Microsoft Copilot Memory Citations: See and Delete Personalization Inputs in Copilot Chat

According to Microsoft 365 Roadmap entry 574111, Microsoft has launched a transparency update for Microsoft Copilot (Microsoft 365) on the web: Copilot response personalization – Memory citations. Listed as generally available for Worldwide (Standard Multi-Tenant) cloud instances as of October 2026, this update gives users a direct look under the hood whenever Copilot Chat uses memory to personalize an answer.

Instead of guessing which past conversation or standing instruction nudged the AI in a particular direction, users can now inspect the specific personalization inputs that shaped the reply—and take action if a stored detail has overstayed its welcome.

How Memory Citations Work in Copilot Chat​

The new feature is designed around visibility and navigation rather than changing how Copilot generates its answers. According to Microsoft’s roadmap description for ID 574111, memory citations only appear when Copilot memory is turned on and actively used to tailor a response.

When those conditions are met, the workflow is straightforward:

1. Locate the Sources control: Beneath a personalized response in Copilot Chat on the web, select Sources.
2. Open the Personalization tab: Inside the Sources view, switch to the Personalization tab.
3. Inspect the influences: The panel displays the personalization inputs that informed the answer, broken down across four categories:
   • Custom instructions (explicit rules you have given Copilot about how to respond).
   • Saved memories (specific facts or preferences saved to Copilot Memory).
   • Memories learned from conversations (inferences drawn from your ongoing chat history).
   • Past chats (prior conversation threads that provided context).
4. Act on the citations: From the Personalization tab, users can click through to return to a cited chat and pick up where they left off, or review and delete saved memories directly.

Section Summary: The new Personalization tab inside Copilot Chat's Sources menu lets web users see which custom instructions, saved memories, learned conversation details, and past chats influenced a response, with direct options to revisit cited threads or delete saved memories.

Personalization Citations vs. Grounding Citations: A Critical Distinction​

Putting memory citations inside the Sources menu is convenient, but it also invites a common misconception. In enterprise AI, the word "citation" usually means factual verification—a link to a Word document, an SharePoint deck, an email, or a public webpage that proves the AI didn't hallucinate a statistic.

Based on Microsoft Support documentation regarding how Copilot answers prompts, it is essential to separate grounding from personalization:

• Grounding citations connect a response to verifiable work data (powered by Work IQ across emails, files, chats, and meetings), attached files, or web search results. They tell you what factual source material anchors the claim.
• Memory citations in the new Personalization tab show why the response was tailored to you. They reveal the behavioral preferences, prior threads, and remembered facts about your working style that shaped the framing or context of the answer.

A memory citation does not validate the factual accuracy of a statement. If Copilot cites a past chat or a saved memory under the Personalization tab, it is telling you, "I framed the answer this way because of what we discussed earlier," not "This past chat proves the numbers in your table are correct." Furthermore, while the roadmap confirms that the Personalization tab appears when memory is turned on, Microsoft has not published granular rendering rules detailing whether every subtle personalization signal is guaranteed to appear on every single prompt.

Section Summary: Grounding citations verify facts against files, emails, and the web, whereas memory citations explain how your preferences and past interactions shaped the style and framing of the answer.

The Broader Personalization Ecosystem in Microsoft 365​

To get the most out of memory citations, it helps to understand what Microsoft 365 Copilot is actually allowed to remember. According to Microsoft Support and Microsoft 365 Message Center documentation (such as MC1492154, rolling out in October 2026), Copilot personalization draws from several distinct buckets:

• Saved Copilot Memory: Key facts remembered from your conversations, such as recurring topics, work goals, and communication style.
• Chat History Inferences: Dynamic conclusions Copilot draws over time from your previous Copilot Chat sessions.
• Work Insights: For users signed in with a work or school account (depending on the organization’s Copilot plan), Copilot can learn details from Microsoft 365 work content, including emails, Teams chats, and meeting transcripts.
• Custom Instructions: Standing directives configured by the user, such as "Use a professional tone" or "Always summarize in bullet points".

Even without waiting for a memory citation to appear beneath a specific prompt, Microsoft Support notes that users can audit their profile at any time by asking Copilot directly: "What do you know about me?" Copilot will generate a summary of the saved information and work insights it uses to tailor responses. If a specific work insight is off-base, you can tell Copilot in the chat to forget that detail or not use it for personalization.

When you want a completely clean slate for a one-off question—say, drafting a message outside your usual role without polluting your personalization profile—Microsoft Support recommends starting a temporary chat via the button at the top-right of the Copilot Chat page. Temporary chats do not access or store personalized information and do not appear in your chat history (though organizational retention policies and admin eDiscovery still apply).

Section Summary: Copilot’s memory spans saved memories, chat inferences, custom instructions, and—for eligible work or school accounts—work insights from emails and Teams. Users can query this profile directly or bypass it using temporary chats.

What Users and IT Admins Need to Know About Deleting Memories​

The new roadmap feature makes it easy to review and delete a saved memory right from the Personalization tab beneath a response. However, based on Microsoft Learn’s administration documentation (Manage Copilot personalization and memory, updated September 2026) and Microsoft Support guidance, removing a saved memory is only one piece of the data-lifecycle puzzle.

For end users and Windows enterprise administrators, the underlying retention mechanics have several sharp edges that a citation UI alone does not change:

1. Deleting a chat does not delete a saved memory: According to Microsoft Learn, saved memories are retained until a user explicitly deletes them (either via the new Personalization citation tab or in Settings > Personalization). Deleting the original conversation thread leaves any saved memories generated from it intact.
2. Fully wiping learned inferences takes two steps: Microsoft Support explains that to completely remove something Copilot knows about you, you must both delete the saved memory in Chat settings and turn off the Chat history toggle (which may be labeled Chat history & work insights for work or school accounts).
3. The 30-day resurrection window: Turning off the Chat history toggle deletes inferences learned from past chats and work content, but Microsoft Learn and Microsoft Support both document a 30-day window. If a user turns the Chat history toggle back on within 30 days, those deleted chat-history and work-content memories are added right back. Permanent deletion of chat-history details after disabling the control takes 30 days (or within seven days if a user manually deletes all specific chats where a piece of information was originally shared or referenced).
4. Where memories live and how Purview treats them: Microsoft Learn specifies that Copilot memories (saved memories, chat history inferences, and custom instructions) are stored in a hidden folder (CopilotMemory, using item class <IPM.Contact>) inside the user's Exchange mailbox. While this means memories inherit Exchange encryption at rest and Customer Lockbox protections, Microsoft Purview retention policies and labels do not apply to Copilot memory. A "delete Copilot chats after 90 days" Purview policy will not purge a user's saved memories.
5. Admin controls and licensing: According to Microsoft Learn, Copilot memory is available to Copilot Chat users both with and without a Microsoft Copilot add-on license (note that while the specific memory-citations interface is marked Launched/GA on the October 2026 roadmap, Microsoft Learn still characterizes the broader Copilot personalization and memory architecture as in preview and subject to change ). Tenant administrators can disable memory features using the Enhanced personalization control (enhancedPersonalizationSetting in Microsoft Graph), which is on by default. Turning it off stops Copilot from applying saved memories and custom instructions and purges chat-history inferences after 30 days, though saved memories and custom instructions remain stored unless explicitly deleted. Admins can search and delete saved and inferred memories via Purview eDiscovery and Microsoft Graph Explorer, though memory actions do not generate Purview audit log entries.

Why This Update Matters​

AI assistants become genuinely useful in an enterprise environment when they stop acting like amnesia patients every time you open a new tab. Yet the more context an AI absorbs from your past chats and instructions, the harder it becomes to troubleshoot why an output suddenly feels overly casual, fixated on an old project, or weirdly specific.

By surfacing memory citations directly inside the Sources > Personalization tab on the web, Microsoft is giving users a practical debugging tool for their own AI profile. Being able to see the exact custom instruction, saved memory, learned inference, or past chat that steered a reply—and delete a stale saved memory on the spot—bridges the gap between hidden background profiling and transparent user control. Just remember that while the Personalization tab shows you how Copilot tailored its voice to you, you still need grounding citations to verify the facts, and a trip to your Chat history settings if you want to wipe out deeper conversational inferences for good.