Self-Direction Method Distills Agent Error Reflection into Lightweight LoRA Adapters

Self-Direction enables agents to convert task failures into persistent lessons by distilling trajectory reflections into low-rank adapters. This approach uses only ~0.02% of base model parameters per adapter, preventing catastrophic forgetting while avoiding the high cost of full-parameter fine-tuning.