# Architecture Board

## Goal
Build a local-first Openclaw agent that becomes more intelligent over time through:
- typed memory
- typed tool graph
- trajectory logging
- reward signals
- shadow meta-controller
- offline policy learning
- sacred eval gates

## Hosts
- Mac Studio: hot-path inference
- openclaw: orchestration and live logging
- Unraid: offline learning, retrieval, replay, eval batch jobs
- Kimi: offline teacher only

## Phase A
1. Typed trajectory schema
2. Reward signals
3. Replay buffer + policy stats
4. Tool graph
5. Uncertainty model
6. Shadow meta-controller