v3 — the Digital Brain
These pages document the v3 cognitive runtime. The features below are implemented and wired at server boot in v3-alpha — each page is the design plus its current behaviour.
Shipped in v3-alpha (Beta). Everything on these pages is implemented and wired at server boot. Interfaces may still change before 3.0 stable. See Roadmap for the milestone breakdown.
How v3 reshapes the memory model
v2 modeled memory as typed knowledge stored in PostgreSQL. v3 keeps that foundation and adds three layers around it:
- Ephemeral state (Redis L0/L1) — working memory and active sessions
- Procedural memory — strategies and evolved prompts, learned from outcomes
- Inferred memory — composed and anticipated knowledge, no LLM needed
The result is an agent runtime that remembers, learns, and anticipates — not just stores and retrieves.
Concept index
Memory model expansion (8)
- Session memory + collapse-to-recall
- Working memory (Redis L0 layer)
- CoT distillation engine
- Self-editing memory
- Procedural memory + prompt evolution
- Hypergraph relations
- Delta-encoded episodes
- Tasks & events as primitives
Zero-LLM intelligence (4)
- Predictive memory (anticipation engine)
- Meta-cognition (self-model)
- Memory compositionality (inference engine)
- Trust-anchor consensus