Why Does Microsoft Copilot Forget What You Told It?

Copilot usually appears to forget for one of four reasons: the memory feature has not reached your account yet, you are on a different Copilot than the one the answer you read was written about, the detail you expected was inferred rather than saved and Copilot discarded it, or a control was switched off in a way that stops memory being applied without deleting it.

The useful question is: which Copilot are you using, and was the thing it forgot saved on purpose or inferred? Those two answers explain almost every case, and Microsoft's own admin documentation is unusually specific about both.

Last verified: September 10, 2026.

Key takeaways

• Copilot personalization and memory are documented by Microsoft as in preview and subject to change, so behavior can move under you.
• Memory covers three different things: saved memories, details inferred from chat history, and custom instructions. They delete differently.
• Memories are stored in your Exchange mailbox, in a hidden folder, which is why mailbox compliance controls apply and Purview retention policies do not.
• Deleting a chat does not delete saved memories generated from it.
• Turning memory off stops it being applied but does not remove it, except for the inferred layer, which is deleted after 30 days.
• There is no way to force-enable the feature if the rollout has not reached you.

Which Copilot are you using?

This is the first question because "Copilot" is four products sharing a word, and a support answer written for one does not necessarily describe another.

• Consumer Copilot at copilot.microsoft.com, tied to a personal Microsoft account.
• Microsoft 365 Copilot, tied to a work or school tenant and governed by that tenant's admin settings.
• Copilot in Windows, surfaced through the operating system.
• GitHub Copilot, a separate developer product with its own context rules entirely.

Microsoft's memory documentation names the control differently depending on where you are: consumer Copilot exposes a personalization and memory toggle, while Microsoft 365 Copilot calls it Enhanced Personalization. If you searched for why Copilot forgot something and landed on advice that does not match your settings screen, the mismatch is usually this.

Does Microsoft Copilot remember previous conversations?

Microsoft says it does, and prints a caveat at the top of the admin guide: "Copilot personalization and memory are in preview and subject to change." The Manage Copilot personalization and memory page, last updated September 2, 2026, is the closest thing to a published contract, and it says Copilot memory is available to Copilot Chat users "with and without a Microsoft Copilot license."

That page groups three separate things under the single word "memories": saved memories that you create deliberately, "details inferred from chat history" that Copilot derives on its own, and custom instructions that act as standing preferences.

The distinction matters more than it sounds, because the three behave differently under every control you might reach for.

Why does Copilot forget something it clearly knew?

Because the inferred layer is designed to change. Microsoft writes that "details from the chat history are dynamic," and that "Copilot might update or discard older details as it learns what's most helpful to retain."

Discarding is documented behavior, not a bug you can file. If the fact you expected Copilot to hold was never saved explicitly, it lived in the layer Microsoft reserves the right to prune. The practical implication is simple: anything you actually depend on should be saved deliberately or written into custom instructions, not left to inference.

This is also why "it remembered last week and forgot today" is a coherent outcome rather than a contradiction.

Why do I not have Copilot memory at all?

Because the rollout may not have reached your account, and there is nothing you can do to hurry it.

On Microsoft's own Q&A, a user asked in January 2026 why the memory feature was not present in their Copilot or Copilot 365. The answer from Microsoft staff the following month was direct: "there's no way to install or force-enable it—it will appear automatically in your settings once your account is included in the rollout."

That same answer is worth reading for expectation-setting as much as for the remedy. It describes the feature as remembering things primarily when you explicitly ask, or inferring interests from activity, rather than as full conversational memory that automatically carries everything across sessions.

If the toggle is absent from your settings, you are not misconfigured. You are waiting.

Where does Copilot keep what it remembers?

In your mailbox. Microsoft states that memories "are stored in the user's Exchange mailbox in a hidden folder," and that they therefore "follow the same security and compliance policies as other mailbox data," naming Customer Lockbox and encryption at rest as examples.

Two consequences follow from that storage choice.

The first is discoverability. Memory can be searched and exported through eDiscovery and Microsoft Graph Explorer using the item class IPM.Contact, with the memory sitting in a CopilotMemory folder. Microsoft notes that ordinary contacts share that item class, so an unfiltered export returns both.

The second is a gap. Microsoft says memory and personalization actions "don't generate audit log entries in Purview." The stored result is discoverable after the fact; the act of storing or applying it is not logged.

Which controls actually delete Copilot memory?

This is where most of the surprise lives, and the three memory types diverge sharply.

Saved memories persist until you remove them yourself. Microsoft says they "are retained until the end-user explicitly deletes the saved memory in Settings > Personalization," and then adds the sentence that catches people out: "Deleting a chat doesn't delete saved memories that were generated from that chat."

Turning the feature off is not an erase. With the admin-level or user-level control switched off, Microsoft says Copilot "is stopped from applying" custom instructions and saved memories, "but doesn't remove" them.

The inferred layer is the exception. With the control off, "all details from chat history are deleted after 30 days." And if you delete every chat where a piece of information was originally shared or referenced, Microsoft says "all traces of that information are deleted from the Chat History details within seven days" — which reaches only the inferred layer, never anything you saved on purpose.

If your goal is a clean slate, clearing history alone will not get you there.

What your administrator can and cannot control

If you are on a work tenant, two common assumptions do not hold.

Enhanced personalization, the control that allows Copilot memory to be used, is on by default. Microsoft is explicit: "No action is required to turn on Copilot memory; you only need to turn off Copilot memory for end-users or the tenant if desired." Where it is switched off, users see the personalization toggles disabled and "can't turn on these settings."

Retention is the first broken assumption. Microsoft says "retention policies and retention labels configured in Purview by organization admins don't apply to Copilot memory," and gives the example that a delete-after-three-months policy on Copilot Chat interactions will not delete the memories. The page states there are no admin controls to enforce retention specifically for Copilot memory.

Content scoping is the second. Asked whether an admin can restrict what information is added to memory, Microsoft answers: "No, admins can't restrict what type of information is added to Copilot memory."

A Copilot memory checklist

Run this before concluding the feature is broken.

1. Identify the surface. Consumer Copilot, Microsoft 365 Copilot, Copilot in Windows, or GitHub Copilot.
2. Confirm the feature exists for your account, from the profile icon under settings. If the toggle is absent, you are waiting on the rollout.
3. Classify what was forgotten. Saved deliberately, or inferred from chat history? Only the first is durable by design.
4. Check whether a control is off at the user or tenant level, remembering that off means "not applied," not "not stored."
5. Promote anything load-bearing into a saved memory or a custom instruction rather than relying on inference.
6. If you need it gone, delete the saved memory explicitly. Clearing chats reaches only the inferred layer.

How Fostera treats the same problem

Fostera builds continuity around a named Soul rather than around a thread or a mailbox, and the design goal is that memory is something you can read.

Retained entries are listed individually in a memory browser where you can inspect, edit, and delete them. Retrieval is relevance-based and can still miss a detail, which is exactly why Custom Rules exist as an explicit home for durable voice, canon, or working preferences instead of leaving them to be inferred and later pruned. Deleting a memory removes it rather than pausing it.

Fostera also routes automatically among eligible engines, with manual selection available as a Premium feature, so a Soul's retained context does not need rebuilding when the engine behind a reply changes. Connected provider terms vary, and memory deletion cleanup may take up to 30 days. A side-by-side of the two models is on the Copilot memory comparison page, and if you are weighing engines more broadly, which AI model should I use covers the trade-offs.

Frequently asked questions

Does Copilot remember previous conversations?

Microsoft documents a memory feature covering saved memories, details inferred from chat history, and custom instructions, and describes it as in preview and subject to change. The inferred layer is explicitly dynamic — Microsoft says Copilot might update or discard older details as it learns what is most helpful to retain.

Why does Copilot forget within the same session?

A single reply is assembled from whatever context that surface supplies, and long threads can be trimmed or summarized before the model sees them. If the detail was in the inferred memory layer rather than saved, Microsoft's documentation allows it to have been discarded.

Can I see or edit what Copilot has stored?

Saved memories appear in personalization settings, where Microsoft says they remain until you explicitly delete them. Custom instructions can be exported manually by the user from the same settings area, though Microsoft notes they are not yet discoverable through eDiscovery or Content Search.

Is Microsoft 365 Copilot memory the same as consumer Copilot memory?

They are different surfaces with differently named controls — Enhanced Personalization in Microsoft 365 Copilot, a personalization and memory toggle in consumer Copilot. Copilot in Windows and GitHub Copilot are separate products again, so guidance about one may not describe another.

How do I delete what Copilot remembers about me?

Delete the saved memory explicitly in Settings > Personalization. Microsoft is clear that deleting a chat does not remove saved memories generated from it, and that switching the feature off stops memories being applied without removing them.

Will Copilot memory get better?

Microsoft labels it a preview feature subject to change, which cuts both ways: capability may improve, and documented behavior may move. Treat anything you depend on as something to record deliberately rather than something to assume will persist.

Fix the context layer, not just the next prompt

Copilot forgetting is rarely a single failure. It is a preview feature, split across four products, storing three kinds of memory in a mailbox, with controls that pause some of it and delete other parts on different clocks.

Once you know which surface you are on and whether the detail was saved or inferred, the behavior stops being mysterious. What remains is a decision about where context that matters should live — and whether you want to be able to open it and read it.