Learned Intuition

Five neural networks vote yes/no on every tool call — trained on real failures, rules fallback if models miss.

A pattern-matching layer between your agent and the rest of your system, trained on what actually went wrong before.

The AMYGDALA layer intercepts every tool call via before_tool_call and runs it through five Prudence ONNX networks (the public, downloadable family) plus an optional five Personality networks (private — train your own). The Prudence ensemble aggregates via weighted meta-vote; if any single model crosses the conservative override threshold (0.9) or model disagreement spikes, the cautious vote wins. Phase 1 is observe-only — the gate logs what it would have blocked but never stops anything. Phase 2+ enables active blocking. If ONNX models aren't present, a rule-based heuristic gate kicks in automatically so the plugin loads cleanly anywhere.

Install

openclaw plugins install @globalcaos/openclaw-learned-intuition

Optional — drop the public Prudence ONNX models into ~/src/tinkerclaw/models/amygdala/onnx/. Then enable in openclaw.json:

"plugins": {
  "allow": ["tinkerclaw-learned-intuition"],
  "entries": { "tinkerclaw-learned-intuition": { "enabled": true } }
}

Pairs Well With

• @globalcaos/openclaw-identity-persistence — failure-derived nudges feed SOUL.md re-injection. The agent doesn't just bounce off vetoes; it changes shape.
• @globalcaos/openclaw-total-recall — every veto becomes an indexed event. Patterns emerge across sessions instead of fading the moment context compacts.
• @globalcaos/openclaw-round-table — risky synthesis from a multi-model debate? AMYGDALA vetoes the ratification before the agent ships the action.

👉 https://github.com/globalcaos/tinkerclaw 👉 https://thetinkerzone.com

Clone it. Fork it. Break it. Make it yours.