Generated by Codex with GPT 5.6 Sol XHigh

Techmeme surfaced OpenAI’s August 14 documentation for Computer History, a macOS feature that turns activity across apps and websites into a timeline and memories that ChatGPT and Codex can use. The original URL is OpenAI’s documentation page, which explains both the workflow and the privacy model.

The feature matters because it attacks a basic weakness of AI assistants: every new task begins with the user reconstructing context. Computer History watches an allowed slice of recent work, summarizes it, and makes those summaries available later. A user can ask what they were doing before a break, find a vaguely remembered document, prepare a standup update, or ask Codex to turn a repeated sequence into a skill or automation.

That is more than better search. It is an attempt to make an agent learn the shape of a person’s work before being explicitly instructed to automate it.

From Activity to Reusable Context

Computer History is off by default in the ChatGPT desktop app for macOS. It is available to Pro, Business, and Enterprise users; Business and Enterprise administrators must first grant access, after which each person still has to opt in. It requires Memories, is unavailable through API keys or Amazon Bedrock, and at launch is not offered in the European Economic Area, Switzerland, or the United Kingdom.

Once enabled, the feature records interaction events from approved apps and websites. These can include clicks, typing, keyboard shortcuts, app switches, and context exposed through macOS accessibility APIs. It does not record screenshots, screen video, microphone input, system audio, or private-browsing activity. This is a rebuilt successor to the earlier Chronicle preview, which did rely on screenshots.

The event stream is periodically converted into text summaries and ordinary Markdown memory files. The timeline groups them by time, identifies the contributing apps, and may suggest a skill or automation when it detects repeatable work. The memories are not meant to replace the underlying sources. If a Slack thread, Google Doc, or file is the real authority, the history can help identify it and ChatGPT or Codex can then read it through the usual access path.

That distinction is important. Computer History is a contextual index, not a new permission system. It can tell an agent where the relevant work happened, but the agent still needs access to the original document or app before acting on it.

The Automation Bet

The most ambitious part is the path from observation to automation. Most people do not document a workflow while performing it, and many small automations never get built because stopping to specify the process costs more time than the task seems worth. Computer History tries to capture enough of the sequence that a user can finish the work first and then say, in effect, “turn what I just did into a reusable skill.”

A first-day test by MacStories shows what that looks like in practice. The system created timestamped summaries roughly every ten minutes, recorded the apps involved, and produced detailed Markdown entries about the author’s drafting and publishing workflow. It correctly answered questions about recent work and helped Codex locate a draft and send it through Messages. The author had not used it long enough to judge reliability, but the experiment illustrates the product thesis: persistent context can reduce both repeated explanation and the cost of discovering automatable routines.

This is also why the feature could become more consequential than a conventional activity log. A log looks backward. Computer History uses the past to propose future behavior. If that loop works, the assistant moves from waiting for a fully specified request to noticing patterns, recovering state, and offering to package repeated work.

The Privacy and Security Bargain

OpenAI has built several layers of control around a feature that necessarily observes sensitive behavior. Users choose which apps and sites are included, can switch from broad exclusions to an allowlist, can pause collection from the menu bar, and can delete individual entries or clear recent or complete history. OpenAI explicitly warns users to pause during communications with other people unless they have prior express consent, and to exclude health, financial, password, and other sensitive sources.

The storage model is local-first but not local-only. Raw interaction-event files are temporarily stored in the ChatGPT app’s isolated macOS container and deleted after no more than 48 hours. An ephemeral Codex session sends those events to OpenAI for processing into memories; OpenAI says it does not retain the event files after processing or use them for training. The resulting Markdown memories remain on the Mac until deleted.

Those local memories are readable files, not an encrypted vault. Other software running as the same macOS user may be able to access them. When ChatGPT or Codex later uses a memory, relevant content can be sent back to OpenAI as prompt context, and that chat content follows the user’s normal data-control settings. “Stored locally” therefore does not mean “never leaves the device.” It describes where the durable memory lives, while generation and later use can still involve server processing.

OpenAI also flags prompt injection directly. A malicious instruction embedded in a visited page or application could enter the recorded context and influence a later ChatGPT or Codex task. The feature expands an agent’s memory, but it also expands the set of untrusted content that may shape the agent. App allowlists, cautious source selection, and ordinary permission boundaries remain necessary even when the summaries themselves look benign.

Takeaway

Computer History is a serious attempt to solve agent continuity. Its most useful idea is not simply remembering what appeared on screen; it is turning a stream of work into an inspectable timeline, human-readable memory, and candidate automation. That could make AI assistance feel less like opening a fresh chat and more like resuming work with a collaborator who knows what happened earlier.

The same design makes the tradeoff unusually clear. To remove the burden of repeatedly explaining context, the system must observe and summarize more of the user’s digital life. OpenAI has made collection opt-in, visible, filterable, and locally inspectable, but the resulting files can still contain sensitive information, server processing still occurs, and prompt injection can travel through the new memory channel.

The feature’s real test will therefore be trust, not novelty. It succeeds only if the saved context is accurate enough to reduce friction, selective enough to avoid becoming surveillance clutter, and safe enough that users are willing to let an agent learn from the routines they previously kept implicit.