ChatGPT Computer History privacy: what OpenAI's Mac activity log records and stores
Computer History is now live in OpenAI's ChatGPT desktop app for macOS. The opt-in feature turns clicks, keystrokes, keyboard shortcuts and app switches into a searchable timeline the assistant can reuse later. It launched on August 13, is off unless a user enables it, and is open only to Pro, Business and Enterprise subscribers. The storage design makes ChatGPT Computer History privacy the central question for anyone tempted to enable it: the memory files the feature writes to disk are unencrypted plain text that other programs on the same Mac user account can read.
Computer History supersedes the earlier Chronicle research preview, which assembled similar records from screenshots. The new pipeline works without images. The raw material comes from what the macOS accessibility framework reports for apps and websites the user has approved; nothing outside that list is logged. OpenAI states that screen recordings, microphone input and system audio never enter the log. At intervals, a short-lived Codex session distills the recorded activity into summaries, writes them to the Mac as Markdown files, and arranges them in a timeline organized by day and time slot.
The feature builds on ChatGPT's existing Memories system, so the summaries it produces behave like the memories the assistant already keeps from conversations, except they now describe activity outside the chat window. Event data lives on the Mac for no more than 48 hours before deletion. The summaries are generated on OpenAI's servers. Interaction events are not used for model training, and generating the summaries consumes tokens against the subscriber's plan allowance. Users can erase history in increments, such as the last ten minutes, and private-browsing tabs are excluded from logging automatically.
The ChatGPT Computer History privacy gap
OpenAI's own documentation identifies the weak point without qualification: local memory files are not encrypted, and any other program running under the same macOS user account can read them. The documentation also acknowledges that Computer History files can contain sensitive information, which makes the plain-text storage harder to dismiss as a theoretical concern. Because those files are distilled records of what a user typed and clicked, including activity inside approved apps that may handle sensitive material, a compromised app or any malware with same-account access gains plain-text visibility into that history with no decryption step involved.
The accessibility-API design widens the exposure. The permission feeding Computer History is the same input-read mechanism keyloggers commonly target, so the feature operates at a privilege level that attracts abuse. The design also carries a prompt-injection risk: malicious web content shown in an approved browser can inject instructions into the event stream, an attack surface a screenshot-based approach would not present.
Comparisons to Microsoft's Windows Recall are inevitable, and the differences matter. Recall builds its history from screen captures; Computer History records interaction events instead. That avoids storing images but keeps raw input data, including typed text, in the record. Both tools give the assistant the ability to answer questions about recent activity, and they part ways on storage: OpenAI has confirmed its memory files ship without an encryption layer.
What gets recorded, and what you control
The feature's data lifecycle, per OpenAI's documentation, breaks down as follows:
| Stage | Detail |
|---|---|
| Recording | Clicks, typing, keyboard shortcuts and app switches via the macOS accessibility framework |
| Raw events | Temporary local files, deleted after 48 hours |
| Memories | Plain-text Markdown files on the Mac, unencrypted, kept until deleted |
| Not recorded | Screenshots, screen recordings, microphone input, system audio, private-browsing tabs |
| Plans | Pro, Business and Enterprise; off by default; Business and Enterprise require workspace-level admin approval |
| Regions | Not offered in the EEA, Switzerland or the UK |
OpenAI also spells out how much leaves the machine. The temporary event files are deleted after processing unless the company is legally required to keep them, and the summaries that replace them are generated on OpenAI's servers. The local unencrypted footprint is therefore not the only copy: a distilled version of your activity also sits on OpenAI's infrastructure, tied to the same account later conversations reference.
The controls are granular enough to limit the blast radius. Approving only work-related apps narrows what lands in the timeline, incremental deletion removes recent history on demand, and the 48-hour window for raw event files means the unencrypted exposure is confined to the distilled summaries after that point. None of those controls changes the core fact: as long as memories exist, they sit on disk as readable Markdown.
The trade-off, and the verdict
The upside is genuine for heavy ChatGPT users. The assistant can reference what you worked on minutes ago without being told, and the system can recognize repetitive processes and propose them as reusable skills or automations. For someone who bounces between many apps and tabs all day, the timeline removes the need to narrate context in each new prompt.
The cost is equally concrete. Enabling the feature places a durable plain-text record of your typing and clicking on local disk, readable by any process with the same account privileges, with deletion left to the user. Business and Enterprise get an admin gate before anyone can opt in; individual Pro subscribers get no equivalent check beyond their own restraint.
The verdict on ChatGPT Computer History privacy, weighed against the convenience, is that most Mac users should leave the feature off. Anyone who still wants it should approve only non-sensitive apps, clear history at the end of each session, and operate on the assumption that anything typed into an approved program can end up in a readable file. The lack of training use and the short raw-event retention soften the concern; the plain-text memory files do not.
Why this matters
Agentic memory features such as Computer History and Windows Recall are making activity logging a routine part of the AI assistant experience. That shifts the privacy burden onto storage design: how sensitive the data is now depends less on what a feature records and more on how the vendor protects it afterward. OpenAI chose plain text for Computer History, so enabling it is also a decision about which other programs on your Mac can read your activity.
Photo by Zulfugar Karimov on Unsplash
Related Articles
- OpenAI Security Update macOS: Urgent Patch Issued After Supply-Chain Attack
- OpenAI Opens ChatGPT Health to All US Adults with Apple Health and Medical Record Integration
- The Apple OpenAI Trade Secrets Case: Injunction as a Weapon in the Hardware Race
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.