# Memp: Exploring Agent Procedural Memory

**Authors:** Runnan Fang, Yuan Liang, Xiaobin Wang, Jialong Wu, Shuofei Qiao, Pengjun Xie, Fei Huang, Huajun Chen, Ningyu Zhang  
**arXiv:** [2508.06433](https://arxiv.org/abs/2508.06433)  
**Date:** 2025-08-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.06433v1/

## Key points

- Authors address limitations of Large Language Models (LLMs) in handling procedural memory
- Proposal of Memp as a solution to provide agents with learnable, updatable, and lifelong procedural memory
- Memp distills past agent trajectories into detailed step-by-step instructions and higher-level script-like abstractions
- Strategies for Build, Retrieval, and Update processes of procedural memory discussed to enhance effectiveness
- Continuous updating and correction of outdated information in sync with new experiences allows Memp to evolve over time
- Empirical evaluations on TravelPlanner and ALFWorld show higher success rates and improved efficiency with refined memory repository
- Transferring procedural memory from a robust model to a weaker one leads to significant performance gains
- Importance of adaptive and evolving procedural memory mechanisms highlighted for enhancing AI agent performance based on Large Language Models

## Summary

In their work titled "Memp: Exploring Agent Procedural Memory," authors Runnan Fang, Yuan Liang, Xiaobin Wang, Jialong Wu, Shuofei Qiao, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang address the limitations of Large Language Models (LLMs) based agents in handling procedural memory. These agents excel at various tasks but struggle with rigidly defined procedural memory that is either manually crafted or constrained by fixed parameters. To overcome this challenge, the authors propose Memp as a solution to endow agents with a learnable, updatable, and lifelong procedural memory. Memp operates by distilling past agent trajectories into detailed step-by-step instructions as well as higher-level script-like abstractions. The study delves into different strategies for the Build, Retrieval, and Update processes of procedural memory to enhance its effectiveness. By implementing a dynamic regimen that continuously updates, corrects, and removes outdated information from the repository in sync with new experiences, Memp evolves over time. Empirical evaluations conducted on TravelPlanner and ALFWorld demonstrate that refining the memory repository leads to progressively higher success rates and improved efficiency in performing similar tasks. Notably, transferring procedural memory from a more robust model to a weaker one results in significant performance gains. This highlights the value of a well-constructed procedural memory system in enhancing an agent's capabilities across various domains. Overall,the findings presented in this study shed light on the importance of adaptive and evolving procedural memory mechanisms for enhancing the performance of AI agents based on Large Language Models.