Field note
ChatGPT Memory for Work: Privacy Questions
Memory in ChatGPT is useful context, not a harmless convenience. Decide what may persist, who controls it and how a user removes it before staff rely on it for work.
Memory in ChatGPT is useful context, not a harmless convenience. Decide what may persist, who controls it and how a user removes it before staff rely on it for work.
What memory does
ChatGPT can remember details across conversations. At the time of writing there are two parts: saved memories, which are specific facts the model has decided to keep, such as your role, your preferred writing style or a project name, and a wider reference to past chat history that shapes answers without a visible list. The exact behaviour and the controls differ by plan and change over time, so check the current OpenAI help pages rather than a screenshot from last year.
Memory sits in the user's own account. Another user in the same workspace cannot read it. A temporary chat does not use or create memories. Users can view, delete or turn off saved memories in the personalisation settings, and administrators of Team and Enterprise workspaces have a workspace-level control over whether memory is available at all.
That is the product behaviour. The privacy question is what your staff put into it.
What the risk is
Memory turns a series of throwaway chats into a persistent profile. The profile can contain things nobody meant to store: a customer's name and complaint, a colleague's health situation mentioned while drafting a message, a salary figure from a spreadsheet pasted for analysis, or a supplier's confidential pricing.
Three things then go wrong.
- It resurfaces. A detail saved during one task appears in an unrelated answer weeks later, sometimes in a document that is then shared externally.
- It outlives the reason for holding it. The customer's complaint was resolved. The saved memory is still there.
- It is held in a place your business does not manage. A personal ChatGPT account used for work keeps its memories after the person leaves. Nobody at the company can see or delete them.
Under UK GDPR, personal data in a saved memory is still personal data that your business is responsible for if it was entered in the course of work. "The tool remembered it" is not a defence.
A realistic example
A twelve-person lettings agency lets staff use ChatGPT to draft tenant correspondence. One negotiator pastes a full email thread about a rent arrears case to get a firm but polite reply. Memory saves the tenant's name, the arrears figure and a note that the tenant is disputing a deposit deduction.
A month later the same negotiator asks for help drafting a general newsletter. The draft includes a line about "handling deposit disputes sensitively, as with recent arrears cases" and, in a later prompt, the model refers back to the tenant by name. The negotiator catches it before sending. Nobody else would have known.
The agency did nothing dramatic. It simply had no rule about what could be pasted, and no habit of checking what had been remembered.
Controls that work for a small business
- Use a business workspace. ChatGPT Team or Enterprise gives you a workspace, admin controls and business data terms. Personal accounts on a company domain should be invited in or closed.
- Decide the memory setting at workspace level. If your work involves customer or staff personal data in most chats, consider turning memory off for the workspace and turning it on only for named roles that benefit.
- Give a plain rule about what may be pasted. Customer identifiers, staff personal data, credentials and full documents need a business reason. Redact names where the task does not need them.
- Teach the temporary chat. For anything one-off and sensitive, a temporary chat does not feed memory. Make it the default for HR and complaints work.
- Schedule a memory review. Once a quarter, each user opens the saved memories list and deletes anything that is personal data, out of date or not needed. Five minutes is enough.
- Cover leavers. When someone leaves, the workspace admin removes the account. If they used a personal account for work, ask them to delete work-related memories and chats before their last day and record that you asked.
- Check connected apps. Memory and connectors compound each other. A connector that can read Drive plus a memory that recalls what it found is a wider footprint than either alone. Our note on ChatGPT HR data boundaries covers the HR side in more depth.
What to write into policy
Keep it short enough that people read it:
- Which plan and workspace staff must use for work, and that personal accounts are not permitted for company data.
- Whether memory is on or off by default, and who can ask for an exception.
- The categories of information that must not be pasted without a documented reason.
- The requirement to use a temporary chat for one-off sensitive tasks.
- The quarterly memory review and who checks it happened.
- The leaver step.
Add memory to your record of processing activities as a place where personal data can persist, with the retention rule you have chosen. If you would like a second pair of eyes on that, an AI governance review includes the memory and connector settings for the tools you actually use.