Why browser and computer-use agents need work memory
Reading the current page tells an agent what is visible now. Recovering earlier research, decisions, and interruptions requires a different layer of context.
この記事を日本語で読む →
The current page is only the current page
A browser agent can inspect a page and act on it. A computer-use agent can also navigate windows and controls. Neither capability, by itself, tells the agent why you opened that page yesterday, which alternative you rejected, or where you stopped after an interruption.
Imagine comparing an API reference, an issue thread, and an internal document before changing a setting. The current DOM can identify the Save button. It cannot reconstruct the investigation that made one value preferable to another.
Different records answer different questions
| Source | Useful for | Important limit |
|---|---|---|
| Current page or accessibility tree | Acting on the interface now | Does not explain earlier work |
| Browser history | Finding pages visited in a time window | A visit is not proof that a page was read or accepted |
| Screens and app activity | Recovering where work happened | May lack the reason behind a choice |
| Agent conversations | Finding earlier questions and suggestions | May omit human research outside chat |
| Work summary | Finding a practical restart point | Must be checked against its sources |
Memory is useful when it connects these clues for a specific question. It should return a relevant slice of work, not send every recorded event to every model request.
Keep control near the person doing the work
Contextberg's desktop app stores its work history locally. You can choose which recording signals are enabled and inspect the Record timeline. The Browser AI History extension separately saves supported AI chat conversations in the browser; its history can help recover what was discussed while researching on the web.
For a question such as “Which documentation did I inspect before changing this configuration?”, Chat or a connected agent can use the available context to suggest a trail. The answer should identify uncertainty: a page that appeared in history is a candidate source until you verify what it says.
Memory complements browser control
Browser control helps an agent perform the next action. Work memory helps it understand what led to that action. Use the two together: recover the previous task, inspect the live page, confirm the relevant facts, and only then make a change.