# Agent Workflow Memory

**Authors:** Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, Graham Neubig  
**arXiv:** [2409.07429](https://arxiv.org/abs/2409.07429)  
**Date:** 2024-09-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2409.07429v1/

## Key points

- Study focuses on enhancing performance of language model-based agents in real-world tasks like web navigation
- Challenge is dealing with long-horizon tasks involving complex action trajectories
- Humans efficiently solve tasks by learning reusable task workflows from past experiences
- Researchers introduce Agent Workflow Memory (AWM) to bridge gap and help agents benefit from similar process
- AWM involves inducing commonly reused routines or workflows for future task-solving assistance
- Abstract, sub-routine-based induction methods using Language Models (LMs) compared to rule-based methods without context and sub-routine abstraction
- LM-based workflow induction more efficient by using fewer steps and preventing unnecessary actions, improving task-solving efficiency
- AWM tested on Mind2Web and WebArena benchmarks, significantly enhancing baseline results with relative success rate improvements of 24.6% and 51.1%
- Online AWM demonstrates robust generalization capabilities across different evaluations, outperforming baselines by up to 14.0 absolute points as train-test task distribution gaps widen
- Importance of abstract, reusable workflows in improving agent performance on complex tasks emphasized

## Summary

This study focuses on enhancing the performance of language model-based agents in solving real-world tasks such as web navigation. The current challenge lies in dealing with long-horizon tasks that involve complex action trajectories. Unlike machines, humans have the ability to efficiently solve intricate tasks by learning reusable task workflows from past experiences and using them to guide future actions. To bridge this gap and enable agents to benefit from a similar process, the researchers introduce Agent Workflow Memory (AWM). This method involves inducing commonly reused routines or workflows and selectively providing them to the agent to assist in future task-solving processes. The study explores how abstract, sub-routine-based induction methods using Language Models (LMs) compare to rule-based methods without context and sub-routine abstraction. Results show that while rule- and LM-based workflow induction perform comparably in terms of success rate, the LM-based method proves to be more efficient by using fewer steps. The finer-grained workflows produced by LM-based induction prevent agents from following unnecessary steps present in rule-induced workflows, thereby improving task-solving efficiency. Furthermore, AWM is tested on two major web navigation benchmarks - Mind2Web and WebArena - covering a wide range of tasks across various domains such as travel, shopping, and social media. AWM significantly enhances baseline results on both benchmarks, with relative success rate improvements of 24.6% and 51.1% on Mind2Web and WebArena respectively. Additionally, online AWM demonstrates robust generalization capabilities across different evaluations, outperforming baselines by up to 14.0 absolute points as train-test task distribution gaps widen. Overall, the study highlights the importance of abstract, reusable workflows in improving agent performance on complex tasks. By leveraging AWM to induce and apply workflows effectively, agents can enhance their problem-solving abilities and adaptability over time in dynamic environments.