OpenAI's Computer History logs your Mac by events, not screenshots

OpenAI is launching Computer History, an optional feature for ChatGPT Work and Codex on macOS available to Pro, Business, and Enterprise users ^[1][2]. With permission, it watches day-to-day activity across apps and websites and turns that into memories and a timeline ^[3]. That matters because it drops a persistent endpoint-capture agent onto managed Macs, and the design decision behind it is the interesting part for anyone who has to approve it.

The pitch is context recall. OpenAI says the feature lets ChatGPT answer questions like "What was I debugging yesterday?" or "Where did I leave off on that PR?" without a user re-explaining the whole situation ^[4]. From that history it can surface files, documents, and conversations relevant to a query, whether they lived in Google Docs, Slack, or an open browser tab ^[5].

The mechanism is what separates this from the obvious comparison. OpenAI says it does this without constantly taking screenshots ^[6]. In the announcement video, OpenAI's Dominik Kundel says "Computer History doesn't rely on screen or audio capture. Instead, it captures interaction events like clicking, typing, app switches, and more, so it can capture memories faster and more efficiently." ^[7] The New Stack reads that as OpenAI learning from Microsoft's Recall, and notes it also differs from OpenAI's own Chronicle research experiment, which pursued the same goal using screen captures and OCR, and from Record & Replay, which turns a demonstrated workflow into a reusable skill ^[8][9].

The controls are more granular than Recall's original shipping state. Users choose which apps and websites the feature can access, with a new menu bar entry to exclude any app and a desktop-app setting to exclude apps and websites ^[10][11]. It is off by default and can be paused from the menu bar ^[12]. Captured data is stored locally, so a user can inspect what was recorded and delete it ^[13].

Here is the part IT should not skim past. Even with local storage and no screenshots, the model still gets access to the summaries generated from that activity ^[15]. Event capture is not obviously safer than pixel capture from a data-governance standpoint; it means keystroke sequences, application switches, and browsing targets are being structured and summarised, and those summaries leave the machine to reach the model. Kundel frames the payoff as agents that understand your work well enough to build skills from what you just did or draft a morning standup update automatically ^[14]. That is also a description of standing, always-on behavioural telemetry tied to a cloud model, which is a different risk register than a local screenshot vault.

The rollout is largely global to Pro, Business, and Enterprise tiers, though OpenAI says users in the EEA, UK, and Switzerland will get access only in the coming weeks ^[16][17].

What to watch: whether OpenAI ships admin-level controls to disable Computer History fleet-wide, what exactly the summaries contain, and how the EEA and UK terms differ once they land.