Knowledge
Persistent entity-relationship graph, five-tier memory hierarchy, and semantic search — maintaining institutional knowledge across the entire agent workforce.
Memory Hierarchy
Entities (person, project, concept, tool, location, org) with typed relationships. 5-hop traversal. Cluster detection. God-node identification for key entities.
pgvector embeddings with IVFFlat indexing. Dense Passage Retrieval (DPR) for precision recall. BM25 full-text as fallback for exact term matching.
Dreamer process synthesises long-term patterns from session data during idle cycles. Identifies recurring themes and promotes them to permanent memory.
Query decomposition into parallel sub-queries. 3 modes: Speed, Balanced, Quality. Source deduplication, content extraction, LLM synthesis with citations.
Tiered access with clearance levels. PII vault for sensitive information. Classification levels control which agents can read what.
Relevant memories automatically surfaced during agent conversations. No manual retrieval needed — the system knows what context is relevant.