolved, so trace-source coverage is partial. Neither these writes nor later loading establishes observed reliance, causal case utility, successful learning or scheduled live deployment (RT-RTE-case-admission, MEM-OBJ-training-data, RT-RTE-retriever-training).

All positive findings retain implementation-level wiring, not observed or causal evidence. Unknown initial control, opaque binaries, arbitrary tool derivations and missing consumer facts remain local units; a resolved witness does not eliminate them from the inventory.