How agents learn and remember winning strategies across tasks

LLM agents solve tasks step-by-step but forget useful strategies once each episode ends. Unified Context Evolution organizes agent experience into four knowledge types—Memory, Strategy, Workflow, Skill—stored in an evolving library that agents retrieve and refine through repeated use. The system learns which knowledge types need reinforcement and automatically strengthens weak areas. On standard benchmarks, it improves task success substantially (96.3% on ALFWorld, up from 75.4%), and learned libraries transfer to different agent models without retraining.