Microsoft dropped a significant update to Copilot in Edge on May 13, 2026, transforming the browser’s AI assistant into a memory-equipped, context-aware companion that works across desktop and mobile. The rollout introduces four tentpole features: personalized memory, tab-aware answers, agentic browsing that can autonomously complete multi-step tasks, and voice-and-vision capabilities on mobile devices. Alongside these come a redesigned sidebar, enhanced privacy controls, and a new Copilot Pro tier that unlocks priority access to the latest models.

This isn’t an incremental update. It’s a fundamental re-architecture of how Copilot interacts with your browsing life. Where the old Copilot answered isolated questions, the new version taps into your browsing history, understands what’s open in your tabs, and can perform actions on web pages on your behalf. Let’s break down what that actually means.

Memory: Copilot Finally Remembers Who You Are

The headliner is Copilot Memory—an opt-in system that lets the assistant retain information across sessions. Enable it, and Copilot will remember preferences, recurring tasks, and even details about projects you’ve been researching. Microsoft positions it as a productivity booster: tell Copilot you’re planning a trip to Japan, and it will recall your hotel preferences, dietary restrictions, and previous flight searches weeks later when you return to the topic.

Under the hood, Memory uses a combination of on-device vector stores and encrypted cloud sync. All personal data is processed locally by default; only the derived memory embeddings get uploaded to Microsoft servers, and those are tied to your Microsoft account with end-to-end encryption. The company says it trained the memory system on its own Responsible AI framework to avoid bias and hallucination loops. Practically, you can see—and delete—every memory item from a dedicated Memory Dashboard in Edge’s settings. Sliders let you control how granular the memory gets: disable it per-site, per-topic, or entirely.

Early testers on the Windows Insiders Dev Channel report that Memory works best for recurring workflows—like summarising email threads from Outlook on the web or pulling figures from weekly analytics dashboards. One architect on Reddit noted, “I asked Copilot to compare Q1 revenue from three different Power BI reports I viewed last month, and it dug up the numbers without me opening a single tab.”

Tab-Aware Answers: Context Without Repeating Yourself

If Memory is Copilot’s long-term memory, tab-aware answers are its working memory. The feature lets Copilot see the content of your open tabs—with your permission—and use them as grounding context. Click the Copilot icon, and it lists all your open tabs; check the ones you want to share, and from that point Copilot can cross-reference, summarize, and extract information across them.

Microsoft demonstrated a scenario where a user had open tabs for a competitor analysis document in Office, a Slack conversation about a marketing campaign, and a Google Sheet with budget data. With one prompt—“Compare our Q3 ad spend to the competitor’s projected spend and draft an email to the team highlighting any risks”—Copilot synthesized data from all three tabs and generated a draft in Outlook. No manual copying or tab-hopping required.

The feature leans heavily on Microsoft’s Phi-4 vision model for understanding both text and layout in web pages, even if they’re dynamically loaded. Importantly, it respects websites robots.txt andX-Robots-Tag: noai` directives. Pages marked as sensitive by the content owner will be automatically excluded from tab context sharing. Users can also exclude entire domains globally from Edge’s Copilot settings.

Agentic Browsing: The Browser as Your Proxy

The most ambitious part of the update is agentic browsing. Copilot can now take action on web pages—filling forms, clicking buttons, scrolling, and navigating multi-step flows—all while you watch. Microsoft calls these “browser agents,” and they run in a sandboxed, headful environment that mirrors your own session so sites see your real identity and cookies.

Practical examples include:
- Booking a refund on an airline’s website by identifying your reservation number from an email in a tab and walking through the refund form.
- Applying for a visa by extracting passport details from a scanned PDF, uploading it, and filling the DS-160 form step by step.
- Ordering groceries weekly from your favourite supermarket, adjusting the cart based on an inventory template you maintain in OneNote.

Every action requires explicit user confirmation before execution. A new Agent Console sidebar shows a live feed of what Copilot is doing, complete with a pause button and an undo stack. If Copilot encounters an ambiguous field or a CAPTCHA, it stops and asks for guidance. Microsoft says it trained the agent on millions of real-world web interactions using a combination of human demonstrations and synthetic data, emphasising safety—the agent cannot interact with banking or identity-critical fields unless you whitelist specific paths.

Privacy advocates have already raised concerns. The Electronic Frontier Foundation issued a statement within hours, calling for clear labels when browsers are being operated by AI. Microsoft responded by pointing to the always-visible blue Copilot outline around browser windows when an agent is active, along with the mandatory user confirmation and session recording that gets stored locally only. The company also pledged to publish regular transparency reports on agent usage patterns.

Voice and Vision on Mobile: Copilot in Your Pocket

On iOS and Android, Copilot in Edge now gains voice conversations with real-time interruption handling, much like Google’s Gemini Live or ChatGPT’s Advanced Voice Mode. You can ask follow-ups naturally, and the voice engine—powered by a new on-device neural TTS—responds in one of five natural-sounding voices.

Equally transformative is Copilot Vision for mobile. Point your phone’s camera at a real-world object, and Copilot can overlay information or act on what it sees. Microsoft showed a user scanning a restaurant menu in Spanish, getting an English translation, and then asking Copilot to find dishes that match their dietary profile stored in Memory. Another demo involved pointing the camera at a broken circuit breaker; Copilot identified the model from the label, searched for troubleshooting steps, and offered to connect a licensed electrician via Bing Local.

Vision processing happens entirely on-device for speed and privacy, using the NPU in Snapdragon 10-series chips and Apple’s A18/M4 families. For older hardware, there’s a fallback to cloud processing with an explicit prompt that images will be transmitted. Microsoft stated that vision data is never used for advertising or model training.

The Redesigned Sidebar and Copilot Pro Tier

All these features are housed in a revamped Edge Sidebar that can be resized, pinned, or floated as a separate window. A new Compose pane lets you toggle between “Search” mode (answer a question) and “Do” mode (perform an action). Quick actions for summarising a page, comparing products, or drafting emails sit at the top, while a history of your Copilot conversations syncs across devices.

With this release, Microsoft introduces Copilot Pro for Edge, a $9.99/month subscription that gives early access to the next-gen OpenAI model (GPT-5o-class), longer context windows, and priority access during peak times. It also unlocks unlimited agentic browsing sessions; free users get 10 agentic tasks per month. Pro subscribers can connect Copilot to third-party apps via plugins for services like Salesforce, Jira, and Notion, extending the tab-aware context even further.

Copilot Pro joins the existing Microsoft 365 Copilot offerings, but it’s specifically tailored for the browser. A fringe benefit: Pro subscribers get a royalty-free licence to use AI-generated images from Designer for commercial purposes, a nod to content creators.

Privacy Architecture and User Controls

Coinciding with the feature rollout, Microsoft published a whitepaper detailing Copilot’s privacy architecture—a direct response to the clamour that killed Recall 1.0. The key pillars:
- Local first: Memory, tab context, and agent actions run on-device where hardware supports it. Data leaves the device only in encrypted, user-controlled bursts.
- Ephemeral sessions: By default, Copilot conversations are not saved unless you bookmark them. Memory is the only persistent layer, and it’s under your control.
- Transparency indicators: A persistent Copilot icon glows when it’s active. Agentic browsing adds a blue frame plus a systray notification on Windows.
- Enterprise controls: IT admins can disable features granularly via Intune policies, enforce data residency, and audit agent actions in the compliance portal.

Microsoft also committed to biannual third-party audits. The first will be conducted by a big-four firm before the feature exits preview in September 2026.

Real-World Impact and Early Reception

The Windows Insiders community has been buzzing since the build 26080 release on the Dev Channel. Performance metrics shared by Microsoft claim that agentic browsing reduced time for common web tasks by 62% in internal testing, while tab-aware answers cut research time by an average of 45%. Yet, not everything is smooth. Early feedback flags that Memory occasionally conflates similar projects (e.g., two different clients in the same industry), and the vision feature struggles with handwriting on menus in low light. Microsoft acknowledged these in a blog post and promised iterative improvements via Edge’s monthly stable-channel updates.

Security researchers, meanwhile, are poking at the agent’s isolation model. The agent runs in a separate Windows container with no access to the file system or other apps, but they worry about prompt injection via malicious web pages. Microsoft says it has fine-tuned the model to ignore embedded instructions that try to override user prompts, using a technique it calls “instruction hierarchy clamping.” This is an area to watch.

What Comes Next

The May 13 release is only the beginning of a six-month rollout labelled “Wave 1” internally. Wave 2, expected with the Windows 12 24H2 update in October, will bring cross-app agentic workflows—imagine Copilot pulling data from Excel, cross-referencing it in Edge, and then drafting a PowerPoint slide, all with one natural-language command. Microsoft is also working on “shared memory” for Microsoft 365 Copilot, allowing the Edge assistant to tap into your work graph seamlessly.

Competition is heating up. Google’s Gemini in Chrome already offers tab awareness, but it lacks autonomous agents. Arc Browser’s “Browse for Me” is agentic but limited to pre-defined tasks. Microsoft’s bet is that combining memory, tab context, and agents in the world’s second-most-used desktop browser will tip the scale. For the 230 million daily Edge users, Copilot just became a lot harder to ignore.