Authoritative and experiential memory live in one graph. Document extraction and turn extraction both produce entity, attribute, and relation records. Provenance (source.kind, source.ref, source.span) records whether a fact came from a document or a turn.
How linking works
Same entity, multiple sources. When a user mentions “AirPods Pro” in chat and the product manual is uploaded as a document, extraction may create or update the same entity (for example Product/airpods_pro) with:
Turn-sourced attributes (
source.kind = "turn") - “purchased last week”Document-sourced attributes (
source.kind = "document") - “return_policy = 30 days”
The reconciler applies authority and calibration rules: cross-provenance conflicts emit uncertainty records instead of silent overwrites; same-provenance updates form supersession chains (valid_until on the prior assertion).
Unified recall. POST /api/v1/{ctx}/query ranks across facts and document passages in one router (structured lookup → response cache → hybrid retrieval → full-context fallback). Filter with include (facts vs passages) when you need a narrower slice.
Elaboration. Background elaboration can add relation edges between entities ingested from different sources but sharing context.
Operational implications
| Task | Approach |
|---|---|
| Ground an agent on manuals | Upload documents; recall with /query |
| Capture conversation | POST /facts or /facts/batch |
| Prefer curated policy over chat hearsay | Trust and confidence floors; higher default trust for documents |
| Debug a conflict | Inspect uncertainty records and entity history |
Manual entity maintenance
GET /api/v1/{context_id}/entities/{entity_type}/{entity_name}
DELETE /api/v1/{context_id}/entities/{entity_type}/{entity_name}Two streams in one store are described in Unified substrate and authority.