Agent Workflow Memory
Proposes Agent Workflow Memory (AWM), which induces and stores reusable workflows, boosting web navigation success rates by 24.6% and 51.1% on Mind2Web and WebArena respectively.
Key Findings
Methodology
This paper introduces AWM, a framework leveraging large language models (e.g., GPT-4) to automatically induce reusable workflows from agent experiences. The core process involves: • extracting sub-routines from action trajectories via prompt-based guidance; • abstracting common patterns into generalized workflows; • storing these workflows in a dynamic memory system; • applying workflows in both offline (pre-collected data) and online (streaming, self-generated data) scenarios. The induction module prompts the language model with carefully designed templates to identify recurring sub-tasks, such as 'search for a product' or 'find a location by name,' abstracting specific details to enhance generality. These workflows are then integrated into the agent's memory, guiding subsequent actions. Extensive experiments on WebArena and Mind2Web demonstrate that AWM improves success rates by 24.6% and 51.1%, respectively, significantly outperforming baselines like BrowserGym, SteP, and AutoEval. The framework also exhibits strong cross-task, cross-website, and cross-domain generalization, especially in online, supervision-free settings, where it rapidly learns from minimal data (around 40 queries) and adapts to new tasks effectively.
Key Results
- On WebArena, AWM elevates success rate from 14.9% to 35.5%, a relative increase of over 50%, while reducing average steps by approximately 2 steps per task, outperforming state-of-the-art methods including BrowserGym and SteP. The success rate gains are consistent across different website splits, indicating broad applicability.
- In Mind2Web, AWM achieves a 24.6% improvement in cross-task success rate and a 14-point increase in cross-domain transfer, demonstrating robust generalization capabilities. The method maintains high element and action accuracy, especially in environments with limited annotated data.
- The online induction process enables the agent to learn effective workflows from minimal interactions, reaching peak success within about 40 queries, showcasing high sample efficiency. This adaptive learning mechanism allows the agent to continuously improve performance as it encounters new tasks and environments.
Significance
This work addresses fundamental limitations of existing web navigation agents, which rely heavily on static examples and domain-specific heuristics. By enabling agents to autonomously induce and store reusable workflows, the proposed AWM framework significantly enhances transferability, robustness, and efficiency. It mimics human-like abstraction of task routines, paving the way for more autonomous and adaptable AI systems capable of handling complex, long-horizon tasks across diverse environments. The ability to generalize across tasks, websites, and domains marks a substantial step toward practical deployment of intelligent agents in real-world scenarios such as customer service, information retrieval, and automation.
Technical Contribution
The main technical innovations include: • a prompt-guided workflow induction mechanism leveraging large language models to extract sub-routines from experiences; • a dynamic memory system that stores and retrieves workflows for task guidance; • a dual offline-online framework allowing pre-collected data or streaming self-generated data to induce workflows; • a process of abstracting workflows to enhance generalization, reducing reliance on specific examples. These contributions collectively enable the agent to learn, adapt, and transfer knowledge more effectively than prior methods, which mostly depend on fixed exemplars or handcrafted rules.
Novelty
This research is the first to systematically leverage large language models for automatic induction and storage of reusable workflows in the context of web navigation. Unlike prior approaches that rely on explicit rules or manual annotations, AWM employs prompt engineering to extract generalized sub-routines, which are then stored in a flexible memory system. This abstraction allows for effective transfer across different tasks and environments, representing a paradigm shift from static example-based learning toward dynamic, experience-driven workflow induction. The combination of online and offline modes further distinguishes this work, enabling rapid adaptation with minimal supervision.
Limitations
- The accuracy of workflow induction heavily depends on the language model's understanding capabilities; in highly complex or ambiguous tasks, the induced workflows may be suboptimal or incomplete, affecting overall performance.
- While abstraction improves generalization, it may sometimes oversimplify task-specific nuances, leading to inefficiencies or errors in certain environments.
- The computational cost of storing and retrieving workflows, especially in large-scale, multi-task settings, can be significant, necessitating further optimization of memory management and retrieval algorithms.
- The current framework primarily focuses on web navigation; extending it to other modalities or more complex environments remains an open challenge.
Future Work
Future research could explore integrating reinforcement learning to optimize workflow induction policies, incorporating multi-modal data (images, audio) for richer representations, and developing more efficient memory architectures for large-scale deployment. Additionally, enhancing interpretability and explainability of induced workflows, as well as applying the framework to real-world applications like autonomous robots or virtual assistants, are promising directions. Combining human-in-the-loop approaches to refine workflows and improve trustworthiness also warrants further investigation.
AI Executive Summary
In recent years, large language models such as GPT-4 have revolutionized natural language understanding and generation, enabling the development of intelligent agents capable of performing complex digital tasks. These agents have shown promise in web navigation, application operation, and information retrieval, but their effectiveness remains limited by reliance on static exemplars and narrow contextual understanding. Traditional methods often struggle with long-horizon, multi-step tasks that require flexible planning and adaptation across diverse environments.
Humans, on the other hand, excel at learning reusable routines or workflows from past experiences. They abstract common sub-tasks, such as searching for a location or filling out a form, and apply these routines to new situations, thereby improving efficiency and robustness. Inspired by this human capability, the authors propose Agent Workflow Memory (AWM), a novel framework that endows AI agents with similar abilities. The core idea is to automatically induce, store, and leverage workflows—sequences of sub-tasks—using large language models guided by carefully designed prompts.
The AWM framework operates in both offline and online settings. In offline mode, it leverages annotated datasets to pre-induce workflows, which are then stored in memory and used during inference. In online mode, the agent continuously induces workflows from self-generated experiences as it encounters new tasks, dynamically updating its memory. This process allows the agent to adapt rapidly, learn from minimal data, and generalize across tasks, websites, and domains.
Experimental results on two major web navigation benchmarks—WebArena and Mind2Web—demonstrate the effectiveness of AWM. On WebArena, success rates increased from 14.9% to 35.5%, a relative improvement of over 50%, while also reducing the number of steps needed to complete tasks. On Mind2Web, success rates improved by 24.6% across tasks, with strong cross-domain and cross-website generalization. Notably, the online induction mechanism enables the agent to learn effective workflows from as few as 40 queries, showcasing high sample efficiency.
These advancements mark a significant step toward autonomous, adaptable AI systems capable of handling complex, real-world tasks. By mimicking human-like abstraction and transfer of task routines, AWM addresses fundamental limitations of prior approaches, paving the way for more robust and scalable intelligent agents. Future directions include integrating multi-modal data, optimizing memory management, and expanding applications beyond web navigation to broader autonomous systems, ultimately contributing to the realization of truly intelligent, self-learning agents.
Deep Analysis
Background
随着大规模预训练语言模型（如GPT系列）在自然语言处理领域的突破，智能代理在网页导航、应用操作等任务中逐渐崭露头角。早期方法多依赖于模仿学习、硬编码规则或有限的上下文信息，难以应对复杂、多变的任务环境。近年来，研究者开始探索利用模型的推理和抽象能力，提升任务迁移和泛化能力。例如，SteP、BrowserGym等方法引入了环境建模和策略优化，但仍受限于静态示例和任务特定的设计。与此同时，如何让模型自主学习、抽取潜在的子任务流程，成为提升复杂任务表现的关键。
Core Problem
现有方法在处理长远、多步骤、跨域的网页导航任务时，表现出明显的局限。主要问题包括：• 缺乏有效的机制抽取和存储可重用的任务流程，导致模型在面对新任务时重复学习，效率低下；• 任务迁移能力不足，难以在不同网站或环境中泛化；• 依赖大量标注示例，缺乏样本效率；• 不能充分利用过去经验中的潜在结构信息，限制了模型的自主学习和适应能力。这些问题严重制约了AI在实际复杂场景中的应用。
Innovation
本研究的核心创新在于：• 提出基于大规模语言模型的流程诱导机制，通过提示工程自动抽取潜在的子流程，减少对人工设计的依赖；• 构建动态记忆系统，将诱导的流程存入模型内部，支持多任务迁移和持续学习；• 设计离线和在线两种场景下的流程诱导与应用框架，增强模型的适应性和效率；• 引入抽象化的流程描述，提升流程的泛化能力，减少对具体示例的依赖。这些创新共同推动了智能代理在复杂任务中的自主性和鲁棒性。
Methodology
- �� 经验采集：收集任务体验，包括自然语言指令和对应的动作轨迹。• 流程诱导：利用预训练的GPT-4模型，通过特定提示模板，自动抽取任务中的潜在子流程，形成抽象描述。• 流程存储：将抽取的流程存入记忆系统，支持后续调用。• 离线应用：在训练阶段，利用标注示例诱导流程，存入模型记忆；在推理阶段，加载流程指导任务执行。• 在线应用：在没有标注示例时，模型通过自我生成和评估不断诱导新流程，动态更新记忆。• 任务执行：在任务执行中，模型根据记忆中的流程，规划和执行动作，提升成功率。• 迁移和泛化：流程的抽象描述支持跨任务、跨网站迁移，增强模型的适应能力。
Experiments
采用WebArena和Mind2Web两个公开网页导航基准，分别评估任务成功率和操作步骤数。WebArena涵盖812个任务，涉及五个不同网站，强调操作的正确性和效率；Mind2Web则关注跨任务和跨域的泛化能力。实验中，比较AWM与现有最优方法（如BrowserGym、SteP、AutoEval、MindAct、Synapse），在不同设置（离线、在线）下进行。超参数包括：模型使用GPT-4（温度0.0），流程诱导的提示模板，流程存储结构等。通过 ablation 研究验证流程诱导和存储机制的贡献，分析不同场景下的成功率变化和效率提升。
Results
AWM在WebArena中，将成功率从14.9%提升到35.5%，相对提升超过50%，且平均步骤减少约2步，显著优于对比方法。在Mind2Web中，跨任务成功率提升24.6%，跨域迁移中成功率提升达14个百分点。在线诱导流程后，模型在少量样本（约40个查询）中快速学习，成功率在短时间内迅速达到峰值，表现出极强的样本效率。跨任务和跨网站测试中，AWM均优于基线，验证了其强大的迁移和泛化能力。这些结果显示，流程抽取和存储机制极大提升了复杂任务的完成效率和成功率。
Applications
该方法适用于自动化网页导航、智能助理、信息检索等场景，尤其在多任务、多域环境中表现出优异的迁移能力。未来可结合实际应用中的用户偏好和环境信息，进一步优化流程抽取和应用策略，推动智能代理在商业、教育、服务等行业的落地。长远来看，结合强化学习和多模态信息，将使得智能体具备更强的自主学习和适应能力，推动智能系统向更高层次的自主性发展。
Limitations & Outlook
当前方法在流程抽取的准确性依赖模型理解能力；面对极端复杂或模糊任务时，流程可能不够精确，影响整体性能。流程抽象虽增强泛化，但在某些环境中可能导致偏离实际场景，降低效率。存储和检索流程的计算成本在大规模、多任务环境中较高，需优化存储结构。模型在极端场景下的鲁棒性和解释性仍有待提升，未来应结合人类专家知识，增强流程可信度和可解释性。
Plain Language Accessible to non-experts
想象你在厨房里做饭。每次做菜，你都会按照一定的步骤，比如洗菜、切菜、炒菜、装盘。随着经验的积累，你会逐渐记住一些常用的步骤组合，比如炒菜的基本流程，甚至会根据不同的菜调整步骤。现在，假设你有一个聪明的机器人助手，它可以通过观察你做菜，学习到这些步骤，然后记在脑子里。下次你让它帮你做菜时，它会根据之前学到的流程，自动帮你安排好每个步骤，甚至可以在不同菜之间复用这些流程。这个机器人就像论文中的Agent Workflow Memory（AWM），它通过学习和存储“做菜流程”，帮助自己更快、更好地完成新菜。这样一来，无论是做新菜还是重复的菜，机器人都能更快完成，效率也更高。这种学习和记忆流程的方法，类似于我们人类从经验中抽象出“做菜套路”，然后用它来应对各种不同的菜肴。
ELI14 Explained like you're 14
想象你在学校的科学实验室里做实验。每次做实验，你会按照一定的步骤，比如准备材料、连接设备、进行操作、观察结果、记录数据。随着你做了很多次，你会逐渐记住一些常用的步骤组合，比如如何连接设备，或者如何观察结果。现在，假设你有一个聪明的助手——它可以观察你做实验，然后记住这些步骤。下次你让它帮你做实验时，它会根据之前学到的流程，帮你安排好每一步，甚至可以在不同的实验中重复使用这些流程。这个助手就像论文中的Agent Workflow Memory（AWM），它通过学习和存储“实验流程”，帮助自己更快、更好地完成新实验。这样一来，无论是做新实验还是重复的实验，它都能更快完成，效率也更高。这就像你在学校学会了很多做实验的套路，之后用它们来应对各种不同的实验任务。
Glossary
Agent Workflow Memory (AWM)
一种利用大规模语言模型自动抽取和存储可重用任务流程的机制，用于提升AI在复杂任务中的迁移和适应能力。技术上结合流程诱导和记忆管理，实现自主学习。
论文提出的核心技术，用于网页导航任务的流程抽取和存储。
Workflow (流程)
一系列有序的操作步骤，用于完成特定任务，具有抽象描述和具体执行两部分。技术上由自然语言描述和动作序列组成，支持跨任务迁移。
存储在模型记忆中的子任务结构。
Prompt Engineering (提示工程)
设计特定的输入提示，引导语言模型生成目标内容或抽取信息。通过模板和示例优化模型输出的相关性和准确性。
流程诱导的关键技术手段。
离线场景 (Offline Scenario)
在任务前利用预标注数据或示例，诱导出流程并存入记忆系统，供推理时调用。
适用于有丰富标注数据的训练环境。
在线场景 (Online Scenario)
在没有大量标注示例的情况下，模型通过自我生成和评估不断诱导新流程，动态更新记忆。
适用于实时或少量数据的任务环境。
迁移学习 (Transfer Learning)
将从一个任务中学到的知识迁移到不同任务中，提升模型在新任务中的表现。
AWM在多任务、多域中的核心优势。
抽象化流程 (Abstracted Workflow)
将具体流程中的元素泛化为更通用的描述，减少对特定示例的依赖，增强迁移能力。
提升模型跨任务泛化能力。
网页导航 (Web Navigation)
在网页环境中，智能体根据指令自动浏览网页、提取信息或完成操作的任务。
论文中的主要应用场景。
成功率 (Success Rate)
模型在任务中正确完成目标的比例，是衡量任务性能的重要指标。
用于评估AWM在不同数据集上的表现。
操作步骤 (Action Steps)
在任务执行中，模型采取的具体操作指令，如点击、填写、滚动等。
衡量任务完成效率的关键指标。
跨任务迁移 (Cross-task Transfer)
将从一个任务中学到的流程或策略应用到不同任务中，以提升泛化能力。
AWM的核心优势之一。
多模态信息 (Multimodal Information)
结合多种信息类型（如视觉、声音、文本）以增强模型理解和决策能力。
未来可能的扩展方向。
流程诱导 (Workflow Induction)
利用模型从经验中自动抽取潜在的子任务流程的过程。
实现AWM的核心技术。
记忆系统 (Memory System)
存储诱导出的流程和经验，用于指导未来任务执行的机制。
模型自主学习和迁移的基础。
提示模板 (Prompt Template)
设计特定格式的输入，用于引导模型生成目标内容或抽取信息。
流程诱导的关键技术手段。
任务抽象 (Task Abstraction)
将具体任务中的操作和目标泛化为更高层次的描述，便于迁移和重用。
增强模型泛化能力的重要手段。
Open Questions Unanswered questions from this research
- 1 尽管AWM在网页导航任务中表现优异，但其在极端复杂或模糊任务中的流程抽取准确性仍有待提升。未来需要研究更鲁棒的流程抽取技术，尤其是在环境信息模糊或任务定义不明确时的表现。
- 2 流程的抽象化虽然增强了迁移能力，但在某些特定环境中可能导致偏离实际场景，影响效率。如何平衡抽象与具体的关系，是未来研究的重点。
- 3 在大规模、多任务环境中，流程存储和检索的计算成本较高，尤其是在高频更新和多域迁移场景下，存储结构和检索机制亟需优化。
- 4 模型在极端场景下的鲁棒性和解释性仍有不足，未来应结合人类专家知识，增强流程的可信度和可解释性。
- 5 如何结合多模态信息（如视觉、声音）丰富流程描述，提升模型对复杂环境的理解能力，是未来的重要方向。
Applications
Immediate Applications
智能网页助手
利用AWM实现自动化网页导航和信息提取，帮助用户快速完成复杂任务，提升效率，适用于企业客服、数据采集等场景。
企业流程自动化
将企业中的重复操作流程抽象化存储，自动指导机器人完成日常任务，减少人工干预，提高生产效率。
个性化智能助理
结合用户历史操作流程，动态学习和存储个性化流程，为用户提供定制化的操作建议和自动化服务。
Long-term Vision
自主智能体系统
构建具备自主学习和迁移能力的多任务智能体，广泛应用于自动驾驶、智能制造、智慧城市等领域，推动AI全面自主化。
跨模态多任务系统
结合视觉、声音等多模态信息，构建更智能、更适应复杂环境的自主系统，实现人类级别的认知和操作能力。
Abstract
Despite the potential of language model-based agents to solve real-world tasks such as web navigation, current methods still struggle with long-horizon tasks with complex action trajectories. In contrast, humans can flexibly solve complex tasks by learning reusable task workflows from past experiences and using them to guide future actions. To build agents that can similarly benefit from this process, we introduce Agent Workflow Memory (AWM), a method for inducing commonly reused routines, i.e., workflows, and selectively providing workflows to the agent to guide subsequent generations. AWM flexibly applies to both offline and online scenarios, where agents induce workflows from training examples beforehand or from test queries on the fly. We experiment on two major web navigation benchmarks -- Mind2Web and WebArena -- that collectively cover 1000+ tasks from 200+ domains across travel, shopping, and social media, among others. AWM substantially improves the baseline results by 24.6% and 51.1% relative success rate on Mind2Web and WebArena while reducing the number of steps taken to solve WebArena tasks successfully. Furthermore, online AWM robustly generalizes in cross-task, website, and domain evaluations, surpassing baselines from 8.9 to 14.0 absolute points as train-test task distribution gaps widen.
References (20)
Autonomous Evaluation and Refinement of Digital Agents
Jiayi Pan, Yichi Zhang, Nicholas Tomlin et al.
HeaP: Hierarchical Policies for Web Actions using LLMs
Paloma Sodhi, S. Branavan, Ryan Mcdonald
WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks?
Alexandre Drouin, Maxime Gasse, Massimo Caccia et al.
Mind2Web: Towards a Generalist Agent for the Web
Xiang Deng, Yu Gu, Boyuan Zheng et al.
Synapse: Trajectory-as-Exemplar Prompting with Memory for Computer Control
Longtao Zheng, R. Wang, Bo An
WebArena: A Realistic Web Environment for Building Autonomous Agents
Shuyan Zhou, Frank F. Xu, Hao Zhu et al.
World of Bits: An Open-Domain Platform for Web-Based Agents
Tianlin Shi, A. Karpathy, Linxi (Jim) Fan et al.
Code as Policies: Language Model Programs for Embodied Control
Jacky Liang, Wenlong Huang, F. Xia et al.
Language Models Can Teach Themselves to Program Better
Patrick M. Haluptzok, Matthew Bowers, A. Kalai
WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents
Shunyu Yao, Howard Chen, John Yang et al.
DreamCoder: growing generalizable, interpretable knowledge with wake–sleep Bayesian program learning
Kevin Ellis, Catherine Wong, Maxwell Nye et al.
Reinforcement Learning on Web Interfaces Using Workflow-Guided Exploration
E. Liu, Kelvin Guu, Panupong Pasupat et al.
Large Language Models as Tool Makers
Tianle Cai, Xuezhi Wang, Tengyu Ma et al.
Zero-Shot Task Generalization with Multi-Task Deep Reinforcement Learning
Junhyuk Oh, Satinder Singh, Honglak Lee et al.
On the Nature of Expertise.
S. Kay
Categorization and Representation of Physics Problems by Experts and Novices
M. Chi, P. Feltovich, R. Glaser
VisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks
Jing Yu Koh, Robert Lo, L. Jang et al.
AutoGuide: Automated Generation and Selection of State-Aware Guidelines for Large Language Model Agents
Yao Fu, Dong-Ki Kim, Jaekyeom Kim et al.
Top-Down Synthesis for Library Learning
Matthew Bowers, Theo X. Olausson, Catherine Wong et al.
Voyager: An Open-Ended Embodied Agent with Large Language Models
Guanzhi Wang, Yuqi Xie, Yunfan Jiang et al.
Cited By (20)
Automata from Agent Traces: Failure and Next-Step Prediction
Dual-Grained Agent Memory and Shapley Context Attribution for Multimodal Agentic Learner
Progressive Agent Skill Generation via Reinforcement Learning
The Path to Self-Evolving Clinical Systems: Scaling Medical Agents from Assistance to Autonomy
On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification
Mechanistic Attention Guidance for Agent Memory Refinement
MCPEvol-Bench: Benchmarking LLM Agent Performance Across Dynamic Evolutions of MCP Servers
The Compliance Trap: Diagnosing How AI Agents Consume Conflicting Memory
Knowledge-Centric Self-Improvement
AgRefactor: Self-Evolving Agentic Workflow for HLS Compatibility and Performance
RESOURCE2SKILL: Distilling Executable Agent Skills from Human-Created Multimodal Resources
Speak Your Network: Automated Network Emulation Construction with Cost-Efficient Multi-Agent Orchestration
Joint Learning of Experiential Rules and Policies for Large Language Model Agents
Memory as a Controlled Process: Learned Adaptive Memory Management for LLM Agents
SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution
Learning on the Job: Continual Learning from Deployment Feedback for Frozen-Weights Agents
From Cognitive Architectures to Language Agents: A Mechanism-Level Review of Lineage, Convergence, and Migration Gaps
Bridging Inference-Time Scaling and Episodic Memory with Action-Centric Graphs
Rethinking Self-Evolution: A Constrained Exploration-Exploitation Process for Mitigating Skill Overfitting
Reusing Past Repairs Through Hierarchical Trajectory Abstraction for Coding Agents