Graphiti profile verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

I checked all ten axes of the memory profile, their unit scopes, coverage assessments, positive findings and evidence bases against the handed records. Reconciliation resolves the superseded epistemic episode, entity, community and saga containers to the memory containers; attribute, summary, watermark and edge parts remain separately usable. No frozen-source read was needed. This judgment adds neither records nor evidence.

  • Storage substrate: Graph properties/relationships, local evaluation exports and the requested in-process checkpoint support graph, files and in-memory findings. Remaining backend/consumer alternatives and checkpoint internals prevent exhaustive coverage rather than negating these witnesses. See MEM-OBJ-graph-metadata, MEM-OBJ-graph-results and MEM-OBJ-memory-saver.
  • Representational form: Readable descriptions, fixed-rule IDs/dates and dense retained vectors support natural-language, symbolic and parametric findings on their consumption paths. Vectors are not provider weights, and readable displays do not classify undisplayed payloads. The mixed-field and checkpoint/export units retain supported positives but coarse coverage, as stated under Limits. See MEM-OBJ-graph-embeddings, EPI-OBJ-entity-attributes and MEM-OBJ-memory-saver.
  • Lineage: Caller acquisition supports imported lineage; conversation-derived graph content supports trace-extracted lineage; summary, metadata and embedding derivations support other-compiled lineage on the recorded paths. Known derivation does not resolve original authorship, supplied vectors or untraced alternatives. See MEM-OBJ-episodic-node, MEM-OBJ-graph-embeddings and EPI-RTE-entity-summary.
  • Behavioral authority: Text supplies knowledge context, text/vectors affect ranking, retained selectors route scope/windows, and invoked cleanup/date predicates enforce exclusions. Retained experience at the graph-update consumer has learning force separately from output facts' knowledge force. Retrieved procedures have afforded instruction force, not wired host compliance. See RT-BAP-facts-context, RT-BAP-memory-instructions, MEM-RTE-saga-summary and EPI-RTE-episode-removal.
  • Write agency: Software/model admission, notebook post-response scheduling and evaluation export support automatic findings. Generic caller identity does not establish a human per-content decision; checkpoint admission internals remain uninspected. Starting an operation is not manual admission of its derived results. See RT-RTE-episode-ingestion, RT-RTE-conversation-loop and RT-RTE-evaluation-comparison.
  • Curation operations: Duplicate reuse supports dedup; standing-fact extraction and summary reduction support consolidate; existing-entry updates support evolve; temporal withdrawal with history supports invalidate; physical removal supports decay without implying a timed schedule. Missing output comparisons leave semantic novelty unresolved rather than supporting synthesize. See EPI-RTE-fact-comparison, EPI-RTE-fact-disposition, EPI-RTE-entity-summary and MEM-CLM-withdrawal-scope.
  • Read-back direction: Requested graph/file returns support pull; automatic retained-context supply to a later model, summarizer or judge supports push within the same chain. Request-driven workflow descriptions do not rule out that downstream push. Requested API returns are not counted again as push. See RT-RTE-context-retrieval, RT-RTE-conversation-loop, MEM-RTE-saga-summary and RT-RTE-evaluation-comparison.
  • Read-back signal: The named triggers, selector inputs and delivered parts support local lexical, embedding, identifier and coarse findings. Prior episode UUIDs and saga membership target retained inputs; time/count/budget filters remain separate coarse operations. The corrected graph-candidate unit no longer emits inferred-judgment. Its partial note and dependent rationale distinguish candidate delivery from subsequent semantic reconciliation. EPI-RTE-fact-comparison and EPI-RTE-fact-disposition do not establish judgment-selected push delivery; optional cross-encoder capability in RT-RTE-context-retrieval does not establish use on that ingestion path. The previous blocker is repaired, not merely caveated.
  • Trace learning: Automatic conversation-fed graph production creates durable descriptions for later model context, supporting wired yes. Session-fed evaluation graph production also has a later judge consumer; production-successor installation is not required. Generic trace origins, checkpoint-only transformation and remaining routes stay unresolved, without invented no findings. See RT-RTE-conversation-loop, RT-RTE-episode-ingestion and MEM-OBJ-graph-results.
  • Trace source: Human/model conversational content and post-tool results support session-logs and tool-traces separately on the notebook route; dataset sessions support session-logs on evaluation. Adapter prefixes and timestamps do not establish event-streams, and a tool loop alone does not establish trajectories. Generic original inputs remain unresolved. See MEM-OBJ-episodic-node and RT-RTE-evaluation-comparison.

All axes remain partial. Included alternatives are retained rather than excluded to manufacture complete coverage. The profile preserves mixed notebook origins, single-versus-bulk raw retention, asynchronous availability and wiring versus observed benefit. These faithfully scoped uncertainties are not additional verification findings. The handed reports contain no unresolved cross-report conflict affecting the classifications. No unsupported emitted value, system-wide absence or complete inventory assertion remains identified.

Blockers

none

Limits

  • The representational-form units “Caller JSON episodes and schema-dependent attributes” and “Notebook checkpoint boundary and evaluation exports” combine operative parts/paths with different coverage rather than assessing each separately. MEM-OBJ-episodic-node, EPI-OBJ-entity-attributes, MEM-OBJ-memory-saver and MEM-OBJ-graph-results support the textual and symbolic positives, but the combined units cannot establish separate coverage for individual schema fields, checkpoint payloads and export associations. The profile already marks these units and their axis partial. The supported form union remains valid; consumers should withhold part-level form/coverage comparisons between those bundled alternatives.