OpenAI Computer History Raises Privacy Concerns Over Unencrypted Mac Memory Files
Table of Contents
OpenAI has introduced Computer History, an opt-in feature for the ChatGPT desktop app on macOS that records selected computer interactions and turns them into a searchable timeline. The feature allows ChatGPT and Codex to use information about a user’s previous activity to provide more context-aware assistance.
Instead of relying on screenshots, Computer History records interaction events such as clicks, typing, keyboard shortcuts and application switches through macOS accessibility capabilities.
The feature replaces Chronicle, an earlier Codex research preview that used periodic screenshots to understand a user’s computer activity.
What Computer History Records
Computer History is designed to understand a user’s workflow without continuously capturing images of the screen.
OpenAI says it does not capture screenshots, screen recordings, microphone input, system audio or video. Activity from private or incognito browser tabs is also excluded.
Users can use the resulting timeline to find things they previously worked on, retrace earlier activity or ask ChatGPT to summarise parts of their workday.
The feature can therefore provide context that would otherwise be unavailable to a conventional chatbot.
How the Data Is Stored
Computer History creates two different categories of data that should not be confused.
Raw interaction events are temporarily stored on the Mac for up to 48 hours. OpenAI says those events are processed on its servers to generate memories and are not retained afterward unless required by law.
The resulting memories remain stored locally on the user’s Mac as Markdown files.
The important security issue is that these local memory files are not encrypted. OpenAI’s documentation warns that other programs operating under the same macOS user account may potentially be able to access them.
That does not mean every application automatically has access, but it creates an additional security consideration if malicious or untrusted software is already running under the same account.
Prompt Injection Is Another Concern
OpenAI has also highlighted a separate risk involving prompt injection.
Because Computer History can capture information about activity across websites and applications, malicious instructions contained in websites or documents could potentially influence what an AI system later encounters in its history.
This is different from the local-file encryption issue. The first concerns unauthorized access to stored information, while prompt injection concerns potentially malicious content influencing an AI system’s behaviour.
OpenAI has advised users to be careful when recording communications involving other people and to use the available controls when sensitive activity is involved.
Privacy Controls Give Users More Control
Computer History is off by default, meaning users must actively enable it.
Users can control which applications and websites contribute to their history and can pause collection. They can also delete stored timeline information, including recent periods of activity.
For Business and Enterprise accounts, administrators must enable access before individual employees can opt in.
The feature is currently unavailable in the European Economic Area, Switzerland and the United Kingdom. OpenAI’s availability information does not establish a specific regulatory reason for those regional restrictions, so they should not automatically be attributed to European privacy regulations.
How Training Fits Into Computer History
OpenAI says the Computer History memory files themselves are not used to train its models.
However, users should distinguish those files from future ChatGPT conversations.
If information from Computer History is later surfaced as context in a conversation, the resulting chat may be subject to the user’s existing data and training settings. The two types of data therefore should not be treated as having identical policies.
Computer History vs. Microsoft Recall
The launch inevitably invites comparisons with Microsoft Recall, another AI-related feature designed to help users retrieve information from previous computer activity.
The approaches are different.
Recall’s initial controversy centred heavily on its use of screenshots, while Computer History relies on interaction events rather than continuous screen captures. OpenAI has also made Computer History opt-in from the beginning.
Nevertheless, both products raise the broader question of how much computer activity users are comfortable allowing an AI assistant to remember.
OpenAI Computer History: Key Facts
| Features | Details |
| Developer | OpenAI |
| Platform | ChatGPT for macOS |
| Status | Opt-in |
| Captures screenshots | No |
| Records | Clicks, typing, shortcuts and app switches |
| Raw event storage | Up to 48 hours |
| Persistent memories | Local Markdown files |
| Memory files encrypted | No |
| Training of memory files | No |
| Regional exclusions | EEA, Switzerland and UK |
Conclusion
OpenAI Computer History is designed to make ChatGPT and Codex more useful by giving them a searchable record of selected computer activity. Its event-based approach avoids screenshots and recordings while allowing users to retrieve information about previous work.
However, the feature also introduces genuine privacy and security considerations. The persistent memory files are stored locally in unencrypted Markdown, while prompt injection creates another potential risk for users who allow websites and applications to contribute to their history.
The feature’s opt-in design and controls give users significant control over whether and what they record. Ultimately, users will need to balance the convenience of AI-powered computer memory against the sensitivity of the activity they allow Computer History to retain.
