Skip to content

Mental model

Unified substrate and authority

SurrealDB Agent Memory stores authoritative and experiential knowledge in one SurrealDB graph. The eight pillars, especially Authoritative and Experiential, show up as records and edges, not as two databases.

One multi-model SurrealDB substrate - documents, turns, entities, attributes, relations, embeddings, traces, and (where enabled) geometry - with provenance explaining which stream produced a record and reconciliation deciding how streams combine.

"Layers" are still a useful picture for authority: curated organisational truth versus conversational input. Both streams are records and edges in the same database, updated under ACID transactions.

StreamTypical source.kindWhat it holdsDefault trust
Authoritativedocument (and operator upsert)Manuals, policies, product data, repos, structured exportsHigh - vetted sources
Experientialturn, plus reflect, elaboration, consolidation, …What users and agents said; synthesised or background-minted factsLower - must earn promotion

Retrieval, elaboration, and consolidation see one entity / relation graph. Hybrid rankers fuse vectors, BM25, graph structure, keyword bridges, geo predicates, and trace-derived features without cross-store joins. A question like "semantically close to X, mentioning Y, linked to entity Z, valid as of last March" is one query over one engine.

When an experiential assertion disagrees with authoritative material, the reconciler:

  1. Does not silently overwrite curated records.

  2. Records the experiential assertion with provenance intact.

  3. Surfaces uncertainty (and/or conflict metadata) so you can see the gap.

That is how the Authoritative pillar wins over casual assertions: authority is a reconciliation policy, not a second copy of the universe hidden in another engine.

  • Transactions across “what the user said” and “what the handbook says” can be reasoned about together.

  • Contradictions become first-class records you can query, not cosine-distance accidents.

  • Traces link decisions back to the sources considered.

Was this page helpful?