Research map / Multi-agent systems and games / LLM agents
Agent Workflow Memory
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Agent Workflow Memory (AWM), which induces reusable task workflows from past agent trajectories with an LM and feeds them back into memory, offline or online, to improve web agents on WebArena and Mind2Web.
Problem
Language model web agents struggle with long-horizon tasks with complex action trajectories and do not learn reusable workflows or adapt from past successes and failures across tasks.
Approach
Induce abstract, sub-routine workflows (with example-specific values replaced by variables) from successful experiences using an LM or rules, store them as auxiliary memory, and optionally run this induction online over streamed test queries without any labeled data.
Key results
- On WebArena, AWM reaches 35.5% total success rate with gpt-4 versus 23.5% for BrowserGym (about 51.1% relative), and beats the human-workflow SteP baseline (33.0%).
- On Mind2Web cross-task, AWM offline with gpt-4 reaches 45.1% step success rate vs 36.2% for MindAct.
- Online AWM on Mind2Web cross-domain improves step success rate to 35.5% vs 18.6% for MindAct, with gains widening as train-test gaps grow.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add online workflow memory | GPT-4 web agent · WebArena | Total SR: 23.5 -> 35.5 | Helped |
| Add online workflow memory | GPT-4 web agent · WebArena | # Steps: 7.9 -> 5.9 | Helped |
| Add online workflow memory | GPT-4 web agent · WebArena cross-template | Total SR: 20.5 -> 33.2 | Helped |
| Add offline workflow memory | GPT-3.5-turbo agent · Mind2Web cross-task | Step SR: 30.6 -> 34.6 | Helped |
| Add offline workflow memory | GPT-4 agent · Mind2Web cross-task | Step SR: 36.2 -> 45.1 | Helped |
| Add online workflow memory | GPT-4 agent · Mind2Web cross-domain | Step SR: 18.6 -> 35.5 | Helped |
| Use LM-based workflow induction instead of rule-based induction (ablation) | AWM · WebArena | Total SR: 35.6 -> 35.5 | No clear effect |
| Use LM-based workflow induction instead of rule-based induction (ablation) | AWM · Mind2Web cross-task | Step SR: 43.4 -> 45.1 | Helped |
| Represent workflow actions as descriptive text instead of code (ablation) | AWM · Mind2Web cross-task | SR: 4.8 -> 3.6 | Hurt |
| Replace NL workflow-state descriptions with filtered HTML (ablation) | AWM · Mind2Web cross-task | Step SR: 34.6 -> 33.8 | Hurt |
| Add filtered HTML to NL workflow-state descriptions (ablation) | AWM · Mind2Web cross-task | Step SR: 34.6 -> 32.9 | Hurt |
| Add induced workflow actions to the agent action space (ablation) | AWM · Mind2Web cross-task | Step SR: 45.1 -> 46.4 | Helped |
Surprising findings
- Rule-based workflow induction performs nearly as well as LM-based induction on WebArena (35.6% vs 35.5%).
- Combining offline and online workflows underperforms either alone, as offline workflows appear to impair online workflow quality.
- Agents invoke workflow actions in only 18.5% of tasks even when they are added to the action space.
Limitations
- Online induction can create incorrect workflows from wrong model-predicted trajectories, degrading performance.
- Fixed workflow action sequences cannot handle dynamic UI states such as pop-ups, and agents rarely call workflow actions.
- Experiments rely on gpt-4 and gpt-3.5-turbo only, with results on closed API models.
Experimental setup
Datasets: WebArena Mind2Web
Benchmarks: WebArena Mind2Web
Metrics: task success rate step success rate element accuracy action F1 number of steps
Baselines: BrowserGym BrowserGym_ax-tree AutoEval SteP WebArena MindAct Synapse CogAgent
Builds on: MindAct Synapse BrowserGym AutoEval
Closest papers by meaning
- AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents (ICLR 2025 · similarity 0.88)
- Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.88)
- WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 2026 · similarity 0.87)
- WebArena: A Realistic Web Environment for Building Autonomous Agents (ICLR 2024 · similarity 0.87)
- AdaMEM: Test-Time Adaptive Memory for Language Agents (ICML 2026 · similarity 0.87)
- Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks (ICML 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.