What happened?

A new feature called Computer History has arrived in ChatGPT's macOS desktop app. It turns a user's actions on their computer into training data: learning how they work, suggesting automations and picking up tasks left half done.

Technically what it does is build a timeline. When the user makes a request, ChatGPT and Codex can reference that timeline. In a short demo video posted by Dominik Kundel of OpenAI's Developer Experiences team, the app looks up the last document he edited, checks whether it was shared with people via Slack, and delivers a recap of how he spent his morning.

Control sits with the user

The most consequential design decision is that the feature is opt-in. The user has to turn it on explicitly. Specific apps and websites can also be excluded from its scope, and for finer-grained control individual entries can be deleted.

Ari Weinstein of OpenAI said on X that Computer History will automatically ignore content in incognito or private browser tabs.

The Windows Recall comparison

The feature inevitably calls Microsoft's Windows Recall to mind. The difference is the point OpenAI emphasises: Microsoft's controversial feature relied heavily on screenshots, whereas OpenAI says Computer History captures no images, video or audio, relying on "events" instead.

That distinction is not trivial. A screenshot records everything on screen at that moment — passwords, private messages and third parties' information included. An event-based record holds what was done in which application, not the contents of that application. Even so, a history of actions is sensitive data in its own right: a record of who did what, and when.

The shape of the feature

  • Where: the ChatGPT desktop app on macOS
  • What it does: builds a timeline of actions from clicks and keystrokes
  • Who uses it: ChatGPT and Codex reference the timeline when a request comes in
  • Default: off; the user has to turn it on
  • Controls: exclusion by app and website, entry deletion, automatic skipping of private tabs
  • Not captured: images, video, audio

What remains on the privacy side

Not capturing screenshots is a meaningful boundary, but not sufficient on its own. A history of actions shows which application was used, when and for how long; that record alone can reveal a person's working pattern, what they are dealing with, and even the hours they do not work.

Where the data is held and how long it is kept were not detailed in the announcement either. Being able to delete entries gives the user a way back; but where that data was processed before deletion stands as a separate question.

Why does it matter?

The core problem agent products run into is missing context: telling a model to "finish this" requires the model to know where the work was left off. Computer History aims to fill that gap.

What it asks in return is a continuous record of what the user does on their computer. That the default is off, and that exclusions are offered, means the trade is being put to the user explicitly — which can be read as a lesson drawn from the Recall controversy.