OpenAI dots launched on September 29, 2026, and the line in the help center that stopped me was not about speed or the 4,000 connected apps. It was this: you cannot view, delete, or edit individual dot memories today. I pay for ChatGPT every month and use it weekly for short reframes and email headlines, but I have not set up a dot, and that one sentence is the reason. For a solo consultant holding confidential material from several clients, OpenAI dots raise a memory question before they raise a productivity one.
In this article
- What OpenAI dots actually are, in plain terms
- Why the OpenAI dots memory model matters more than the model behind it
- How Auto-review and Custom Rules split the control
- What the first-month pricing does and does not tell you
- Five checks before any client data goes near a dot
OpenAI Dots Are Always-On Agents With Their Own Computers
The short version is that a dot is a persistent agent rather than a chat. OpenAI describes dots as always-on agents, each running on GPT-6 Astra and working from its own cloud computer and browser. You give it a goal, connect the apps it needs, decide what it may do on its own, and it brings results back for review.
That is a different shape from anything in my current stack. Claude, Notion AI and Claude Code all wait for me to start something. OpenAI dots are built to keep going while you are away, and you can reach one from ChatGPT on desktop, web and mobile, or from Slack and Teams.
A few launch facts frame the rest of this piece:
- Rollout is to ChatGPT Pro and Business Premium subscribers, and some accounts may wait several days
- Pro access excludes the European Economic Area, Switzerland and the UK at launch
- Dots can reach more than 4,000 apps through OpenAI’s plugin ecosystem
- Local computer access is optional and starts turned off
- Setup happens on desktop or web; you cannot create a dot from mobile
I covered the model underneath in my checks on GPT-6 Astra before its wider rollout, so I will not re-argue capability here. The interesting part of OpenAI dots is not what they can do. It is what they keep.
The OpenAI Dots Memory Model Is the Real Product Decision
Memory, not intelligence, is the part of OpenAI dots a consultant has to price first. According to OpenAI’s help center, a dot inherits your existing ChatGPT memories and builds new ones from the apps you connect. When you are not working with it, the dot runs what OpenAI calls proactive research over those connected apps, using read-only access.
Then come the two sentences that matter. You currently cannot view, delete, or directly modify individual dot memories. And removing an app stops new access but does not delete what the dot already learned from it. The only full undo OpenAI documents is Reset, which deletes the dot along with its conversations, saved memories and scheduled tasks.
Think through what that means for client work. If I connected a shared folder for a B2B SaaS founder, finished the engagement, and disconnected the folder, whatever the dot had absorbed from it would still be there. My only clean exit would be to wipe the agent and every routine I had built on top of it.
That is not a scandal. It is an early product that has not built granular memory controls yet, and OpenAI is candid about it. But it inverts the usual habit, where disconnecting an integration feels like closing the door. With OpenAI dots, disconnecting closes the door and leaves the furniture inside.
With OpenAI dots, disconnecting an app closes the door. It does not take back what the agent already learned in the room.
Auto-Review and Custom Rules Split the Control in Two
The control model is more careful than the memory model, and it deserves credit. Actions that could affect your accounts or share information are checked by Auto-review against your instructions, your Custom Rules and OpenAI’s safety requirements before they run. The help center’s own example is an email: Auto-review checks the recipient and message to catch a wrong address or an unintended disclosure.
Custom Rules are where you shape what the dot does alone and what it hands back. OpenAI’s setup material describes four behaviours you can assign to an action: take it without asking, take it if pre-approved, ask first, or hand it off to you. Some tasks, such as password changes, always stay with the user.
The limits on Custom Rules are worth reading closely. They cannot turn off core safety requirements, cannot change Auto-review, and cannot lift the restrictions on proactive research. In other words, you can make a dot more cautious, but not less. For a one-person operation that is the right default, because nobody else is going to catch the mistake.
Two smaller details also lean in the user’s favour. OpenAI says credentials go directly to the dot’s browser environment without being exposed to the model, and a monitoring system can pause or stop a dot over safety concerns. OpenAI still tells users to review consequential work because the agents can make mistakes, which matches what I found when I looked at the ChatGPT Work agent earlier this year.
First-Month Pricing Tells You Little About Month Two
The launch offer is generous, and it is also temporary. Your first dot is included at no extra cost on Pro and Business Premium, and dot usage does not count against eligible plan allowances for the first month after launch.
What comes after is not published yet. Ordinary dot conversations do not count toward ChatGPT usage limits, but deeper background work draws on a plan allowance whose post-launch terms OpenAI has not disclosed. Additional dots, plus more speed or monthly capacity for existing ones, are planned at prices that have not been announced.
That gap matters for anyone tempted to build routines during the free window. My rule from earlier pieces still holds: a paid tier has to return roughly three times its cost in recovered hours. You cannot run that maths on OpenAI dots yet, because the second number in the equation does not exist.
On training, the defaults depend on the plan rather than on dots specifically. OpenAI does not use content from Business, Enterprise or Edu workspaces for training by default. On personal plans, including Pro, the “Improve the model for everyone” toggle in ChatGPT’s data controls decides. If you are on Pro, check that setting before connecting anything.
Five Checks Before Client Data Goes Near a Dot
My conclusion is that OpenAI dots are worth watching and not yet worth trusting with client material. If I set one up next month, these are the checks I would run first, in this order:
- Start the dot with no client connectors at all. Use it for public research and personal admin until you have seen what it remembers.
- Remember that it inherits ChatGPT memory. Review and prune your existing ChatGPT memories before the dot copies them.
- Treat every connected app as permanent. Since disconnecting does not delete learned material, only connect what you would be comfortable resetting the whole agent over.
- Set every outbound action to “ask first.” Auto-review is a safety net, not a substitute for reading the email before it leaves.
- Write down what Reset would destroy. If a client offboarding ever requires a clean wipe, you should know in advance which schedules and routines go with it.
None of this is unique to OpenAI dots, and the same logic sits behind the rules I already apply to keeping client data local. What is new is the persistence. An agent that keeps working while you sleep also keeps remembering while you sleep, and that is the trade OpenAI dots are asking you to make.
For me, OpenAI dots stay on the watch list rather than in the stack. The permission design is thoughtful, the free first month is real, and a proactive agent in Slack is the kind of thing I could see helping with research triage. But until I can view and delete individual memories, client work stays with Claude, Notion AI and my local models. When OpenAI ships granular memory controls, I will run the five checks above and report back.
Sources
AI-assisted research and drafting. Reviewed and published by ToolMint.