Agent Workflow Memory

People who use a website often learn routines: how to search for a product, how to find an order, how to filter results. Web agents usually do not; they get a fixed prompt and solve each task as if it were the first. Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried and Graham Neubig at CMU and MIT built Agent Workflow Memory (AWM), which lets an agent extract such routines from its own experience and use them on later tasks.

A workflow has a short description of the goal ("find a place by its name") and a series of steps, each with a description of the page state, the reasoning, and an executable action. Workflows are induced by prompting the model with past trajectories and asking for common sub-routines at a finer grain than whole tasks, with example-specific values abstracted into placeholders, so "dry cat food" becomes {product-name}. Induced workflows are added to the agent's prompt as memory, grouped per website.

AWM runs in two modes. Offline, when annotated training examples exist, it induces workflows from all of them first and gives every test task the same memory. Online, with no labels, it processes test tasks as a stream: after each task, an LLM evaluator judges whether the attempt succeeded, and if so, the trajectory is turned into workflows for the next task. Workflows compound: "find a place by its name" becomes a step inside "get the zip code of a place".

On WebArena, online AWM with GPT-4 solved 35.5% of 812 tasks, above BrowserGym's 23.5% (the best autonomous agent reported), and above SteP's 33.0%, which uses 14 workflows written by human experts. It also took fewer steps. On Mind2Web, AWM improved step-level success in cross-task, cross-website and cross-domain tests, most where the test domain differed from training.

Agents that operate a customer's admin consoles, CRMs or ticketing systems repeat the same sub-tasks all day. AWM shows a cheap way to make them learn those routines: no fine-tuning, just text in the prompt, induced from runs that worked.

Sub-routines, not whole tasks. The useful unit is "open the order by ID" or "filter tickets by assignee", reused across many requests. Whole-task examples transfer poorly and bias the agent toward the specific values in them.

The success judge is the gatekeeper. Online learning stores whatever the evaluator calls a success. A judge that accepts failures fills memory with misleading routines. Use real signals where you can: a confirmation page, a database change, an API response.

Keep memory per application. AWM keeps workflows per website, which keeps prompts small and relevant. The same applies to per-system memory in a customer's environment.

Expect UI changes to break workflows. Workflows encode page structure. When the application changes, stored routines go stale; plan for re-verification.