Practical, but still safe
Use ChatGPT Computer History to recover work, notice patterns, and find repeatable workflows without opening your whole computer at once.
One low-risk source. One real task. One short test.
What made me keep testing
I had been using Computer History for less than 24 hours when I asked what it noticed about how I operate. It correctly warned that one day is a snapshot, not a diagnosis.
Then it identified my main constraint as attention architecture, counted roughly 196 foreground app transitions in one busy 90-minute stretch, and surfaced repeatable workflows that could become Skills.
That was useful because it gave me something specific to inspect. The goal is not to ask AI to judge your life. It is to recover real work, notice a repeated sequence, or see one attention leak you can do something about.
Computer History is optional. Skipping it is a valid decision.
Use this when
Skip it for now when
Current availability: Computer History is available in the ChatGPT desktop app for macOS on Pro, Business, and Enterprise plans. Business and Enterprise administrators must grant access first. OpenAI expanded access to Pro users in the EEA, Switzerland, and the United Kingdom on August 20, 2026. It remains off by default and requires Memories.
Do not make all three broad at the same time.
Three separate permissions
Which apps and websites may contribute clicks, typing, app switches, and other interaction context?
Start with one dedicated, low-risk app.Which activity becomes a summary or local memory that may appear in a future ChatGPT or Codex conversation?
Inspect the memory before you trust it.What is ChatGPT or Codex allowed to change, send, publish, delete, purchase, or run using connected tools?
Give the first test no external actions.Put it to work
Best first sourceA dedicated low-risk writing app with unrelated documents closed. A public research page can work too, but only when both the browser and the specific public website are included.
Keep excludedEmail, messages, calls, meetings, banking, health, payroll, passwords, client portals, confidential work, and any communication involving another person without prior express consent.
Open Settings → Data Controls and decide whether new consumer-chat content may be used to improve OpenAI's models. Computer History also requires Memories.
Open Settings → Computer history → Permissions. Choose Include only these apps and Include only these websites. Add the smallest source that can complete the test.
Use a real, non-confidential task. Do not perform a fake demo. Pause Computer History from the macOS menu bar when the session ends.
Use one of the prompts below. Require it to separate observed evidence from inference.
Open Settings → Computer history → History. Review the summary, contributing apps, and suggested workflow. Reveal the local memory file in Finder if you want to inspect the exact text.
Delete the item or clear the last hour if the capture was broader than expected. Expand to another source only when the first source saved enough repetition to justify it.
Based only on my Computer History, what was I working on, where is the real source, and what is the smallest next action? Separate what the history directly shows from anything you are inferring. Point me back to the source before suggesting an action.
Review the current Computer History window as a limited snapshot, not a diagnosis. Identify one repeated work pattern, one costly context switch, and one unfinished loop. For each, cite the observed activity that supports it. Do not make claims about my personality, health, or long-term habits from this short window.
Find one sequence I repeated that may be worth turning into a Skill or scheduled task. List the observed steps, the source apps, the decision points that still require me, and the evidence needed to call the workflow complete. Do not build or automate it yet.
Show me which apps and websites contributed to this answer. Flag any detail that appears sensitive, involves another person, or came from a source I should consider excluding. Tell me what is direct evidence, what is a generated summary, and what is your inference. Do not take any external action.
The privacy setting is part of the workflow.
Plain-language data map
Events may include clicks, typing, keyboard shortcuts, app switches, and context exposed through macOS accessibility. OpenAI says it does not capture screenshots, screen recordings, microphone input, or system audio. Private browsing is excluded.
The event stream is stored temporarily on the Mac in the ChatGPT App Group. OpenAI says it processes those files on its servers to create memories, does not retain them afterward unless legally required, and does not use them for training.
The summaries are plain-text Markdown files stored locally. Computer History does not encrypt them. Other programs running as your macOS user may be able to access them.
If ChatGPT or Codex later uses a memory in a conversation, that conversation follows the account's normal data controls. Computer History also uses tokens while creating summaries and memories.
More access is not the default reward.
Make one decision
It recovered something useful, captured what you expected, and needs no additional source yet.
It helped, but one app or website exposed more than the value justified. Exclude it before the next session.
The feature may be useful later, but your current work cannot be observed safely.
The capture was inaccurate, too broad, or not useful enough. Delete the relevant history and turn it off.
Done means: You know whether one narrow source reduced repetition, you inspected what was kept, and you left with no more access enabled than the test earned.
Sources and limits
This guide uses OpenAI's current Computer History documentation and OWASP's independent warning about persistent memory as an attack surface. Product availability, controls, and retention terms can change. Recheck the official documentation before relying on it for regulated or confidential work.
OpenAI Computer History documentation
OWASP on persistent memory risk