Hands-On with Computer History: OpenAI’s Take on Agent Memory

Late yesterday, OpenAI revealed the sort of automation catnip I love: Computer History. It’s a new feature for Pro, Business, and Enterprise subscribers that creates continuity for ChatGPT and Codex by converting the actions you take on your Mac into Markdown memory files that serve as context for subsequent interactions. The idea is that with the additional context, ChatGPT and Codex will better understand your requests, requiring you to provide less detail up front while getting better results. Anyone familiar with OpenClaw, Hermes Agent, and similar projects will know where this is going.

Here’s how it works.

Computer History is currently a Mac-only feature that is turned off by default, so if you want to try it, head to the Integrations section of the desktop app’s settings and toggle it on. History entries can be cleared individually, too. Before you get started, ChatGPT explains that you can pause or clear your history and manage the apps and websites that Computer History accesses at any time. I excluded Passwords, 1Password, and Keychain Access to start and may add to that list as I use the feature more and get a better sense for the sort of information that it collects.

That opening alert also says you agree to pause Computer History during video and audio calls with others to avoid recording them without their consent. I appreciate the reminder that recording people without consent is something you shouldn’t do and is even illegal in some places, but as with the many meeting summarization tools on the market, I doubt that very many people will heed the warning and comply even though doing so is just a click of the menu bar away.

Still, I appreciate the alert up front, because turning Computer History on requires trust that the information won’t be misused by OpenAI. I’m sure that’s a bridge too far for some, but one of the most frustrating parts of using a tool like Codex is explaining yourself over and over. Every thread is like Groundhog Day for an agent, and while there are techniques that help provide continuity, making it automatic with Computer History is an exciting prospect.

If you’re intrigued by the potential, it pays to read OpenAI’s documentation and understand how the feature works. First of all, Computer History is not the same as Chronicle, a retired research preview feature that relied on taking screenshots as you worked. In contrast, Computer History records interaction events but not screenshots, audio, or private web browsing sessions, according to OpenAI’s documentation. In practice, what that means is that as you work, ChatGPT will record your switches from app to app and what you’re doing in apps. Here’s what it’s captured as I started work on this story:

As you can see, every 10 minutes, ChatGPT creates a brief, timestamped summary of what I’ve been doing and lists the apps I’ve used by their icons. If I hover over an entry, I can immediately delete it or open the source file in the Finder for more detail on the information that ChatGPT is collecting.

Each entry in Computer History is a Markdown file with YAML front matter that includes a title, description, and list of the apps you’ve used followed by a structured memory document. Here, you’ll see a summary of what you’ve done:

The user continued writing computer history.md in Obsidian’s MacStories Vault, refining the section that explains how Computer History differs from the earlier Chronicle preview and adding a practical description of what the feature records while the user works. The user referenced the OpenAI Learn Computer History documentation in Safari, briefly exchanged messages with Federico Viticci, checked ChatGPT/Codex settings, and then opened MacStories Images to queue a screenshot named CleanShot 2026-08-14 at 06.08.44@2x.png for upload. The upload was queued by the end of the segment, but no completed upload or copied Markdown image link was observed.

You’ll also find details that ChatGPT deems important and non-obvious, such as the following:

Important non-obvious context about the user

• MacStories Vault - Obsidian vault where the Computer History article draft was active.
• computer history.md - active note being drafted for the Computer History article.
• CleanShot 2026-08-14 at 06.08.44@2x.png - screenshot selected from @ Screenshots and queued in MacStories Images near the end of the window.
• MacStories Images - internal Safari image uploader used for the screenshot; the page showed one queued item and an Upload 1 control.
• Federico Viticci - Messages conversation briefly active during the writing session.
• ChatGPT/Codex settings - ChatGPT app showed settings around 10:08Z, likely related to the same Computer History article context.

That’s followed by a more detailed blow-by-blow accounting of what I did over the course of the 10 minutes covered by the report. In all, it’s just shy of 700 words, which, by the end of the day, will be just one short journal entry among dozens tucked away in a folder on my Mac.

Those files are used as context for ChatGPT and Codex. I haven’t used the feature long enough to thoroughly test it, but as you can see from the screenshot above, it provided concise, correct answers when asked about what I’ve been working on today and my conversation with Federico.

That’s not too different from what the new Siri app can do on macOS Golden Gate, but as I suspected, the seams start to show when you move beyond questions to tool use. When I asked Codex to send a draft of my story to Federico, it called on Computer Use to find the Markdown file and drop it in my Messages thread with him, which is exactly what I wanted. By contrast, the Siri app pasted the text of the draft into the message itself as a wall of raw Markdown text with \n characters for line breaks. Spoiler: that’s not what I wanted.

Siri comparisons aside, Computer History is the sort of feature that deserves a check-in down the road because its true value and limitations aren’t going to reveal themselves from a few contrived examples. OpenAI’s documentation explains that it can be used to jog your memory when resuming a project, find things you recently worked on, and create workflows, which is the most interesting bit.

So many potential automations die before they’re ever created because we all have deadlines and too many things to get done. Stopping to build a skill or other automation for a task breaks your flow. Computer History promises to reduce that to a simple request like, “I just did X. Turn it into a skill.” Again, how well that works in practice remains to be seen, but it’s exactly what I want from an agent.

I’ve already touched on some of the privacy implications of Computer History, but it’s worth digging into how these memories are generated before you use the feature. The Markdown files I described are built from the interaction events that Computer History records locally as you use your Mac. OpenAI says the event files are deleted after 48 hours and are used by Codex in an ephemeral session to generate the Markdown memory files that are also stored locally. Then, when you ask a question in ChatGPT or Codex that calls on your local Markdown memory files, they are used as context for the prompt, transmitting the data as part of the request to OpenAI. So if you’re not comfortable with letting Codex use your Mac interactions to generate memories and then sending those memories to ChatGPT or Codex as context, you shouldn’t turn on Computer History.

Me, I’m excited by the prospect of Computer History. I work across a huge number of apps every day on multiple projects, and it’s too easy to lose track of it all. But most of all, I’m intrigued by the possibility of reducing the friction of creating repeatable skills to automate everyday tasks. I’m just a few hours into my Computer History experiment, and I will report back after I’ve had time to live with it for a while.