OpenAI has replaced Chronicle, the activity history of screenshots, with a new opt-in feature called Computer History. The change is for the ChatGPT desktop app on macOS for Pro, Business and Enterprise users. The feature does not record the screen, but interaction events: clicks, typing, keyboard shortcuts, app switches, and other context that macOS exposes thru its accessibility APIs. Those events are converted into text summaries and local memory files and placed on a timeline of activity.

If you're a UI shipper, the technical shift is easy to imagine. Chronicle took the exhibit at face value. Computer History considers the accessibility tree and the input stream to be the truth. That's how assistive tech sees an app already. Focus changes, control names, key events, window switches, not pixels. OpenAI's pitch is that this history can make for better ChatGPT responses, surface automation opportunities and make it easier to pick up where you left off.

Who can switch it on

Default is off for Computer History. Pro users can turn it on themselves. Business and Enterprise users require an admin to approve it first. It is not available in the United Kingdom, European Economic Area or Switzerland. It's not available if you call ChatGPT via an API key or thru Amazon Bedrock. The feature was published on the Register on August 14, 2026.

That list of availability matters if you work across regions, or if you're mixing desktop ChatGPT with API-backed tools. A teammate on Pro can opt in outside those restricted areas. A teammate in the UK is not capable, at least not in the account described here. The feature has nothing to do with getting the same models thru Bedrock or an API key.

What was different than Chronicle

Chronicle compiled comparable histories with screenshots. Computer History has no screen images. OpenAI said the new feature doesn't capture screen images, mic input, system audio or private-mode browsing. The Register describes the swap as event capture—clicks and typing—vs. Recall-style screenshot surveillance.

The comparison is useful even if you have never used Chronicle. Screenshot history grabs everything that was on the screen, including stuff you didn't type and UI you didn't want to save. Event history is what you did, and what the accessibility system exposes That's a narrower pipe but it's still a continuous log of interaction. It is not a place for screenshots. It is not a blank slate, either.

What is put down and what not

According to OpenAI's documentation, Computer History creates an interaction-event stream from allowed apps and websites. Events are clicks, typing, keyboard shortcuts, app switches, and the context that macOS exposes thru its accessibility system. From time to time those events become summaries in text, and local memory files.

That stream doesn't include screen images, microphone input, system audio and private-mode browsing, as OpenAI says. The company does not claim that all other classes of data are not included. Accessibility APIs can expose control names and values, selected text, and other UI metadata depending on the app. The source does not include all the fields that end up in an event file.

If you've ever wired an assistive client to a custom widget, you already know the gap between "we do not take screenshots" and "we do not record the screen's meaning." Accessibility can read a well labeled form, data table or chat transcript in its entirety without having to store a single pixel. That channel is the foundation of Computer History.

The local files for 48 hours

Interaction events are kept locally, unencrypted, for a maximum of 48 hours. And then Codex and Chatbot delete those event files. Computer History files can contain sensitive data, are not encrypted by Computer History, and can be accessed by other programs running as the same macOS user, according to OpenAI.

For a frontend developer, that warning is real. Anything else running as you - another desktop app, a helper, a debugger, malware that already has user-level access - can read those files while they exist. Local does not mean private from other processes on same account . Unencrypted means the files can be read if a process can open them.

The event files are to be deleted by ChatGPT and Codex after 48 hours. The source doesn't explain what happens if that window is closed when the app is quitting, crashing or offline. Neither the file format nor the path on disk is specified.

Memories vs. Event Files

Event files and memories are not one and the same. Events are local interactions that are stored for up to 48 hours. Those events are sent to the servers of OpenAI as well, so that the service can build memories. Memories can be stored locally for prolonged periods . Memories may also be used in future conversations, which are sent back to OpenAI.

OpenAI says it doesn't keep the files after processing them, unless it's required to do so by law, and that it doesn't use them for training. That claim is about event files after they are processed . It is not about memories . It is not about chat logs that later contain those memories .

According to The Register, OpenAI has provided chat logs to legal process even as it resists some demands. That is a different fact from the training claim. If a memory is then used in a chat, that chat is a log that OpenAI has in other instances created under legal process.

The source does not say how long memories are kept on disk, whether they can be edited or deleted separately from chats, or what "processing" means.

Price

Computer History uses tokens to record activity and to generate memory data. That adds expense. The source doesn't provide a price per hour or a token estimate, nor a view to see that usage separately. If you turn the feature on, you'll probably use a few more tokens than for normal chat and Codex usage. It's not yet clear how big that extra spend is.

Injection prompt

Computer History raises the risk of prompt injection from content in apps and websites, OpenAI says. The example given: if you go to a website that has malicious instructions, ChatGPT or Codex might execute those instructions.

Anyone who has wired model output to tools knows that is a well-known class of problem. The new piece is the path of collection. If the event stream and the summaries have page text, button labels or other content the accessibility system exposes, then instructions sitting in a page or an app can get into the same pipeline that builds memories and later chats. You don't have to paste the page into ChatGPT for that content to be context.

OpenAI's advice is operational advice, not architectural advice. Turn off the feature when communicating with other people unless you have the other people's prior express consent. Consider stopping it or leaving off apps that have sensitive health, financial or personal information.

The source doesn't mention how the exclusion is implemented in the UI, how comprehensive it is, or whether excluded apps can still leak context via window titles or app-switch events.

What we don't know

The public account leaves out a lot of practical details.

We do not know the exact schema of an event file, nor what accessibility attributes are serialized. We don't know how often is meant by "periodically." We don't know if a user can review events before sending to OpenAI. We don't know how long memories stick around locally, or whether they can be exported, inspected, or deleted independently. It is not known if Windows or other clients are planned. We don't know how much admin approval Business and Enterprise gets – per user, per workspace or per app list.

We also don't know how Computer History behaves when ChatGPT or Codex aren't running, when the machine is asleep, when the 48 hour deletion job can't run. "OpenAI does not retain your files after the event is processed and does not use the files for training," it says. The source does not independently verify for retention or training use

The note on the legal process is also short. Open AI has given chat logs if asked. The source does not report whether event files or memory files have been requested or generated.

The mental model for doing UI work is still a useful one. Chronicle recorded screen pictures. Computer History maintains an interaction log, keeps that log on disk unencrypted for up to two days, sends events to OpenAI to build memories, and may reuse those memories in future chats. It is opt-in, off by default, only available for ChatGPT desktop on macOS and not in the EEA, UK and Switzerland. It does not take screenshots, microphone, system audio, private mode browsing. It does add token cost and a bigger prompt-injection surface…

If you already trust ChatGPT desktop with your working context, Computer History is a more focused capture than Chronicle, and a more transparent one: events and accessibility data, not pixels. Please read the local-file warning before enabling it. Treat memories as something that can reenter future chats, and by extension, legal process.