What is agent workflow memory (AWM)?

Agent Workflow Memory is a technique for teaching an LLM-based agent to remember and reuse multi-step task solutions (“workflows”) from its past experience. As the agent interacts with an environment (i.e. a web browser), AWM monitors its action sequences and induces common subroutines that led to success. These extracted workflows are stored and later injected as guiding plans when a similar task or goal is encountered. By leveraging previously discovered action sequences, AWM seems to improve long-horizon task success rates (by 51% on the WebArena benchmark) while reducing steps needed. It works both offline (mining training trajectories) and online (during live use) to continually refine the agent’s competence (paper.