Summary

GBT (Gated Behavior Tree) is an OpenClaw plugin that transforms execution logs into a reusable "experience tree," significantly reducing agent token costs and improving efficiency. It targets OpenClaw users facing repetitive tasks, enabling agents to leverage past successes and failures, thus avoiding redundant planning and reasoning. The core benefit is a more economical and personalized agent experience through persistent storage and refinement of learned behaviors.

How It Works

Based on "Traversal-as-Policy" research, GBT distills completed OpenClaw runs into reusable macro nodes within a persistent experience tree, storing both successful and failed trajectories. For recognized "covered" tasks, GBT routes them to a cheaper executor, guided by learned experience. Failures trigger a self-evolution process: during idle periods, GBT replays failed tasks, identifies and repairs broken steps, verifies fixes against real data, and updates the tree, creating a self-improving system.

Quick Start & Requirements

• Prerequisites: OpenClaw, Node.js, Python 3.10+, openai Python package, configured OpenClaw model, and OpenAI authentication (for LLM analysis).
• Installation: Install Node.js dependencies (npm install, npm run build), then install the plugin into OpenClaw (openclaw plugins install . or tarball) and enable it (openclaw plugins enable gbt-skill).
• Configuration: Requires plugins.entries.gbt-skill.config in OpenClaw, specifying pythonExecutable, stateSubdir, cheaperModel, coverageThreshold, and idleMinutes.
• Docs: Research paper: https://arxiv.org/abs/2603.05517.

Highlighted Details

• Distills runs into reusable, non-task-specific macro nodes.
• Stores both success and failure paths for repair and reuse.
• Switches covered tasks to a cheaper executor model based on coverageThreshold.
• Implements self-evolution: replays, repairs, and verifies failed trajectories during idle time.
• Provides step-local guidance for known tasks.
• Facilitates long-term personalization via user-specific experience trees.

Maintenance & Community

The README provides no details on maintainers, community channels (e.g., Discord, Slack), or roadmap. The project is grounded in academic research.

Licensing & Compatibility

The README does not specify a software license. This omission may hinder adoption, particularly for commercial or closed-source applications.

Limitations & Caveats

The current release intentionally omits the safety-gate mechanism from the research paper. System effectiveness depends on OpenClaw's runtime capabilities (tool access, workspace). Self-evolution replay requires proper authentication, and complex repairs may need multiple attempts.