You tell Microsoft 365 Copilot you prefer Python, that you are running Project Alpha, that you write in a plain style, and weeks later it still knows. That memory is not sitting in some abstract AI cloud. It lives in a hidden folder inside your own Exchange Online mailbox, named CopilotMemory, tucked beside your address book and tagged with the exact item class Outlook uses for a contact card.
Most coverage of Copilot memory stops at the marketing line: it remembers your preferences so you repeat yourself less. The part nobody screenshots is the plumbing. The memory is a mailbox object. It is discoverable in eDiscovery. And deleting the chat that created it does not delete the memory. Those three facts change how you should think about what Copilot keeps on you, and they are the difference between feeling clever about clearing your history and actually clearing anything.
Copilot memory lives in a hidden mailbox folder called CopilotMemory
Memories are stored in the user's Exchange mailbox in a hidden folder. Microsoft says so directly in its admin documentation, the Manage Copilot personalization and memory article on Microsoft Learn. The folder is named CopilotMemory, it sits inside your Contacts, and the items inside it carry the item class IPM.Contact, the same class Outlook assigns to the people in your contacts list. That overlap is not cosmetic. When an administrator exports your IPM.Contact items, your Copilot memory comes out in the same pile as your real contacts and has to be separated out by folder.
Because the data sits in the mailbox, it inherits the mailbox's protections rather than getting its own. Microsoft confirms Copilot memory follows the same security and compliance policies as other mailbox data, including Customer Lockbox and encryption at rest. No separate vault. No separate key. Your memory is mail-shaped, governed by mail rules, and it rides along with whatever your Exchange Online tenant already enforces.
Your Copilot memory is not a setting. It is a contact card in a hidden folder in your mailbox.
Memory has three parts, and only some of it is intentional
Copilot memory bundles three different things: saved memories you confirmed, details inferred from your chat history, and custom instructions you typed. Saved memories and inferred details live together in that mailbox folder. Custom instructions are stored too, but they behave differently for export and discovery, which matters a great deal once an admin or a legal team comes looking.
What kinds of things end up saved? Microsoft's own launch post groups them into a few buckets: technical preferences like a default language, communication styles such as a formal tone in email, project context like the fact that you are on Project Alpha, and creative preferences for things like image generation. Copilot memory reached general availability in July 2025, so this is not a future-tense feature. If your tenant runs Copilot Chat, it is already deciding which of these about you are worth keeping.
The interesting filter is what makes the cut. Copilot uses what Microsoft calls intent-driven storage. A durable preference gets remembered. A one-off task does not. The official example is exact: "I prefer Python for all data science tasks" is remembered, while "Write Python code for k-means clustering" is not. The first is a standing fact about you. The second is a job that ends when the code ships.
If you want Copilot to forget something, do not phrase it as a stated preference. Standing facts about you are exactly what the intent filter is designed to keep.
The point of that filter is to stop the memory from filling with noise. A model that wrote down every line you ever typed would drown in dead context: the function you debugged once, the date you needed last Tuesday, the city you booked a single flight to. Intent-driven storage keeps the kept set small and durable. The tradeoff is predictability. You cannot always tell, in the moment, whether a sentence reads to Copilot as a passing instruction or as a fact about you, and the system does not ask before it decides. That is the reason it pays to know where the result lands.
One more thing surprises people who assume memory is a premium toy. Copilot memory is not limited to people who pay for the add-on. Microsoft makes it available to Copilot Chat users both with and without a Microsoft 365 Copilot license. Some of the broader Frontier early-access features do require a license and a tenant opt-in, but the core memory behavior described here reaches a wide set of work accounts. If your organization touches Copilot Chat at all, this storage model very likely applies to you.
Deleting the chat does not delete the memory
This is the gotcha. Microsoft states plainly that deleting a chat doesn't delete saved memories that were generated from that chat. You can wipe your conversation history and still leave behind a saved memory of what you said in it. Saved memories persist until you explicitly delete each one in Settings, under Personalization. There is no bulk clear tied to your chat log, and no automatic expiry on the saved layer.
Inferred details from chat history behave differently. They are dynamic, so Copilot may update or discard them as it learns what is worth keeping. If you delete every chat where a piece of information was originally shared or referenced, Microsoft says all traces of that information are removed from the chat-history details within seven days. There is a second timer too: if the Chat history control is turned off, at the admin level or by you, all details from chat history are deleted after 30 days. So the inferred layer eventually follows your deletions. The saved-memory layer does not. It waits for you to act on each item by hand.
| Memory type | Where it lives | How it gets removed |
|---|---|---|
| Saved memories | CopilotMemory folder, item class IPM.Contact | Manual delete per item in Settings; deleting the source chat does not remove them |
| Chat-history details | CopilotMemory folder, dynamic and updated over time | Purged within seven days of deleting all source chats; deleted after 30 days if the control is turned off |
| Custom instructions | Stored, but not yet discoverable in eDiscovery or Content Search | Manual delete; turning off the control stops use but does not remove them |
An administrator can read your Copilot memory through eDiscovery
Saved memories and inferred memories are discoverable using Microsoft Purview eDiscovery and Microsoft Graph Explorer. An admin searches the item class IPM.Contact, then narrows to the CopilotMemory folder, because your real contacts share that same class. From there they can search, export, and delete memory data. Custom instructions are the exception: they are not yet discoverable in eDiscovery or Content Search, though a user can export them by hand from Settings.
Three more constraints are worth knowing, and they cut against the usual governance instincts. Purview retention policies and retention labels do not apply to Copilot memory, so a "delete after three months" rule on Copilot chats will not age out the memories those chats produced. Microsoft is blunt about it: there are no admin controls to enforce retention rules specifically for Copilot memory. Memory and personalization actions do not generate audit log entries in Purview, so there is no built-in trail of who read or changed a memory. And an admin cannot restrict what types of information land in memory in the first place; the filter is the model's, not the tenant's.
A retention policy that deletes your Copilot chats after three months does nothing to the memories those chats created. They outlive the conversation.
For a data subject request, this combination is the practical recipe. An admin runs an eDiscovery or Content Search across the IPM.Contact item class to find the memory, exports it for review, then uses Microsoft Graph Explorer together with Purview to delete it. Temporary chats are not a loophole, either. Microsoft confirms temporary chats are available in Purview the same way normal chats are. The one corner the tooling does not yet reach is custom instructions, which an admin cannot discover and which the user has to export themselves.
How to view, manage, and turn off Copilot memory
As a user, open Settings and go to Personalization. You see three controls there, Custom instructions, Saved memories, and Chat history, each of which you can turn off, plus a list of saved memories you can delete one at a time. Turning a control off stops Copilot from applying that memory to new responses. It does not erase what is already stored. To actually clear a saved memory, you delete that item directly. Knowing the difference between off and gone is the whole point.
At the tenant level, the master switch is the Enhanced personalization control, and by default it is turned on. No action is required to turn Copilot memory on; an admin only acts to turn it off. When the control is off, end users see their Personalization settings as disabled and cannot switch them back on. Admins can configure all of this programmatically through Microsoft Graph, using the enhancedPersonalizationSetting resource type, so a tenant can script the policy rather than click through it. As of June 2026, Copilot personalization and memory are still in preview under Microsoft's Frontier program and subject to change.
What this means if you ever leave the tenant
Everything above describes a memory bolted to one mailbox in one Microsoft 365 tenant. It is invisible outside that tenant, it cannot be queried by anything except Microsoft's own tooling, and it does not travel with you. Switch jobs, change tenants, or move part of your work to Claude or ChatGPT, and the context Copilot built about you simply stays behind in an Exchange folder you can no longer reach. There is no export-and-import-elsewhere path, because the memory was never a portable file to begin with. It was a contact card in someone else's directory. The convenience was real. The ownership was never yours.
MemX takes the opposite design. It is a memory layer you own and can export, sitting outside any single vendor and feeding the AI you actually use, whether that is ChatGPT, Claude, or Gemini. It is private by architecture: per-user isolation, encryption at rest, and your memory is not used for training. Your context belongs to you, not to whichever mailbox happens to hold it this year.
Frequently asked questions
01Where does Microsoft 365 Copilot store memory?
In the user's Exchange Online mailbox, inside a hidden folder named CopilotMemory that sits within Contacts. The items use the item class IPM.Contact, the same class Outlook uses for contact cards, so memories sit alongside your address book.
02Does deleting a Copilot chat delete the memory?
No. Microsoft states that deleting a chat does not delete saved memories generated from it. Saved memories persist until you delete each one in Settings under Personalization. Only chat-history details are purged, within seven days of deleting all source chats.
03Can my admin see my Copilot memory?
Yes. Saved and inferred memories are discoverable through Microsoft Purview eDiscovery and Microsoft Graph Explorer by searching the IPM.Contact item class. Admins can search, export, and delete them. Custom instructions are not yet discoverable this way.
04How do I turn off Copilot memory?
Open Settings, go to Personalization, and turn off Custom instructions, Saved memories, or Chat history. Turning a control off stops Copilot from using that memory but does not erase stored items, which you delete separately.
05What decides whether Copilot remembers something?
Intent-driven storage. A standing preference like "I prefer Python for all data science tasks" is remembered. A one-off request like "write Python code for k-means clustering" is not, because it ends when the task does.
The takeaway
Copilot memory is not a black box. It is a hidden CopilotMemory folder in your Exchange mailbox, tagged IPM.Contact, discoverable by your admin in eDiscovery, exempt from your retention policies, and stubborn enough that deleting a chat leaves the memory behind. Know where it lives, delete saved items by hand when you mean to clear them, and remember that none of it follows you out the door.
