Key takeaways
- AI agents from Google, OpenAI and others now keep what they learn as they work, and the vendors' own documentation gives different answers on whether your company can read that learning or take a copy. OpenAI says people using its new ChatGPT agents, called dots, currently can't view the agents' individual memories, while Microsoft lets administrators search and export users' memory in Microsoft 365 Copilot, in preview.
- For one workflow where an agent is already learning, have the workflow's owner and one IT person answer three questions per agent: do you know where what it learned is kept, can someone at the company read it, and can you get a copy someone could use. Treat any unanswered question as a no until someone shows it working, and let an agent with a no help with the job without becoming the only way it gets done.
- Keep a dated record of how to do the work in storage your company controls, and test it by having another employee do one real task from it without the agent. A pass shows that task can be done without the agent, and anything the agent learned that you can't read or export still stays with it.
In this article
Your operations team probably runs on rules nobody wrote down: this client is invoiced in euros, that supplier’s delivery notes always arrive late so there’s no point chasing them before Thursday, refunds above a set amount go to the finance lead first.
Those rules usually live in people’s heads. Now an AI agent can learn them too, one with its own email address and a place in the company directory, and when the team corrects it, the agent can keep the correction. That’s how it gets better at the work.
Corrections like these are often given in passing, in a chat or a comment, and then nobody writes them down anywhere else.
So before it learns much more, three things are worth knowing: where it keeps what it learned, whether anyone at your company can read it, and whether you could take a usable copy with you. Every correction it keeps from here on is one more piece of the job that may live only there.
The new colleague keeps notes
On October 8, 2026, Google described a coworker agent that “receives its own Workspace account, including an email address, calendar, Drive, and presence in your company directory” (Google Cloud blog, October 2026). Google says Gemini, the agent behind it, keeps four kinds of memory, one of them “procedural memory for how a job gets done, including skills it writes for itself.”
Nine days earlier, OpenAI launched dots, personal agents in ChatGPT that “learn from feedback over time” (OpenAI, September 2026). According to OpenAI’s help pages, a dot “keeps its own notes about preferences, decisions, and ongoing work,” and those notes are separate from ChatGPT’s regular memory.
What sits in those notes and skills can include the kind of know-how in the rules above. Plenty of organizations know what happens when that kind of know-how lives with one person or inside one tool, and then the person leaves or the tool is replaced.
OpenAI, Google, Microsoft and Anthropic
We read what four vendors publish about this. We didn’t test the products, so what follows is what their documentation said on October 11, 2026, and several of these features are still in preview or beta.
OpenAI is the most direct. Its dots privacy FAQ says: “You currently cannot view, delete or directly modify individual dot memories.” The one way the FAQ gives to clear what a dot has picked up is to delete the dot.
Its admin guide, meanwhile, tells administrators to “Review saved memories and reset options” when someone leaves, without saying how. We found no documented way for a company to export a dot’s notes. In Business and Enterprise workspaces, export of ChatGPT data goes through the workspace owner rather than self-service, and nothing we found says it includes those notes.
Google documents controls for the regular memory in Gemini Enterprise: users can review and remove what they’ve saved, and administrators can switch memory off for everyone. Skills that people write themselves can be suspended or deleted by an administrator.
But we found nothing in Google’s documentation on how an administrator reads or exports the procedural memory Google’s agent writes for itself. The launch post does say “your context, your skills, and your data stay put,” and that sentence is about Google swapping the model underneath. It says nothing about leaving Google.
Microsoft goes furthest on access. Memory in Microsoft 365 Copilot sits in a hidden folder in each user’s mailbox, and administrators can “search, export, and delete users’ memory data” with eDiscovery, Microsoft’s compliance search tool (Microsoft Learn, preview). Custom instructions, the standing instructions a user gives Copilot, are the exception, and only the user can export those.
In Copilot Studio, where companies build their own agents, a user’s memories are private to that user. The person who built the agent can’t see them, and they’re deleted after 28 days without use.
Anthropic sits in between. In the Claude app, organization owners “can’t view or edit a user’s individual memories,” yet memory entries “are included in standard conversation history exports.” Skills are plain folders with an instruction file, and when Claude builds one from a screen recording, it proposes the skill “for you to review before you save it.”
So what each vendor calls memory comes with a different answer, and the answer can change between two products from one vendor. The question isn’t whether to let agents learn. It’s whether you know which answer you’ve got, agent by agent.
Memory of this kind lives outside the model’s context window, the text it can see while it answers, so a tool has to put it back there for each new conversation.
Now the agent writes them down
In June 2026 we argued that the corrections and definitions a company feeds its AI are an asset worth keeping portable, in plain files you can carry out the door. That advice assumed your people write those definitions down.
With agents like these, the agent writes them, and sometimes in a place your company may not be able to open. Giving an agent its own identity makes it easier to see what it did. In Google’s version it also gets its own persistent storage, and nothing we found in Google’s documentation says whether its learning is kept there.
Start with one workflow
The steps below are ours, built from what the documentation says. The vendors don’t prescribe them, and we haven’t tested them in a company. Two people run the check, one more employee does the handover task, and you read one sheet.
-
Pick one recurring workflow where an agent remembers or writes skills. One is small enough to finish, and it shows you what the answers look like before you take on the rest. Name the manager who owns it, who knows which rules and exceptions matter, and one person from IT, who knows what an administrator can reach.
-
Answer three questions for each agent, about what it learned from the work. Label each agent with the exact product and plan, because the answers differ between them.
- Do we know where it’s kept, and under whose account?
- Can someone at the company read it?
- Can the company get a copy someone could use?
A partial answer isn’t a yes, and a chat history export counts only if it holds the learning itself. Anything unanswered or contradictory counts as a no until someone shows it working, and goes on the sheet as “unresolved” rather than “can’t”, one line per agent with the three answers, the date and the open question.
-
Let the answers decide how much of the job the agent carries, and keep a dated record either way. If all three answers are yes, the agent can take on more of the work. If any answer is no, the agent can still help, but a named person stays able to do the job, and the agent doesn’t get the next workflow until the answer changes.
Either way, the owner writes a dated record of how to do the work, starting from the product’s export where there is one. It holds the rules, the corrections the team accepted, the exceptions, two or three examples of good output and any template or file the work needs, and it lives in storage the company controls. Whoever accepts a correction, in a chat, a comment or a meeting, sends it to the owner, who adds it to the record the same week.
-
Run a handover check. It comes after the record because it tests the record, since a file can hold the words and still miss what someone needs to do the job. Another employee does one real task, ideally one with an exception or a judgment call in it, using the record and their normal tools and data, without the agent or help from the owner, and the owner checks the result. If it fails, the owner adds what was missing to the record and the check runs again.
-
Check again before the moments that erase memory. Someone leaves or an account is removed, memory is reset or switched off, or you move to another tool: in Claude, switching memory off for an organization deletes every user’s memory immediately. On a schedule you choose, compare what the agent has learned, where you can see it, against the record, and keep what’s useful.
Before you lean on the record
A record your manager writes can cover what the team knows. Anything the agent learned that you can’t read or export still stays with the agent.
One handover task shows that one task can be done without the agent, and only that one. And nothing here settles who owns an agent’s memory under your contract, which is a question for whoever reviews your vendor agreements.
The sheet on your desk
What comes back to you is small. The sheet tells you, agent by agent, how much of the job it can carry. The handover result tells you whether someone could still do that task on Monday if the agent disappeared tomorrow.
No vendor’s documentation can give you that second answer, because it’s about your team.
Questions this article gets
Does this mean we shouldn't let AI agents learn?
No. Learning is what makes an agent better at the work. The check decides how much of the job you let rest on learning you can't see, and makes sure there's a record of how to do the work that your company controls, whatever the agent keeps.
Isn't an export of our chat history enough?
Only if it contains the learning itself, in a form someone can read. OpenAI, for example, keeps a dot's notes separate from ChatGPT's saved memory and says they aren't a complete transcript, so check what an export actually holds before you count it.
Who should run the check?
The manager who owns the workflow and one person from IT. The manager knows which rules and exceptions matter, and IT knows the accounts, the settings and what an administrator can reach, while one more employee does the handover task and the CEO reads one sheet.
