ChatGPT can finally learn how you work.
Decide what it gets to see.
The feature I wanted can recover lost work and reveal repeated patterns.
The same context that makes it useful creates the boundary you cannot skip.
Start include-only, inspect what it kept, and clear the test.
Start with the useful partGet the six decisions that matter this week.
See the whole weekRead everything I kept.
What I read. What I tested. What I left alone.
My goal: save you time, reduce noise, and help you make steady progress.
Overwhelm beat
The feature I wanted creates the decision I cannot skip
I have said some version of this more than once: if AI could learn how I do things and what I am doing, I would lose fewer good ideas.
The hard part of knowledge work is not always creating the idea. It is remembering which tab, chat, document, voice note, or half-finished project held it when you finally have time to come back.
ChatGPT answered that problem this week with Computer History. I turned it on yesterday afternoon.
Less than 24 hours later, I asked it to give me feedback on how I operate day to day. It warned me that the sample was still short, then named something I recognized immediately:
My primary constraint is attention architecture, not ability or effort.
In one 90-minute stretch, it counted roughly 196 foreground app transitions. Some were system windows, but the larger pattern was accurate. I was moving between a launch, member support, product repair, pricing checks, document review, and several unfinished work loops.
It also found three repeatable sequences that could become Skills. That feels like gold to me. It reflected how I work, showed me where attention was leaking, and found work that should not require me to remember every step next time.
One day is not a diagnosis. It is enough to recognize a pattern that was already costing me something.
The benefit and the risk come from the same thing: more of your activity can become context.
Let Computer History help you find the work. Don’t make it the only place the work lives.
Use it to recover what happened, find the real source, and notice repeated work. Keep the durable version of your knowledge in files and systems you own and can move.
Read this first
This issue in 30 seconds
Computer History can recover work between your chats.
Start include-only with one low-risk source. Do not give it your whole computer on day one.
Google Drive is now inside ChatGPT Library.
Bring one active folder into a real job instead of uploading and explaining the same files again.
You can change an existing Project's memory boundary.
Use Project-only memory when one client, offer, or body of work should stay separate.
Claude is adding an invisible watermark to future text.
There is no detection tool yet and no reader task today. Keep your existing work where it is.
Gemini's agent layer got broader and faster.
Retest one workflow only if Spark or one of the named services is already part of your stack.
A benchmark cannot pick your permanent AI platform.
Use the model that does your real work best now. Keep your context portable.
If you do one thing: run the 30-minute Computer History boundary test near the end of this Brief.
ChatGPT can now remember the work between your chats
This is more useful than a longer chat history, and more sensitive.
Computer History arrived in the ChatGPT desktop app for macOS on August 13. It is off by default and available to Pro, Business, and Enterprise users. Business and Enterprise administrators have to grant access before each person can choose to turn it on. It is not currently available in the EEA, United Kingdom, or Switzerland.
Computer History creates an event stream from the apps and websites you allow. OpenAI says those events can include clicks, typing, keyboard shortcuts, app switches, and context macOS exposes through its accessibility system. ChatGPT and Codex turn that activity into a searchable timeline and local memory files.
Where this could help
- Recover the document and next step you lost when the day interrupted you.
- Ask what you worked on yesterday and prepare a status update.
- Find the source you remember by subject but not by filename.
- Notice a process you repeat every week and turn it into a Skill or automation.
- Reconstruct the path from a half-formed idea to the files you already started.
What it does not record
OpenAI says Computer History does not capture screenshots, screen recordings, microphone input, or system audio. Private-mode web browsing is excluded.
That does not mean nothing sensitive can be captured. Typing, clicks, app changes, and accessibility context can still reveal a lot.
OpenAI tells people to turn it off during communications unless the other people involved have given prior express consent. It also advises pausing or excluding apps with health, financial, or personal information.
Email, messages, video calls, client portals, banking, payroll, health apps, and password tools do not belong in a casual first test.
Where the information goes
The temporary event stream stays on your Mac for up to 48 hours. Computer History turns parts of it into plain-text Markdown memory files that remain on your computer until you delete or clear them.
Those memory files are inspectable, which I like. They are not encrypted by Computer History itself. Another program running as your macOS user may be able to access them.
OpenAI also processes the temporary event files on its servers to create the memories. The company says it does not retain those files after processing unless legally required and does not use them for training.
If ChatGPT or Codex later pulls one of those memories into a conversation, that conversation follows your normal ChatGPT data controls. If you do not want new consumer-chat content used to improve OpenAI's models, check Settings → Data Controls → Improve the model for everyone.
The security risk most people will not see
OpenAI explicitly warns that Computer History increases prompt-injection risk. A website or app can contain text designed to manipulate an AI that reads it.
OWASP, an independent application-security organization, describes persistent memory as an attack surface because misleading or hostile content can influence later work across sessions.
That does not mean someone has breached Computer History. It means a memory that can carry useful context forward can also carry bad context forward.
- Observe: What can the AI see?
- Remember: What can it carry forward?
- Act: What can it change, send, publish, or run?
Do not make all three broad at the same time.
My decision: Use it, but make it earn every additional source.
Google Drive is now inside ChatGPT Library
A useful connection when it removes repeated uploading and re-explaining.
ChatGPT can now browse connected Google Drive files and folders directly from Library. You can add a Drive file to a conversation with an @mention, keep a Doc, Sheet, or Slide open beside the chat, and select a folder when the job depends on several related files.
Where supported and authorized, ChatGPT can update the original Drive file instead of producing another disconnected copy.
The rollout is reaching Plus, Pro, Enterprise, Edu, Healthcare, and Business users on the web. It currently includes My Drive and files or folders shared directly with you. Shared Drives are not included yet, and some editing and collaboration features are still missing.
Where this could help
- Review a launch folder and create the one missing checklist.
- Compare a proposal, client notes, and a pricing sheet without uploading each file again.
- Keep the source document beside the conversation while you revise it.
- Ask ChatGPT to identify which file contains the decision you remember making.
Start with one active folder and one review-ready output. A connection is useful when it removes repetition. It is not an invitation to give ChatGPT every file you have.
You can change an existing Project's memory boundary
Better memory is not always broader memory.
ChatGPT now lets you switch eligible unshared Projects between Default memory and Project-only memory after the Project already exists.
Project-only memory keeps the boundary tight. ChatGPT can use conversations inside that Project, but it will not pull in memories or conversations from elsewhere. Information from the Project also stays out of memory used in chats outside it.
The control is available on all plans. Shared Projects remain Project-only, and changes may take a few hours. ChatGPT Work is not available inside a Project using Project-only memory.
Use Project-only memory when
- A client or employer's work should not mix with another body of work.
- You want an experiment to use only its own files and decisions.
- Your personal memories would make the answer less accurate.
- You need to know whether the source material is complete enough on its own.
This is the control that should sit beside the excitement about more context. Sometimes the most useful context is the context you deliberately kept out.
Claude will mark text from future models with an invisible watermark
The detection system is not available yet, so there is no reader task today.
Anthropic says every future Claude model will include an imperceptible text watermark from launch. The company also plans to add the marking to older models over time. Supported generated files will receive C2PA content credentials.
Anthropic says the text watermark does not add tokens or change speed, price, readability, or output quality.
The important limit is that Anthropic has not released the detection API or checking tool yet. A watermark is also not proof of who wrote, approved, or published a piece of work.
What to do
Nothing new today. Do not rewrite old work. Do not stop using Claude. Keep following the disclosure rules that already apply to your client, platform, workplace, or type of content. Watch how the detection system works once people outside Anthropic can actually use it.
Gemini's agent layer got broader and faster
Retest one workflow only if it touches a service you already use.
Google made Gemini 3.7 Flash generally available on August 13 and says it now powers Gemini Spark for Google AI Pro and Ultra subscribers in more than 160 countries.
Google also announced another group of services coming to Gemini, including Granola, Otter.ai, Wix, OpenTable, Ticketmaster, Angi, Thumbtack, and Zocdoc. Rollout timing and access vary.
What to do
If Spark or one of those services is already part of your work, retry one workflow that was too slow, too disconnected, or too limited before. Do not switch platforms because of a benchmark or a connector that has not reached your account.
My take
The model can change. Your context still needs to come with you.
Three months ago, most of my serious AI work was happening in Claude.
Since GPT-5.6 Sol came out, I have moved much more of that work back into ChatGPT and Codex. For the work I am doing right now, OpenAI feels back on top to me.
I do not expect that sentence to stay true forever.
Independent benchmarks do not show one model permanently winning every kind of work. The configuration matters. The surrounding product matters. Your own work matters more than any company's launch chart.
That is why I keep building everything to be platform agnostic.
The model can change. My context, decisions, source files, and finished work still need to come with me.
Computer History makes ChatGPT and Codex more useful because they can reconstruct more of the work around the chat. It also makes portable context more important because passive memory can quietly become another reason you feel unable to leave a platform.
I like that Computer History's generated memories are local Markdown files I can inspect. I still would not make them my company brain.
My business context belongs in the Notion Brain, source files, portable context documents, and finished assets that I can use with Claude, ChatGPT, Gemini, or whatever earns more of my work next.
Your model is rented. Your context should be yours.
Put it to work
The 30-minute Computer History boundary test
Use this if repeated explaining and interrupted work are real bottlenecks. Skip it if you cannot create a low-risk test without exposing confidential or other people's information.
Use this when
You have one low-risk task, one app or website, and a real question about whether passive context will reduce repetition.
Skip it when
The work includes client, employer, health, financial, password, or another person's communication data.
Open Settings → Data Controls and decide whether you want new consumer-chat content used to improve OpenAI's models.
Open Settings → Computer history → Permissions. Choose Include only and add one low-risk app or website tied to one unfinished task.
Do the actual task. Do not perform a fake demonstration for the feature.
What was I working on, where is the real source, and what is the smallest next step? Based only on what you observed, name one repeated sequence that could become a Skill or scheduled task, one costly context switch, and one unfinished loop. Separate evidence from inference. Do not build or automate anything yet.
Open the timeline, review the item, and reveal the local memory file. Check whether it recovered useful context, captured anything unexpected, and stayed accurate enough to guide a next step.
Delete the timeline item or clear the test period. Expand to another source only if the value was worth the access.
Done means: You know whether one narrow source saved enough repetition to justify keeping. You have not turned on your entire computer or built a new system around a three-day-old feature.
Free resource
The Computer History Safe Start Guide
Decide whether Computer History fits your work, keep observing, remembering, and acting as separate permissions, and run a low-risk test with a clear stop-or-expand decision.
Quick uses and checks
Useful if they touch your account
USE | Think is now live on ChatGPT Free and Go
The Think control reached Free and Go users on the web on August 14. Use it when a harder request needs more reasoning. Do not upgrade only to get a thinking control.
USE | ChatGPT can search restaurant reservations on every plan
Ask for a location, time, party size, budget, dietary needs, and atmosphere. OpenTable works globally, Resy in the United States, and Yelp in the United States and Canada. Book through the provider.
CHECK IF THIS IS YOU | ChatGPT ads reached five more countries
Logged-in adults on Free and Go in the United Kingdom, Mexico, Brazil, Japan, and South Korea may now see ads. Paid consumer, Business, Enterprise, and Education plans remain ad-free. If affected, review ad personalization once.
CHECK | o3 leaves ChatGPT on August 26
API access is not changing with this retirement. If you still use o3 in ChatGPT, move one real task to the current model before the deadline.
CHECK | The official DALL·E GPT leaves August 30
Download images you want to keep and use ChatGPT Images going forward. User-created GPTs with image generation are unaffected.
No homework required
Safe to leave alone
GPT-5.6 Sol Ultrafast API preview
Relevant to developers evaluating API speed. It does not change the ordinary ChatGPT task on your screen.
OpenAI's Daybreak cyber models and AWS access
Important platform and security news without a practical reader assignment this week.
Pixel 11 hardware and Pixel-first Gemini Intelligence
Know it exists if you are buying a phone. Do not reorganize your stack around a device launch.
Astra math and cyber-safety reporting
Large capability and safety stories. There is no ordinary-reader action today.
Leadership, finance, and infrastructure news
Anthropic and OpenAI company news may shape the field later. It does not belong in Monday's workload.
Model-race claims
A benchmark result matters only when it changes a real decision in your work. Keep your context portable and let the models keep moving.
Source note: The Computer History facts come from OpenAI's public documentation and release notes. The risk framing adds OWASP's independent work on persistent-memory security. A general memory-security risk is not evidence of a known Computer History exploit. OpenAI's statements about retention and training are identified as company claims.
What I'm watching
Whether Computer History continues to recover meaningful work after the novelty wears off, and whether its most valuable contribution is retrieval, honest reflection, or finding workflows I should stop carrying in my head.
One question
What is the one low-risk source you would let Computer History observe first?