Experience Memory Graph: One-Shot Error Correction for Agents
EMG turns agent recovery into graph matching, then applies the fix in one execution.
The paper converts failed and expert trajectories into directed action decision graphs, compares them, and stores reusable correction paths in a memory graph. Those paths specify concrete edits such as adding, deleting, or relabeling actions under an observation. At test time, the agent retrieves the relevant memory instead of running reflection loops. The authors report stronger success rate and average reward than reflection baselines on ALFWorld and ScienceWorld, with no test-time trial-and-error. ArXiv · AI/CL/LG's note
The paper converts failed and expert trajectories into directed action decision graphs, compares them, and stores reusable correction paths in a memory graph. Those paths specify concrete edits such as adding, deleting, or relabeling actions under an observation. At test time, the agent retrieves the relevant memory instead of running reflection loops. The authors report stronger success rate and average reward than reflection baselines on ALFWorld and ScienceWorld, with no test-time trial-and-error. ArXiv · AI/CL/LG's note
score 5