Retrieval, graph traversal, and hints

Recall builds a bounded candidate pool from full-text search, embeddings when available, prospective hints when enabled, structured paths, and graph traversal. The graph is a candidate shaper, not a replacement for evidence or ordinary search.

Before content is loaded, reranked, summarized, or access-tracked, candidate IDs are authorized against the caller’s agent, visibility, project, and scope. A broad vector or traversal result is not permission to read a memory.

Structured knowledge

The current graph is a scoped ontology: entities, aspects, grouped claim slots, versioned attributes, and directed dependencies. Dreaming is its only automatic semantic writer. Explicit ontology operations can also apply or propose audited changes.

The retired relations extraction path is not the current semantic-authoring model. Do not treat relation counters or compatibility tables as current graph truth.

Hints and ranking

Optional prospective hints are alternate query phrases associated with an existing memory. They can rescue vocabulary mismatches, but they remain a derived retrieval signal. They do not create semantic facts.

After authorization, recall may apply structured evidence shaping, reranking, dampening, currentness annotation, and bounded supplemental source results. These stages are best-effort: a failed secondary channel degrades toward simpler safe retrieval rather than failing the whole request.

For the ontology model, see Knowledge architecture. For the public request and response contract, see Memory and recall and API reference.