Graphiti memory profile

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

Comparison rationale

This comparison classifies accepted records, not fresh source evidence. All positive implementation findings are wired unless the unit explicitly identifies the afforded MCP assistant role. No observed or causal finding is available. Partial coverage is not weakened evidence: it names retained parts and alternatives whose inventory, admission, origin or consumer remains unresolved.

Canonical identities and retained parts

Reconciliation supersedes the epistemic episode, entity, community and saga containers with MEM-OBJ-episodic-node, MEM-OBJ-entity-node, MEM-OBJ-community-node and MEM-OBJ-saga-node. The profile cites these canonical containers. Epistemic attribute, summary and watermark parts remain separately usable; the broader edge/metadata/vector containers have different extents and were not superseded. Those retained parts supply the form, authority and lineage distinctions rather than an extra storage substrate.

Dense graph vectors are parametric encodings on graph storage, not graph-owned model weights. Inference-generated vectors support other-compiled lineage locally, while MEM-OBJ-graph-embeddings leaves initial caller vectors and community variants unresolved. Readable MCP/evaluation displays omit vectors and cannot stand in for the numerical search payload.

Direction and consumer force

A request-driven overall workflow can contain a push to a later consumer. RT-RTE-context-retrieval returns material explicitly requested by library/service callers: that is pull. In RT-RTE-conversation-loop, notebook code requests facts, then automatically selects and inserts retained user facts into the model's system message. The model did not request those facts: this downstream operation is push. This distinction does not invent an unsolicited Graphiti service notification. Similarly, a write request triggers automatic prior-episode/candidate supply to extraction, and a saga update supplies previous text and cursor-selected episodes to a summarizer. Requested API delivery is not counted again as push.

Push signals retain actual selected parts. User/product center identities affect graph-distance selection; prior episode UUIDs replace temporal selection; saga identity/membership and cursor select its next episode window. Budget/count/time filtering is coarse even alongside identity selection. The ingestion push retains lexical and embedding signals, but no inferred-judgment finding. EPI-RTE-fact-comparison and EPI-RTE-fact-disposition establish judgments consumed for identity reuse and temporal disposition after candidate retrieval, not a judgment selector choosing retained parts for downstream push delivery. Optional cross-encoder capability in RT-RTE-context-retrieval does not establish its use on this ingestion path. The unsupported judgment-signal finding is withdrawn; the wired semantic reconciliation mechanisms remain unchanged.

Authority is local to a consumer. Fact strings give the answering model knowledge; fact/name text and vectors give search ranking input; retained groups/cursors route selection; cleanup predicates and invoked date filters enforce exclusions. Automatically retained conversational experience gives the graph-update consumer learning input without granting factual truth. RT-BAP-memory-instructions affords stronger instruction force for retrieved preferences/procedures only at an external host following those rules. Host loading, tool-name mapping and compliance are not wired by the service. Static rules and response schemas are outside the accumulated-memory inventory, so their enforcement/validation is not attributed to retained graph content.

Curation and trace sources

Extraction of standing facts from retained episodes, summary reduction and saga folding support consolidation. Literal summary append and typed overlays/replacements support evolution; duplicate reuse supports deduplication. Model-generated descriptions do not by themselves prove synthesis of a claim absent from inputs. The semantic alternatives remain unresolved in EPI-RTE-fact-extraction and EPI-RTE-entity-summary, without negating the implemented reduction/update operations. Temporal invalidation retains history; physical deletion/clear forget selected content. Neither establishes complete retraction from summaries, checkpoint state or already delivered context; see MEM-CLM-withdrawal-scope.

The notebook's qualifying graph-learning chain is automatic trace admission, extraction into durable descriptions and later retrieval into model context. It can consume human/model conversation and tool-result material, including recirculated facts. Its user-name prefix cannot make all inputs human session turns. Session-logs and tool-traces therefore remain separate positive source findings. No event-stream value follows merely from timestamps or message-source typing, and no trajectories value follows merely from a multi-step tool loop. The evaluation route also automatically derives retained graph results from recorded sessions for later model judgment; this is not an automatic production-successor installation or improved capacity claim.

MEM-OBJ-episodic-node distinguishes single-path raw suppression from unchanged-content bulk writes. Bulk inputs can therefore remain available to later extraction and saga consumption even when single ingestion would suppress text. MEM-RTE-saga-summary establishes recurrence but retains bounded selection/truncation and cursor limits; no losslessness is inferred. Generic callers, other example origins, checkpoint internals and remaining backend/consumer variants prevent system-complete trace-source, admission and authority conclusions. No local or global trace-learning no is invented without bounded absence records.

Source-read log and checks

No frozen-source reread was needed. This member uses only the supplied accepted reports and reconciliation; it declares no records, quotes or new evidence. Structural validation checks the classification form and record references, not semantic completeness or observed operation.