Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

Experience Memory Graph: One-Shot Error Correction for Agents

Source

arXiv

arxiv.org/abs/2607.13884v1

Read the source

Source

Plain-Language Brief

Researchers propose a new framework called Experience Memory Graph (EMG) to help large language model agents recover from failures more effectively in complex, long-horizon tasks. Instead of using prompt-based reflection that can be brittle and costly, EMG reformulates failure recovery as a graph matching problem. During training, successful exploration trajectories and expert trajectories are converted into directed action decision graphs. These graphs are then matched to extract common subgraphs (successful workflows) and graph edit paths that indicate how to correct failures. At test time, EMG retrieves relevant insights and guides the agent in a single execution without requiring any test-time trial-and-error.

What the Document Says

  • LLM agents generate sequential trajectories of states, actions, and observations for autonomous decision-making. However, they often suffer from compounding errors and struggle to recover from failures in complex tasks.
  • Existing self-correction mechanisms rely on prompt-based reflection but are brittle, costly due to iterative trial-and-error loops, and produce task-specific memory that may not generalize well.
  • The authors propose Experience Memory Graph (EMG), a framework that reformulates agent failure recovery as a graph matching problem.
  • At training time, EMG converts both failed exploration trajectories and successful expert trajectories into directed action decision graphs.
  • By matching these graphs, EMG extracts common subgraphs (successful workflows) and graph edit paths that indicate how to correct failures.
  • These extracted insights are stored in a memory graph with intra-task nodes and cross-task edges.
  • At test time, EMG retrieves relevant insights and guides the agent in a single execution without requiring any test-time trial-and-error.
  • Experiments on ALFWorld and ScienceWorld show that EMG outperforms state-of-the-art reflection baselines in success rate and average reward.

Key Concepts Mentioned

  • Large Language Model (LLM) agents
  • failure recovery
  • graph matching problem
  • Experience Memory Graph (EMG)
  • directed action decision graphs
  • common subgraphs
  • graph edit paths
  • success rate
  • average reward

Claims Made by the Document

  • LLM agents struggle with compounding errors and failure recovery in complex tasks.
  • Existing self-correction mechanisms have limitations.
  • EMG reformulates agent failure recovery as a graph matching problem.
  • EMG outperforms state-of-the-art reflection baselines in success rate and average reward.

Evidence Used

  • The text describes the difficulties LLM agents face in complex tasks and mentions existing self-correction mechanisms have limitations.
  • The authors present their proposed solution, EMG, and explain its functionality in detail.
  • Experiments on ALFWorld and ScienceWorld are mentioned as evidence of EMG's superior performance.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Large Language Model (LLM) agents have shown remarkable capabilities in autonomous decision-making by generating sequential trajectories of states, actions, and observations. However, in complex, long-horizon tasks, these agents frequently suffer from compounding errors and struggle to recover from failures. Existing self-correction mechanisms rely on prompt-based reflection, which is inherently brittle, incurs heavy time and API costs due to iterative trial-and-error loops, and produces task-specific memory that may be hard to generalize to new scenarios. To address this, we propose Experience Memory Graph (EMG), a framework that reformulates agent failure recovery as a graph matching problem. At training time, we convert both failed exploration trajectories and successful expert trajectories into directed action decision graphs. By matching these graphs, we extract common subgraphs (successful workflows) and graph edit paths that explicitly indicate how to correct failures (e.g., which actions to add, delete, or relabel under a given observation), and store them in a memory graph with intra-task nodes and cross-task edges. At test time, EMG retrieves relevant insights and guides the agent in a single, loop-free execution. Experiments on ALFWorld and ScienceWorld show that EMG consistently outperforms state-of-the-art reflection baselines in success rate and average reward, while requiring no test-time trial-and-error.