SAGE memory profile

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

Comparison rationale

The profile classifies accepted records, not new source facts. Reconciliation establishes no supersessions: containers, their numerical access parts, expert products and separately isolated route stages remain canonical. The axes preserve these scopes rather than merging every similarly named function list.

The internal training findings are wired capacities, conditional on reaching the AppWorld loader. The shipped two-node configuration instead selects an incompatible Parquet reader; it does not establish launch-to-learning operation. This qualification applies to paired reuse and RL trace learning alike (RT-RTE-paired-rollout, RT-RTE-policy-update). No observed or causal basis is asserted.

Code extraction is consolidation of retained action material even though it can omit bindings or damage a multiline definition. Such faithfulness defects do not establish synthesis. Same-name replacement supports local evolution, but key collisions and first-name delivery filtering do not establish near-duplicate retained-content curation. No dedup finding is emitted. Appended evaluation records do not guarantee withdrawal of older entries. Function source is symbolic at the interpreter and interpreted reference/guidance at the model. Numerical access tensors are parametric representations on files/in-memory substrates, not evidence of a vector database (EPI-RTE-function-extraction, RT-OBJ-skill-library-functions, MEM-OBJ-skill-embeddings).

Learning authority belongs to trace-consuming update routes; downstream code advice and execution are different authority paths. Trace learning therefore includes extracted reusable code as well as parameters. It asserts no improved competence. The original conversational, execution and trajectory contents remain distinct from imported demonstrations and task text (RT-RTE-skill-install, MEM-OBJ-rollout-functions, MEM-OBJ-expert-dataset, RT-RTE-sft-admission).

Initialization supplies code without a model retrieval request. Paired/scenario whole-library availability is coarse selection; evaluation's actual task-ID match is identifier selection; alternative vector and query-text scores support embedding and lexical selection. The cache instead fulfills a requested completion, so its argument matching is not classified as push targeting (RT-RTE-evaluation-loop, MEM-OBJ-query-list, MEM-RTE-response-cache).

Every axis remains partial. Unknown predefined preparation, initial weight provenance, provider retention and unread alternate backend behavior cannot inherit a known route's classifications. No absence record supports a blanket trace-learning negative. Predefined load versus output write, missing query-list serialization, and operator selection of the SFT output prevent stronger restart or automatic pipeline conclusions, without erasing wired subsequent transformations (RT-OBJ-skill-library-functions, MEM-OBJ-query-embeddings, RT-RTE-sft-admission). Static context and the API-prediction/retry branch do not independently qualify as accumulated memory (MEM-RTE-api-context, EPI-RTE-code-retrial).

Operator selection of an existing SFT model path is retained as a requested load, not manual write agency. The accepted records establish neither per-content human approval nor a separate replacement/admission mechanism at that selection edge (RT-RTE-sft-admission, RT-CMP-sft-model).

Recovery comparisons remain incomplete: latest-iteration tracking and saved dataloader state have resume-routing support, but the storage, form and lineage inventories do not separately enumerate all non-parameter recovery parts. The affected axes remain partial; their positive parameter findings do not license complete per-part recovery statistics (RT-RTE-policy-update).

Source-read log: none. This classification uses only the supplied accepted members and their annotations; no frozen-source ambiguity required a source read.