Experience Memory Graph: One-Shot Error Correction for Agents
Merged summary
TL;DR — Experience Memory Graph (EMG) is a framework for correcting LLM agent failures by reformulating recovery as a graph-matching problem, enabling one-shot error correction without costly test-time trial-and-error loops.
- Converts failed and successful expert trajectories into directed action decision graphs, then extracts common subgraphs (successful workflows) and graph edit paths that specify corrective actions (add/delete/relabel) under given observations.
- Stores insights in a memory graph with intra-task nodes and cross-task edges, enabling generalization across scenarios rather than brittle, task-specific prompt-based reflection.
- At test time, retrieves relevant insights to guide the agent in a single, loop-free execution, avoiding the API/time costs of iterative reflection.
- Reported to outperform state-of-the-art reflection baselines on ALFWorld and ScienceWorld in success rate and average reward, with no test-time trial-and-error (specific numbers not provided in the abstract).
Sources (1)
Experience Memory Graph: One-Shot Error Correction for Agents
TL;DR — Experience Memory Graph (EMG) is a framework for correcting LLM agent failures by reformulating recovery as a graph-matching problem, enabling one-shot error correction without costly test-time trial-and-error loops.
- Converts failed and successful expert trajectories into directed action decision graphs, then extracts common subgraphs (successful workflows) and graph edit paths that specify corrective actions (add/delete/relabel) under given observations.
- Stores insights in a memory graph with intra-task nodes and cross-task edges, enabling generalization across scenarios rather than brittle, task-specific prompt-based reflection.
- At test time, retrieves relevant insights to guide the agent in a single, loop-free execution, avoiding the API/time costs of iterative reflection.
- Reported to outperform state-of-the-art reflection baselines on ALFWorld and ScienceWorld in success rate and average reward, with no test-time trial-and-error (specific numbers not provided in the abstract).