ACE memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The profile classifies retained outputs and their later consumers, not the whole runtime. MEM-OBJ-playbook-content, MEM-OBJ-bullet-metadata, MEM-OBJ-best-playbook and MEM-OBJ-call-logs supply the local retained inventory. Model components expose no comparable inventory of provider retention. That included uncertainty remains a separate unit on every axis; it is not evidence for a provider memory store. Unavailable AppWorld implementation is outside this main-repository classification, so this is not a whole-system AppWorld profile.
Automatic trace learning is wired through response/feedback → last reflection → admitted advice → later generator, and through response-conditioned tags → retained counters → later prompt consumers. RT-RTE-sample-adaptation, RT-RTE-batch-adaptation and MEM-RTE-counter-update support these routes without establishing benefit. The original linked reasoning, prediction and correctness feedback supports trajectories; neither role names nor raw JSON logging establish additional source values. External task/oracle provenance remains unresolved beside that positive. Log retention alone does not establish an additional learning route.
The whole-playbook generator route is coarse push. Its model-internal relevance interpretation does not establish another recorded selector. The reflector route is identifier push because it actually matches reported IDs to retained lines. Analyzer maintenance separately selects retained groups with embeddings. Cross-run file return is pull; automatic delivery of the returned text to a later role is push. These are distinct operations within one chain, not competing system labels (RT-BAP-playbook-guidance, RT-RTE-sample-adaptation, RT-RTE-bulletpoint-merge, RT-RTE-playbook-reload).
Analyzer deduplication and replacement support dedup and evolve, including destructive fallback. Consolidation is only claimed for its requested concise combination: the records do not establish meaning preservation. Likewise new-insight instructions do not prove synthesis of a claim absent from the inputs. Selection of a best snapshot promotes a complete retained artifact for final testing; it does not promote every bullet or roll back current training state (RT-RTE-bulletpoint-merge, EPI-RTE-curator-admission, RT-RTE-evaluate-select).
Reconciliation supersedes EPI-OBJ-reflection with MEM-OBJ-reflection-content, EPI-OBJ-playbook with MEM-OBJ-playbook-content, and EPI-RTE-counter-update with MEM-RTE-counter-update. These identity resolutions affect the trace, lineage, authority and admission citations; they do not upgrade evidence strength. The combined epistemic tags object remains distinct from persisted metadata, so this profile uses the latter for retained-field classification. Last-tag-per-ID admission is not a tally of all batch judgments.
Every axis retains unresolved included units. Supported unions therefore include only local positive findings at their stated bases, never complete-value conclusions. There is no warranted local trace-learning no: the set declares no bounded absence record for these alternatives. Missing initial authorship prevents manual/authored classification; missing actual generated lessons prevents stronger semantic curation conclusions; missing remote retention and consumers prevents a complete inventory. No observed operation or causal improvement is asserted.
Source-read log: none. This classification used only the supplied accepted members, reconciliation, types and record contract; it added no source evidence or records.