Experience Memory Graph Improves Agent Error Correction.

Researchers have developed Experience Memory Graph (EMG), a novel framework designed to significantly improve the error correction capabilities of Large Language Model (LLM) agents. Traditional self-correction mechanisms, often relying on prompt-based reflection, are frequently brittle, costly, and produce task-specific memory that struggles to generalize. EMG addresses these limitations by reframing agent failure recovery as a graph matching problem. During a training phase, both failed exploration trajectories and successful expert trajectories are converted into directed action decision graphs. By comparing these graphs, EMG identifies common successful workflows and explicit graph edit paths that detail how to correct failures, such as which actions to add, delete, or relabel under specific observations. This knowledge is then stored in a memory graph. At test time, EMG efficiently retrieves relevant insights from this memory graph, guiding the agent in a single, loop-free execution to correct errors. Experiments on complex environments like ALFWorld and ScienceWorld demonstrate that EMG consistently surpasses state-of-the-art reflection baselines in success rate and average reward, all while eliminating the need for iterative trial-and-error during inference.