How Roux works
Data flow, not magic.
Roux treats memory as a structured data problem. When a session ends, an episode detector classifies the interaction into typed events — facts learned, relationships formed, significant moments. These aren't embeddings of raw text; they're structured, typed records.
Storage is deliberately heterogeneous. Factual memories live in Postgres with pgvector for semantic retrieval. Relationships — who knows whom, what was promised, how trust has evolved — live in Neo4j, where graph traversal outperforms any embedding similarity search.
The retrieval layer composes context from both sources before each session. Context assembly is deterministic, auditable, and bounded — no surprise token bloat.
Episode detection is the design choice that matters most. Instead of summarizing everything, Roux decides what's worth remembering — a fundamentally different question that keeps storage costs bounded as usage scales.