Supermemory memory profile

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

Comparison rationale

The comparison separates service objects from their physical substrate. Remote exact-file documents have inspected API consumers, but the graph/vector engine is declared rather than inspected. Dense representations are not the graph display's layout coordinates. The readable fact/profile payload does not expose every retained relation or review field. See MEM-OBJ-document-content, MEM-OBJ-memory-state, MEM-CLM-graph-display and MEM-RTE-mcp-read.

Every axis retains unresolved included alternatives. A known unit means the named path or part can be classified, not that it operated. The engine's physical storage, initialization, embedding provenance, exact ranking and extra maintenance remain unresolved. Alternative adapters are not presumed Vercel-equivalent; MEM-CLM-adapter-variants bounds those findings. Known declared operations coexist with opaque implementation, rather than being removed or upgraded by it.

Trace learning has a supported claimed session route: inspected capture and later context delivery surround declared automatic document-to-fact/profile production. The derivative is durable behavior-shaping context; new knowledge, parameter training and observed benefit are unnecessary for this classification. The same retained experience supplies learning authority at extraction, while its output supplies knowledge at the answering model. LiveKit callbacks do not change the original conversational content into an event-stream source. Arbitrary model-authored files and connector documents do not establish a trace source without recorded input provenance. See MEM-RTE-engine-dreaming, MEM-RTE-livekit-session, MEM-RTE-claude-file and MEM-RTE-connector-sync. No evidenced-absence record supports a local trace-learning negative; the derived system union supports only the claimed positive, not complete coverage.

Requested tool and MCP prompt returns are pull. Automatic middleware insertion is push downstream of an API pull. The push selector's namespace/profile and cache-key identity matches are distinct from unresolved remote semantic ranking. An embedding's existence does not establish which push selector uses it. See RT-RTE-vercel-call, RT-OBJ-memory-cache and RT-RTE-mcp-context.

Curation evidence is operation-specific. Local normalized deduplication and exact-file editing are wired; remote consolidation, conjectural synthesis, expiry, update, invalidation and confirmation promotion remain claimed. Review confirmation raises product status but does not establish evidence-consuming truth acceptance. Retained capability checks establish validation/enforcement only at upload admission, not content truth. See EPI-RTE-memory-format, MEM-RTE-inference-review, EPI-CLM-review-facts and RT-RTE-upload-session.

Supersessions affect the file and review citations: the canonical object/route are MEM-OBJ-claude-file, MEM-RTE-claude-file and MEM-RTE-inference-review, not their superseded epistemic counterparts. The broader epistemic graph object and narrower graph transformation remain valid containers/parts, not replacements for the memory objects or complete dreaming chain. This preserves arbitrary-file-content and review-acceptance limits without allocating records.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no source reread, added evidence, target execution or report revision occurred.