The University of Maryland, Baltimore (UMB) is preparing its community for a major shift in Microsoft 365’s AI assistant. In a notice sent to faculty, staff, and students, the institution revealed that Microsoft Copilot Chat will soon offer personalization through dedicated Chat settings—allowing users to adjust memory and customize instructions. The announcement signals that enterprise-grade memory features, previously seen only in consumer versions of Microsoft’s AI, are finally coming to higher education and business tenants.

While the exact launch date remains unconfirmed, UMB’s proactive warning suggests the rollout is imminent for 2026. The email, obtained by windowsnews.ai, reminds users that all Copilot interactions are subject to the university’s acceptable use policy and data governance rules—a clear nod to the privacy and compliance challenges that such persistent memory might introduce.

What Copilot Chat personalization means

Microsoft has been gradually infusing its AI tools with the ability to recall user preferences. In the consumer space, Copilot’s “memory” feature stores facts like your name, dietary restrictions, or preferred coding languages, using them to tailor responses without repeating yourself. The new M365 implementation, as hinted by UMB, appears to bring that same capability to enterprise accounts.

But there’s a twist: the personalization goes beyond mere facts. UMB’s notice mentions “custom instructions,” a feature that would let users permanently nudge Copilot’s behavior—specifying tone, response length, or professional context. Together, memory and instructions could transform Copilot from a stateless chatbot into a persistent digital coworker that remembers departmental acronyms, project deadlines, or compliance requirements across sessions.

The governance tightrope

UMB’s announcement walks a careful line. On one hand, it highlights the productivity potential. On the other, it underscores that this power comes with responsibility. “All Copilot Chat interactions remain subject to the University’s Acceptable Use Policy and data classification standards,” the email states, according to a source familiar with the communication. This is no boilerplate disclaimer—it targets the very real risk of faculty inadvertently storing student data, health information, or proprietary research in an AI’s memory.

Universities are particularly sensitive because they handle a messy mix of regulated data. HIPAA, FERPA, and export-controlled research all coexist on the same tenant. A persistent AI memory that bleeds context across conversations could accidentally surface protected information to unauthorized users. UMB’s early head’s-up suggests IT administrators will need to tightly manage these settings, possibly via new admin controls in the Microsoft 365 compliance center.

How the feature likely works

Based on Microsoft’s existing Copilot architecture and UMB’s description, the personalization will appear under a “Chat settings” pane within the Copilot sidebar. Users will be able to toggle memory on or off, view what Copilot remembers, and delete individual memories. Custom instructions will probably accept plain-text input—similar to the system prompts that power users already craft for ChatGPT.

Significantly, these memories are tied to the user’s Microsoft 365 account, meaning they could follow a person across devices and applications. A professor who tells Copilot, “I teach graduate-level biostatistics every Tuesday,” might get meeting summaries tailored to that class without re-explaining the context. But that also means the memory could surface in unrelated work chats unless boundaries are clearly set.

Privacy questions linger

Microsoft has yet to publish detailed documentation on M365 Copilot memory privacy, which explains why UMB is moving cautiously. The key questions: Where are memories stored—in the user’s mailbox, a separate encrypted database, or the Azure OpenAI service? Are memories end-to-end encrypted? Can they be e-discovered or subpoenaed? And crucially, does the AI use these memories to train future models?

For enterprise and education users, the default answer to that last question is usually “no.” Microsoft has repeatedly stated that customer data from M365 Copilot is not used to train foundation models. However, memories introduce a gray area. If a user stores the fact “Project X has a $2M budget” in memory, that becomes part of their tenant’s data sphere—but its persistence makes it more vulnerable to accidental exposure. UMB’s reminder to follow data classification standards hints that high-risk data (Level 3 or 4 in many university frameworks) should never be trusted to an AI’s memory.

The role of IT admins

For IT teams managing M365 tenants, this feature rollout will demand immediate attention. Microsoft typically adds new Copilot capabilities with a default-on setting, leaving admins to disable them if needed. UMB’s notice suggests they may be pre-configuring policies or preparing to push out group policy objects (GPOs) that disable memory for certain roles.

PowerShell modules and the Microsoft 365 admin center will likely gain new toggles under the Copilot section. Expect granular controls: admins might disable memory for guest users, restrict it to specific departments, or require periodic reviews of stored memories. Some institutions may opt to block the feature entirely pending a security review, while others will embrace it with training modules.

Real-world implications for faculty and staff

For end users, the 2026 update could be a double-edged sword. A research coordinator who frequently requests literature summaries will appreciate a Copilot that remembers their preferred citation style and database access. But a faculty member serving on multiple committees—some involving confidential student discipline cases, others public curriculum planning—could accidentally create a dangerous memory overlap.

UMB’s guidance will probably stress that users should regularly audit their Copilot memories, much as they would check browser history for saved passwords. Microsoft may build in transparency tools: a “What does Copilot remember about me?” dashboard could become standard, along with prompts like “I see you mentioned clinical trial data. Is it okay to remember that?”

The broader M365 roadmap

Copilot Chat personalization fits neatly into Microsoft’s 2026 vision of an “AI-first” workplace. The company has been teasing deeper contextual awareness since Ignite 2024, with demos showing Copilot proactively surfacing relevant documents before you even ask. Memory makes those demos technically feasible.

Other features likely arriving in the same wave include semantic indexing of enterprise data and cross-application orchestration—Copilot coordinating between Outlook, Teams, and Excel to handle multi-step tasks. Personalization ties them together: the AI remembers that you prefer Excel forecasts in dollars and Teams messages in Spanish, eliminating repetitive configuration.

But competitors aren’t standing still. Google’s Gemini for Workspace is also developing memory features, and enterprise ChatGPT is testing persistent custom instructions. The race to own the “AI assistant that knows you” space will intensify through 2026, with privacy and compliance becoming the true battleground.

What users should do now

The UMB notice is effectively an early warning for the entire M365 ecosystem. If you’re a user or admin, start by reviewing your organization’s acceptable use policy for AI tools. Identify data types that should never touch a memory-enabled system. Communicate with your security team about whether a pilot program makes sense before the full rollout.

For individuals, it’s worth experimenting with the consumer version of Copilot memory to understand its boundaries. Note how often the AI misinterprets or over-applies stored facts, and imagine those errors amplified in a professional setting with legal and ethical stakes.

Ultimately, Copilot Chat personalization represents the next logical step in human-AI collaboration. UMB’s transparency gives the rest of us a preview—and a reminder that features this powerful demand a deliberate, policy-first adoption.