Copilot memory, citations and personalisation: what to set first

On 14 October 2026 the Microsoft 365 roadmap added an item, already marked launched, called 'Copilot response personalisation: memory citations' (roadmap 574111). On its face it is a small interface change: a new tab under Sources that explains why an answer came out the way it did. Behind it sit two earlier entries, both still rolling out, that have Copilot learning from a user's previous communications and chat history. For an Australian business, a system that retains inferences about a worker and acts on them raises questions a search box never did. This note sets out what the three entries say, where they are silent, and what Frontrow recommends settling before memory reaches staff.

Everything below about the product comes from the three roadmap entries; where they are silent, the article says so. Frontrow is not a law firm, and the Privacy Act section poses questions for a business's own adviser rather than answering them.

The three roadmap items, in plain language

What Copilot now remembers

The citation entry names four things. Custom instructions are what the user has told Copilot about how to answer. Saved memories are facts the user has asked it to keep. Memories learned from conversations are inferences Copilot has drawn without being asked. Past chats are the history that entry 503955 says Copilot now draws on. The third category is the one to notice: it is memory the user did not author.

Entry 499153 says the memories are private to the user. The entries do not say where memories are stored, how long they persist, whether they fall inside retention policies or eDiscovery, or what happens to them at offboarding. Frontrow has not found those answers in the sources reviewed and will not invent them; the Microsoft Learn page on data, privacy and security for Microsoft 365 Copilot is the place to confirm them, and /insights/copilot-data-residency-au-sovereign-cloud covers where Copilot data sits.

How the citation shows why an answer was shaped by memory

The mechanism is deliberately ordinary. Beneath a response, the user selects Sources, the same control that lists the documents and web pages behind an answer. A Personalisation tab now sits alongside and lists which custom instructions, saved memories, learned memories and past chats fed into that response. Each cited chat can be reopened, and saved memories can be reviewed and deleted from the same place.

That matters beyond convenience. 'Why did Copilot assume I wanted a one-page summary in a formal tone?' now has an answer, visible to the person it is about: a user-facing audit trail for personalisation. One line in the entry deserves a second read: citations appear only when memory is turned on. The sources do not say whether a response with no Personalisation tab means memory is off or merely unused for that answer, so the absence of the tab is not proof that nothing was remembered.

How a user views or clears memories

The entries describe two routes: the Personalisation tab, where saved memories can be reviewed and deleted in the context of the answer they influenced, and the updated settings from entry 503955 for managing what Copilot remembers and the user's personalisation preferences. The entries do not give the settings path, and they say 'saved memories' when they describe deletion. Whether memories learned from conversations can be removed one at a time, or only by turning memory off, is not stated, and it is the first question Frontrow would put to the tenant.

How an admin controls it

This is where honesty matters more than reassurance. All three entries describe controls that belong to the user. None describes a tenant-wide switch, a group-scoped policy, a retention setting or an audit report, and the Copilot release notes for 11 August to 6 October 2026 do not mention memory at all. 'Citations appear only when memory is turned on' implies a toggle exists, but the entries do not say whose it is.

Frontrow's advice is to assume nothing in either direction. Before any policy says 'IT can disable Copilot memory' or 'memory cannot be disabled', an administrator should open the Microsoft 365 admin centre, read the Copilot settings the tenant actually shows, and record the finding with a date. If no tenant-level control exists, the posture is set by communication and user instruction rather than configuration, and the policy should say so.

The Australian angle: the questions to answer

The Privacy Act 1988 and the Australian Privacy Principles do not mention AI memory, but several principles are engaged the moment a system retains inferences about an identifiable person. The Office of the Australian Information Commissioner's guidance on commercially available AI products, linked in the sources, is the starting point. These are the questions Frontrow would want answered, with the client's own adviser, before memory is on for the workforce.

• Is a memory personal information about the worker? A memory learned from conversations may record a person's role, working pattern, or a leave or health detail mentioned in passing. If it is personal information, APP 1 (privacy policy), APP 5 (notification at collection) and APP 6 (use for the purpose collected) are the principles to work through.
• Does the employee records exemption reach it? The Act exempts a private sector employer's handling of an employee record directly related to the employment relationship. Whether a Copilot memory in a worker's own account is such a record, or falls outside the exemption because the worker is a contractor or the organisation is an agency, is a question for the adviser.
• What about memories about someone else? A memory learned from a chat may concern a client, a candidate or a colleague. It is private to the user as a matter of product design, but as a matter of privacy law it is personal information about a third party, retained by the business's system.
• What does deletion mean? APP 11 asks that personal information be destroyed or de-identified when no longer needed. The sources do not say whether deleting a saved memory removes it from backups, logs or model context, so a business should not promise more than it can show.
• Can the business answer an access or correction request? APP 12 and APP 13 give individuals rights to access and correct personal information. The Personalisation tab shows a worker their own memories; the sources describe no report a privacy officer could use to answer a request from anyone else.
• Does memory touch a decision about a person? The automated decision-making transparency provisions from the Privacy and Other Legislation Amendment Act 2024 are covered by Frontrow's scope check at /insights/privacy-act-adm-rules-copilot-scope-check. If a memory shapes an output that feeds a decision about an individual, that check applies.

What to write into the AI use policy

Most AI use policies were written for a stateless assistant. The document-handling side is covered at /insights/copilot-dlp-sensitivity-labels-policy-recipes; these additions cover the memory side.

• A plain statement that Copilot remembers across conversations, naming the four categories, and that staff can review and delete saved memories under Sources > Personalisation.
• An instruction that anything which must be reproducible or defensible is grounded in a document, not in memory, and the document is what gets cited.
• A rule for other people's information: a chat about a client, a candidate or a colleague is handling personal information and the usual limits apply, because Copilot may retain inferences from it.
• A line on leavers: who checks what memory exists at offboarding and what the business can do about it.
• A named contact for memory questions, usually the privacy officer, and a commitment to update the policy once the tenant's admin controls are confirmed.

What Frontrow recommends setting before it reaches staff

1. 1Confirm in the Microsoft 365 admin centre what memory controls the tenant exposes, and record the finding with a date and a screenshot. Everything else depends on it.
2. 2Decide the default posture: on for everyone, on for the pilot cohort only, or off until the policy is approved. If no tenant control exists, the posture is a communication decision and should be minuted as one.
3. 3Have the privacy adviser read the questions above against the current privacy policy and collection notice, and update both if needed.
4. 4Brief staff with a one-page note: a screenshot of the Personalisation tab, how to delete a saved memory, and what not to put in a chat in the first place.
5. 5Add memory to the Copilot pilot's weekly review: ask the cohort what Copilot remembered about them and whether it was right. Wrong memories are a quality problem before they are a privacy one.
6. 6Diary a re-check for when entries 503955 and 499153 move to launched and when Microsoft documents the admin side.