OpenAI's feature called Computer History records the clicks you make, the keys you press and your switches between applications on a Mac. Those records are turned into a searchable timeline for ChatGPT and Codex.

What it is for

The purpose is to give the model context. For an assistant to understand a reference like "that file you were working on yesterday" or "that bug you looked at last week", it has to know about those moments. Computer History closes that gap by keeping a record of the time a user spends on the computer.

In practice it removes the burden of re-explaining history every time you talk to the assistant. The model already knows what you did and when.

Where the data sits

The technical detail matters: the data is stored on the device, as unencrypted Markdown files.

That cuts both ways. On the positive side the record does not go to a server, and the files sit in a readable format — you need no special tool to see what was captured, delete it or export it.

The negative side comes from the same place: an unencrypted file can be read by any software with access to the device. Keystroke records, applications visited and working patterns, taken together, form a record describing a user's day in detail.

The training data question

OpenAI says the data is not used for AI training. But The Decoder points to a distinction: memories that feed into chats may still end up as training data.

The chain runs like this: the record stays on the device, a memory derived from it enters a chat, and chats are a channel that can be used for training. The record itself may not travel, but information extracted from it can.

What it means for users

The feature's value and its risk come from the same thing: how much the model knows about you. A useful assistant requires that; privacy requires the opposite.

Three things are worth checking when assessing it:

  • Scope — which applications are recorded and which can be excluded.
  • Retention — how long records are kept and whether deletion is automatic.
  • Egress — what information moves from the device into a chat, and under what conditions.

Features of this kind are neither the first nor the last in the industry; assistants need context to be useful. The real issue is whether the user can see where that context came from.

A separate problem on work machines

The feature has an equivalent on work computers, and there the question stops being one of personal preference. An employee's screen may hold customer data, contract text or an unannounced product plan; all of it enters the same record.

An unencrypted local file adds weight in that context: if the device is lost or picks up malware, a single directory holds a readable account of what the employee did for weeks. Corporate disk encryption covers part of this but offers no protection against a process running on an unlocked device.

A point of comparison

Microsoft attempted a similar approach on Windows with a screenshotting feature and had to delay it after heavy criticism. The lesson from that episode was less about the record existing than about it arriving switched on by default.

How Computer History is received will largely turn on the same point: not the feature itself, but whether the user knowingly turned it on.