Copilot Memory Problems: Microsoft's AI Forgets Who You Are

Updated January 2026 | 6 min read

You tell Copilot you prefer Python for data work. It suggests R the next day. You explain your project naming conventions. A week later, it creates files with completely different patterns. You outline your writing style preferences. It reverts to corporate boilerplate by the next session.

Microsoft just launched memory for Copilot—a persistent layer that's supposed to remember your preferences, working style, and project context across the entire M365 ecosystem. It went into general availability in January 2026. The pitch is strong: AI that learns and adapts to you over time.

In practice, it's memory trapped inside Microsoft's walls.

How Copilot's Memory Actually Works

Microsoft introduced two connected systems in late 2025:

Saved Memory stores preferences from your conversations. You say "I prefer bullet points in my writing," Copilot logs it. You mention "I'm working on Project Alpha," it remembers. You see a subtle "memory updated" notification. You can view, edit, or delete memories from the Settings pane. You can turn memory off entirely.

Work IQ is the deeper layer. It tracks your role, company structure, recurring projects, and work patterns across Word, Teams, Outlook, and other M365 apps. Microsoft calls it "long-term memory" that maintains "continuous awareness" of your context.

Memories live in a hidden folder in your Exchange mailbox. They follow the same security and compliance policies as your email—Customer Lockbox, encryption at rest, standard retention rules. Microsoft isn't training on your data. It's storing personal context in a proprietary format.

For developers using Visual Studio, Copilot memories let you save coding preferences in personal user files or version-controlled repo instructions. The AI learns project guidelines, code style, and tooling choices. It's the most functional version of the feature because it's tied to specific codebases.

What Copilot Gets Wrong About Memory

The system only works inside Microsoft 365. If you use Slack for communication, it doesn't know. If you manage projects in Notion or Asana, Copilot's blind. If your notes live in Obsidian, that context doesn't exist to Work IQ.

Saved Memory relies on chat history. If you don't explicitly tell Copilot something in conversation, it won't remember. There's no central file you can edit to seed the system with your full context. You have to teach it piece by piece, conversation by conversation.

Work IQ pulls from your M365 activity—emails sent, documents edited, meetings attended. That's useful for Microsoft-centric workflows. It's useless if your real work happens in other tools. You can't point Copilot at an external knowledge base and say "learn this."

Memory fragmentation is the bigger issue. You use Copilot in Word to draft a document. You use it in Teams to summarize a meeting. You use it in Outlook to respond to emails. Each app has access to your Saved Memory, but the context from one session doesn't always flow into the next. Work IQ connects the dots in theory. In practice, you still repeat yourself.

And if you turn memory off—which you can—Copilot becomes session-only. Every conversation starts cold. The whole system resets to amnesia mode.

The Workarounds (And Why They Fall Short)

Tell Copilot everything explicitly. You start every week with a chat: "Remember I'm working on X, my clients are Y, my preferences are Z." It updates your Saved Memory. Great—until you have ten projects and fifty preferences. The list gets long, Copilot's memory becomes cluttered, and you're still manually managing it.

Use Copilot in Visual Studio with repo-level instructions. If you're a developer, this works. You document coding standards, project structure, and tooling in version-controlled files. Copilot reads them. But this only applies to code. It doesn't help with business context, client details, or cross-functional work.

Keep all work inside Microsoft 365. Move everything to Word, Teams, SharePoint. Let Work IQ absorb your activity. Now Copilot has full context—but you've locked yourself into one ecosystem. If a client uses Google Workspace or if you prefer other tools, you're out of luck.

Regularly review and prune your Saved Memory. You open Settings, check what Copilot remembers, delete outdated entries. It's maintenance overhead. And it doesn't solve the core problem: Copilot doesn't have a master document that defines you.

None of these workarounds give you centralized control. You're either feeding Copilot piecemeal through conversations or surrendering to Microsoft's app structure.

The Alternative

Microsoft's approach treats memory as scattered preferences and activity logs. It learns from what you do inside M365. File-based memory flips the model: you define the context upfront, in one place.

CLAUDE.md is a single markdown file that lives in your Obsidian vault. It contains your role, tools, clients, preferences, workflows, everything. Claude Code reads it at the start of every session. No teaching required, no memory fragmentation, no ecosystem lock-in.

Copilot's Work IQ infers your context from behavior. CLAUDE.md states it explicitly. You write down who you are, what you do, how you work. The AI starts with full knowledge. No waiting for it to learn, no hoping it picks up the right details.

It's tool-agnostic. Obsidian syncs across devices. CLAUDE.md works with Claude Code today, adapts to other AI tools tomorrow. You're not dependent on Microsoft's infrastructure or data formats.

You own the file. You control every line. If context changes, you edit the file—not hunt through conversation history and Saved Memory entries. One source of truth, always current, always accessible.

Microsoft's betting on persistent memory across their ecosystem. File-based memory already does this across any ecosystem. Your vault connects your notes, projects, and daily work. The AI reads all of it, not just activity logs from approved apps.