New framework helps LLM agents learn from experience and adapt faster to new environments, boosting performance by nearly 8%

Experiential Reflective Learning (ERL) enables LLM agents to improve over time by reflecting on past task outcomes and generating reusable heuristics

ERL retrieves relevant lessons from previous interactions and injects them into the agent's context to guide execution on new tasks

Achieves 7.8% improvement in success rate over ReAct baseline on the Gaia2 benchmark with better task completion reliability

Addresses a key limitation of current autonomous agents: they typically approach each task from scratch without leveraging accumulated experience