OpenClaw-native memory with shadow learning

March 2026 | Jonathan Gu

v12.2.6 is the point where the site needs to stop sounding like a generic memory paper and start describing the actual product direction: OpenClawBrain is a memory and routing layer for OpenClaw that becomes useful quickly and keeps learning in the background.

OpenClaw -> activate + compile -> learned route_fn -> bounded compiled context -> answer

historical + ongoing sessions
-> scanner/harvester labels
-> async teacher labels
-> Ultimate Policy Gradient
-> better runtime route_fn later

What matters in practice

• use the brain immediately after `init` and `serve`
• keep the async teacher off the hot path
• train continuously from historical and ongoing OpenClaw work
• treat the learned runtime route_fn as the actual product surface

Proof boundary

The mechanism can be reproduced locally. Strong product claims still need recorded-session eval, shadow traffic, and narrow online rollout. The current site should state that directly.

For the full current framing, read the v12.2.6+ series, keep the claim boundary in /proof/, and use /paper/ for the canonical paper route and direct PDF.