AriGraph memory profile

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

Comparison rationale

This profile classifies retained objects and actual consumption paths, not the whole runtime's permissions or every model component. All ten axes are partial because included alternatives remain unresolved. Local known units retain positive implementation evidence without upgrading opaque alternatives. The basis is wired for shipped local routes and afforded for alternate library mechanisms. No finding upgrades delivery to activation or attributed scores to causal benefit.

Inventory and reconciliation limits

The reconciliation supersedes the epistemic triplet, episode-text and summary declarations with MEM-OBJ-triplets, MEM-OBJ-obs-episodic and MEM-OBJ-summary. Those canonical objects carry the corresponding values here. The Hypergraph thesis/event container is not superseded; its content and index parts are classified separately using MEM-OBJ-hypergraph-content and MEM-OBJ-hypergraph-index.

The unresolved inherited Hypergraph spatial-store omission is retained as an included unresolved unit on every axis. RT-RTE-hypergraph-update supports movement-derived insertion but explicitly leaves its relation to the other state unassessed. This profile neither invents the missing object record nor assigns that alternate store to the main ContrieverGraph object. A complete inventory needs the declaring analysts' additional record and consumer account. The omission can be stated faithfully as a limit here; it prevents every axis from becoming system-complete.

MEM-RTE-observed-update leaves generic caller admission, initial provenance and downstream consumption unresolved. MEM-OBJ-logs distinguishes known text logging from arbitrary caller JSON. Storage findings separate wired append-only text logging from afforded replacement JSON writes; the unused JSON capability cannot inherit the shipped text branch's basis, and displayed text cannot classify its unknown serialized payload. MEM-ABS-state-readers bounds the lack of shipped snapshot replay and accessor calls, not all external consumption. Provider internals remain uninspected under RT-CMP-gpt-agent and RT-CMP-llama-agent; this profile does not assert an experiential provider store exists.

Mapping choices

Process-local retention can shape later invocations without surviving process restart. Dense vectors are parametric representational form, not evidence of a vector-service substrate or model-weight training: MEM-OBJ-graph-embeddings. Their local derivation is known while upstream checkpoint provenance remains limited by RT-CMP-retriever.

Game facts and static QA paragraph extraction have different origins. Game-derived facts are trace-extracted; dataset text is imported and its extraction is other-compiled. MEM-OBJ-triplets also records deterministic spatial compilation, which does not prove the inferred relation is true. The indeterminate semantic edges in EPI-RTE-triplet-parsing and the extraction account do not justify a synthesis finding. Summary consolidation describes the implemented reduced-memory transformation, not guaranteed faithful meaning preservation: RT-RTE-summary-cycle.

Authority belongs to each actual consumer. Old retained facts and prior summaries feed automatic memory updates (learning), selected text supplies decision context (knowledge), plan priorities guide later roles (instruction), vectors/text/importance influence relative selection (ranking), and spatial edges select navigation (routing). These paths are separately warranted by RT-BAP-memory-decisions, RT-BAP-plan-decisions, RT-RTE-graph-navigation and MEM-OBJ-rag-metadata. The recent exclusion list has an exact-match veto, but its inventory-augmented strings need not match pre-inventory episode keys: MEM-OBJ-history, MEM-OBJ-obs-episodic. No validation authority is inferred from static parser criteria or diagnostic answer metrics.

Requested graph retrieval is pull, including its internal automatic selectors. Separately, drivers push histories, current summaries/plans and selected smart-RAG entries into model requests. RT-RTE-smart-rag supports embedding and retained model-importance selectors plus coarse recency/recent-entry selection; dictionary/hash identities alone establish no identifier push. Unknown external delivery cannot become inapplicable merely because an inspected API returns requested content.

Qualifying automatic game fact, summary, plan and smart-RAG metadata writes retain behavior-shaping outputs for later consumption. This is trace learning under the comparison contract, not demonstrated improvement. The source is environmental events or linked game experience, not session conversation merely because the model receives messages. The inspected QA route instead extracts static paragraphs, and full history only appends raw traces: RT-RTE-qa-cycle, RT-RTE-full-history. Neither route has a canonical bounded-absence record warranting local no; both remain not-determinable without cancelling other routes' yes. Library APIs without recorded original trace content and later behavioral consumers remain unresolved.

Source reads and checks

No source checkout reads were performed. Classification uses only the supplied records and reconciliation; no records, quotations or evidence were added. Deterministic validation is reported by the required invocation check, not taken as evidence of analytical completeness.