Don't Forget What You Just Learned: Self-Directed Repair Turns Agent Mistakes Into Lasting Lessons
Summary
- Researchers introduce Self-Direction, a mechanism allowing AI agents to learn from mistakes by combining in-context retries and lightweight low-rank adapters.
- The method freezes base model weights and trains additional parameters equivalent to about 0.02% of the base model size for each adapter.
- Adapters are saved in a persistent library and automatically routed to solve related problems in future tasks.
- Experiments using Qwen-family models across GAIA, Claw-Eval, and Terminal-Bench 2.1 improved success rates by 8.9 to 17.0 percentage points.
Why it matters
This approach offers a cost-effective alternative to full fine-tuning, allowing AI agents to permanently retain lessons learned from mistakes without overwriting base capabilities or exceeding context limits.