Copilot can now cite the memories that shaped an answer — what should Privacy tell staff before we leave Memory on?

You’re right to treat “Copilot remembers me” as a privacy conversation, not a productivity footnote. Citations make the personalisation visible — which is good for trust — but they also make it obvious that stored preferences and past chats can shape what colleagues see on screen.

Here’s what’s going on. Roadmap 574111 (Copilot response personalization - Memory citations) is Launched, with general availability October 2026 (Worldwide, Web). The MRC description matches the product promise: when Copilot Chat uses memory to personalise a response, users can select Sources beneath the answer, open the Personalization tab, and see custom instructions, saved memories, memories learned from conversations, and past chats that shaped the reply — then jump back to a cited chat or review and delete saved memories. Citations appear only when memory is turned on.

Learn’s Manage Copilot personalization and memory page (still flagged as preview / Frontier where Microsoft says so — check Message Centre for your rings) is the admin reality check. Memories — saved memories, details inferred from chat history, and custom instructions — live in the user’s Exchange mailbox in a hidden folder, so they inherit mailbox security patterns such as encryption at rest and Customer Lockbox. Enhanced personalization is the tenant control that allows memory; it is on by default. Turning it off stops users from enabling Custom instructions, Saved memories, and Chat history personalisation. Importantly for GRC: Purview retention policies and labels do not apply to Copilot memory; there are no admin controls to enforce retention specifically for memory; memory and personalisation actions do not generate Purview audit log entries; and admins cannot restrict what types of information go into memory. DSRs lean on eDiscovery / Content Search and Graph (item class for memory search is documented under <IPM.Contact> / CopilotMemory folder patterns on Learn — verify the current procedure before you brief Legal).

This is Schofield’s First and Third Laws with a mailbox attached: if you can’t explain how memory leaves (user delete, admin DSR path, or Enhanced personalization off), don’t pretend “citations” alone are the control. Steady before scale puts Privacy and GRC ahead of “leave the Frontier default and hope.”

Options that actually work:

1. Decide the tenant posture in writing. Keep Enhanced personalization on for Wave 0 productivity or turn it off for cohorts that handle regulated client data until you’ve walked the DSR and eDiscovery path. Document the choice — default-on is still a decision.
2. Teach the Personalization tab as hygiene, not novelty. Brief staff: open Sources → Personalization when an answer feels oddly specific; review and delete saved memories they wouldn’t want in a mailbox folder; remember deleting a chat does not always delete saved memories derived from it (Learn is explicit on that split).
3. Run a Privacy tabletop on DSR and retention gaps. Practice export/delete via the Learn-documented eDiscovery/Graph route; note custom instructions may need user export from Settings → Personalization; accept that interaction retention policies do not sweep memory.
4. Assume citations mean Memory is “governed like SharePoint.” It isn’t — mailbox storage without memory-specific retention or audit is a different risk register line.

12 Panes action / comms: Tell Privacy: 574111 makes memory influence visible under Sources; storage is mailbox-side; retention labels for Copilot Chat do not delete memories. Tell GRC: Register Enhanced personalization state, DSR runbook, and the “no memory audit events” limitation on the AI register. Tell staff: You can see and delete what Copilot remembered about you; treat saved memories like notes in your mailbox. Leadership: Citations improve transparency — they do not invent admin purge for free.

Benefits / risks: Users can finally see why a reply felt personalised; preview/Frontier labels on Learn may still apply in some rings even when MRC says Launched — verify in-tenant; no type-of-data restriction means oversharing into memory is a behaviour problem; eDiscovery complexity for <IPM.Contact> needs a tested playbook, not a slide.

I’d suggest a one-page Steady brief this week: tenant default (on/off), three screenshots of Sources → Personalization, and the DSR owner. Once Privacy and the service desk can answer “how do I forget what Copilot knows about me?”, leaving Memory on becomes a managed choice rather than an accident of the default.

DIGITAL SKILL WORKS LTD / 12 PANES · You ask, we answer