Agent Workflow Memory

One-Minute Overview#

Agent Workflow Memory (AWM) is an innovative system that enables AI agents to learn and reuse workflows through memory mechanisms, particularly effective for web automation tasks. Researchers and AI developers can utilize it to improve task completion rates of intelligent agents in complex environments, achieving excellent performance on benchmarks like WebArena and Mind2Web.

Core Value: Significantly improves success rates and efficiency of AI agents in web automation tasks by memorizing and reusing workflows.

Quick Start#

Installation Difficulty: Medium - Requires environment setup and data preparation, but detailed instructions are provided

# Run AWM on WebArena
cd webarena
python pipeline.py --website "shopping" # choose from: ['shopping', 'shopping_admin', 'reddit', 'gitlab', 'map']

# Run AWM on Mind2Web
cd mind2web
python pipeline.py --setup "offline" # or "online"

Is this suitable for my scenario?

• ✅ Researching agent workflows: AWM offers a new approach to studying how AI agents learn and utilize workflows
• ✅ Web automation tasks: Particularly effective in WebArena and Mind2Web environments
• ❌ Need plug-and-play solutions: Requires substantial environment setup and data preparation
• ❌ Simple web scraping: Overly complex for basic automation, designed for complex tasks

Core Capabilities (Optional)#

1. Workflow Memory and Reuse - Enhance Task Completion Efficiency#

Stores and reuses workflows through agent memory, avoiding relearning the same task steps Actual Value: Significantly reduces the number of attempts needed for AI agents to complete tasks, improving success rates and efficiency

2. Offline and Online Learning Modes - Flexible Adaptation to Different Scenarios#

Supports offline mode (learning from annotated data) and online mode (learning from real-time experiences) Actual Value: Allows selection of the most appropriate training method based on available data, maximizing adaptability and practicality

3. Cross-task Generalization - Adaptable to Diverse Websites and Domains#

Particularly effective in the Mind2Web environment, generalizing across various tasks, websites, and domains Actual Value: Reduces the need for retraining on new tasks, lowering the application barrier

Technology Stack and Integration#

Development Language: Python Main Dependencies: Not explicitly listed, but requires WebArena and Mind2Web environments Integration Method: Code library, usable by running specified pipeline scripts

Maintenance Status (Optional)#

• Development Activity: Actively developed, with the last commit on December 22, 2025
• Recent Updates: Recent feature additions and optimizations
• Community Response: Moderate community engagement, with 3 open issues and 3 closed issues

Documentation and Learning Resources (Optional)#

• Documentation Quality: Basic documentation with a quick start guide
• Official Documentation: Accessible via the repository README
• Example Code: Provides example code for two primary use cases: WebArena and Mind2Web