ChatGPT Now Remembers What You Did on Your Computer — the Lock-In and the Risk

"What did I do yesterday?" You can now ask an AI that. On Aug. 13, 2026, OpenAI shipped a Computer History feature in the ChatGPT app for Mac, extending AI memory from conversations into work activity.

What It Remembers

The feature watches and remembers what you do on your computer — which apps you used, which sites you visited, what you clicked and typed — and assembles it into a timeline.

Which makes things like this possible:

• "Where's that proposal file I was looking at earlier?"
• "Summarize what I did yesterday for a report."
• "List what I handled today and where each item stands."

You don't need the file name. "That document I was reading after lunch" is enough for ChatGPT to search the timeline. It removes the routine of explaining context from scratch and attaching files every time. When it spots a repeating work pattern, it can proactively offer to automate it.

Isn't This Surveillance? Microsoft Already Found Out

This isn't the first attempt. In 2024 Microsoft announced Recall for Windows, which screenshotted the screen every few seconds and made it searchable. The backlash — "you're saving pictures of my entire screen?" — delayed the launch by nearly a year.

OpenAI appears to have studied that episode and designed differently from the start.

• It doesn't record the screen. No screenshots, screen recording, microphone or audio. Only behavior: clicks, typing, app switches.
• It's off by default. Nothing happens until the user turns it on. Opt-in.
• You choose what gets recorded. You pick which apps are tracked, and can delete the last 10 minutes, hour or day. Pages opened in private browsing aren't recorded at all.
• Records expire after 48 hours. Collected data is held temporarily for a maximum of 48 hours, and OpenAI says it isn't used for model training.

The Weaknesses OpenAI Admits

First, the log files aren't encrypted. Other apps running on the machine could access them.

Second, prompt injection risk rises. That's the technique of hiding instructions inside a webpage or document that only the AI reads, then getting the AI to follow them. The more the AI knows about your activity, the more sophisticated those attacks can get.

Which is why OpenAI's own guidance is to exclude apps that handle sensitive information. Shipping a feature alongside instructions for how not to use it signals awareness of both the utility and the risk.

Why It Matters

Memory has entered the AI competition. OpenAI has spent the year expanding memory across past conversations, files and mail. Google and Anthropic are doing the same. Capability is no longer only "who is smarter" but "who remembers me better."

Memory creates powerful lock-in. After months of use, moving from an AI that knows your work patterns to a blank one is hard. The more memory accumulates, the harder leaving becomes. Accumulated data becoming switching cost is exactly the structure behind Neko Health's 75% rebooking rate on time-series data.

You now decide what it remembers. The question shifts from "what should I ask the AI to do" to "how much do I expose to it." Banking apps, salary negotiation emails, medical appointments — that's a personal call.

For Marketers

An AI that holds work context sharply lowers the entry cost of repetitive tasks. But teams handling client material should decide the app inclusion list first. Deciding what to exclude before turning convenience on is cheaper than cleaning up afterward. As the points where AI waits on humans shrink, the boundaries you have to manage grow.