Meta switched the assistant inside WhatsApp, Instagram, Messenger and Facebook to new models on Friday, and the selling point is memory. Meta AI and Muse, Meta’s agent app, now run on “specialized” models trained using Muse Wonder, the company said on November 6, 2026. Mark Zuckerberg described the upgrade as “bringing Active Memory and Active Focus from Muse Wonder to everyone.”
That reaches a lot of people. Meta said in October 2025 that more than 1 billion people use Meta AI every month, and an assistant that carries details from one conversation to the next also keeps more about the person typing. Meta’s announcement didn’t mention any new privacy controls. And WhatsApp’s own help page for the Meta AI memory feature now says that feature is gone. Below: what changed, what Meta’s numbers do and don’t show, and how to see and delete what Meta AI and Muse keep.
- What changed: since November 6, Muse and Meta AI run on new models trained using Muse Wonder. It’s free and there’s nothing to install, but Meta hasn’t named the models or said which countries get them.
- Meta’s numbers: in Muse, 90% of test tasks used critical context correctly, up from 76%, and Meta claims a 67% improvement on tasks that hit a snag. Meta ran every test, and it published none for Meta AI.
- To delete it: in Muse, tap Assistant, Identity, Memory, or reset Muse under Settings, Data controls. In Meta AI chats on WhatsApp, Messenger and Instagram, send /reset-ai (one chat) or /reset-all-ais (all chats).
- The catch: Muse’s help page says it “may still remember information it learned from what you deleted.” Outside the EU, the UK and South Korea, Meta has used AI chats to pick ads since December 16, 2025, with no opt-out.
What Meta switched on November 6
Meta’s research post, “Bringing Specialized Models to Muse and Meta AI,” says the company used Muse Wonder, its most capable model family, to train improved versions of the models behind Muse and Meta AI. It doesn’t name those models, give their size or say how quickly the rollout reaches every account. Muse Wonder itself is a paid product. It launched October 12 on Meta’s Model API at $4 per million input tokens and $20 per million output tokens, as we reported when Meta priced Muse Wonder and kept its weights closed. Meta said then that it would later use Wonder “to develop specialized models that we can bring to everyone using Muse.” Friday’s update delivers on that, and it reaches Meta AI too.
Two capabilities carry over. Meta’s Muse Wonder launch post defines Active Memory as “carrying important context about you across conversations and actively retrieving the details relevant to the task at hand.” Meta says it trained the new models to tell temporary information from permanent information and to separate what’s still relevant from what has gone stale. Active Focus is about persistence: the models were trained on “a variety of bottlenecks reflective of the real world” so they try harder, recover or find another route when a task goes wrong.
How that shows up depends on the app. In Muse, Meta says the models know when to push on a goal alone and when to work through it with you, and they can take corrections in the middle of a running task. For Meta AI, the chat assistant, Meta’s AI account on X says the training helps it “break complex problems into steps, ask sharper follow-up questions and adapt when you change direction.” Meta’s description of the Meta AI changes doesn’t mention memory.
Meta’s numbers are Muse numbers, and Meta ran them
Meta published four figures. All come from its own evaluations, and all describe Muse:
| Meta’s test (Muse) | New model | Previous model |
|---|---|---|
| Active Memory eval: tasks that correctly used critical context | 90% | 76% |
| Forgetting critical information | 2.4x less often | baseline |
| Personal Superintelligence Index | 47% | 38% |
| Completing everyday tasks when unexpected issues come up | 67% improvement | not published |
The 2.4x figure holds up against Meta’s own chart. Going from 76% to 90% means the model misses critical context in 10 tasks out of 100 instead of 24, and 24 divided by 10 is 2.4. The rest is harder to read. Meta doesn’t say how many tasks either test contains, what its Personal Superintelligence Index measures, or what completion rate the 67% improvement starts from. Nobody outside Meta has run these tests. And there are no figures at all for Meta AI, which is the version most people meet, inside WhatsApp or Instagram.
The Meta AI memory catch: WhatsApp says the old feature is gone
Meta AI has had a memory before. In January 2025, Meta said it was rolling out a feature that let Meta AI remember details people shared in one-on-one chats on WhatsApp, Messenger and Facebook in the US and Canada, such as being vegan, and that people could delete those memories at any time. As of November 9, 2026, WhatsApp’s help page titled “About Meta AI Memory” is a single sentence: “Meta AI Memory and any details Meta AI may have remembered are no longer available.”
Meta hasn’t explained how that fits with an announcement promising Active Memory “to everyone.” It hasn’t said whether Meta AI now remembers across conversations the way the old feature did, whether the old memories were deleted or folded into something else, or where people would review what it keeps. The November 6 post doesn’t describe a memory screen or setting for Meta AI.
There’s also a quieter kind of memory. WhatsApp’s Meta AI help page says that in some countries, Meta AI may use information from accounts you’ve added to your Meta Account, such as Facebook or Instagram, including “your interests, age, location, and profiles,” to make its answers more relevant.
What Muse keeps about you, file by file
Muse is more open about it. Meta’s Muse data help page says Muse “learns about you over time.” It lists what the app stores: your conversations, information pulled in through Connectors such as email and calendar, goals, files it creates, reminders and three plain-text files. Memory.md holds what it saved to personalize answers, Soul.md describes its personality and boundaries, and Identity.md holds its name and “vibe.” You can open and edit each one.
The same page has the line worth reading twice: “After you delete something from Muse, Muse may still remember information it learned from what you deleted.” Its “forget” skill removes information “to the best of its ability.” The setting that lets Meta train on your Muse conversations, “Help improve our AI models,” is on when you first use the app. Meta says it removes some identifiers, such as names, email addresses and phone numbers, from conversations it uses for training.
Muse’s memory isn’t only about the user, either. WebProNews noted on November 7 that Muse instructions a security researcher extracted and shared with Wired in October showed the agent creating “a page for every person in the user’s life.” Meta’s November 6 post doesn’t address that.
How to see and delete what Meta AI and Muse remember
In Muse, Meta’s help page lists these controls:
- Ask it. Meta says the simplest route is to ask Muse what it has saved about you, or to tell it to forget a topic.
- Edit the memory file: tap the Assistant icon, then Identity, then Memory, and change or delete entries in Memory.md.
- Stop training on your chats: go to Settings, then Data controls, and turn off “Help improve our AI models.” Meta says the change also applies to earlier conversations.
- Start over: Settings, Data controls, Reset Muse. It permanently deletes chat history, files and active tasks, and it can’t be undone.
For Meta AI on WhatsApp, Messenger and Instagram, deleting the chat isn’t enough, because Meta keeps its own copy. WIRED’s guide lists two commands that work across the three apps: send /reset-ai in a chat to reset that chat, or /reset-all-ais in a chat with Meta AI to reset all of them, group chats included. WhatsApp’s help center still refers to /reset-ai. For anything you don’t want remembered, WhatsApp’s Incognito chat with Meta AI runs on Meta’s Private Processing, which WhatsApp says even Meta can’t read, and by default the chat disappears when you leave it. The /reset-ai command isn’t supported there.
Training on what you type is a separate fight from training on what you publish. Meta wasn’t among the companies that backed the shared robots.txt AI-training opt-out we covered this month, so for now the Muse toggle is the switch Meta gives individual users.
The ads link you can’t switch off
Memory also feeds Meta’s main business. Since December 16, 2025, Meta has used people’s interactions with its AI features to personalize content and ads across its apps, under a policy it announced in October 2025. TechCrunch reported there’s no opt-out, and that the policy doesn’t apply in the EU, the UK and South Korea. Meta says it doesn’t use AI conversations about religion, sexual orientation, politics, health, ethnicity, philosophical beliefs or union membership for ads, and that Meta AI chats on WhatsApp only shape other apps if WhatsApp is added to the same Accounts Center.
That’s why the memory upgrade matters beyond convenience. A model that’s better at keeping track of your plans, preferences and problems is useful, and Meta’s own help pages show the same information can stick around after you delete it. For people who use WhatsApp for work, anything typed to Meta AI about customers or colleagues falls under the same rules.
What to watch next
Three things would settle the open questions. Meta could publish evaluation numbers for Meta AI rather than Muse, name the countries that get the new models, and replace WhatsApp’s one-line memory page with instructions for reviewing what Meta AI keeps. Until it does, the most reliable check is the one Meta suggests for Muse: ask the assistant what it knows about you, tell it to forget what it shouldn’t have, and ask again a day later.
Frequently asked questions
Does Meta AI remember my conversations?
Meta says the new models behind Meta AI, live since November 6, 2026, are trained to recognize when earlier information matters and to carry important context across conversations. But WhatsApp’s help page says the older Meta AI Memory feature and the details it saved “are no longer available,” and Meta hasn’t described a memory setting for Meta AI. In the Muse app, saved details sit in a Memory.md file you can edit.
How do I delete what Meta AI knows about me on WhatsApp?
Deleting the chat doesn’t remove Meta’s copy. Send /reset-ai in a chat with Meta AI to reset that chat, or /reset-all-ais to reset every Meta AI chat, including groups. For questions you don’t want stored, use an Incognito chat with Meta AI, which by default disappears when you leave it.
Can I stop Meta from training on my Muse chats?
Yes. In Muse, go to Settings, then Data controls, and turn off “Help improve our AI models.” Meta says the setting is on by default and that turning it off also covers past conversations.
Does Meta use Meta AI chats for ads?
Yes, outside the EU, the UK and South Korea. Since December 16, 2025, Meta has used interactions with its AI features to personalize content and ads, and TechCrunch reported there’s no opt-out. Meta says sensitive topics such as health, religion and politics aren’t used for ads.
Is Meta’s Active Memory free?
Yes. The new models run behind the free Meta AI assistant and the Muse app, with nothing to install. Muse Wonder, the model they were trained with, costs $4 per million input tokens and $20 per million output tokens on Meta’s Model API.
Sources
- Meta AI Research: “Bringing Specialized Models to Muse and Meta AI,” November 6, 2026 (rollout, Active Memory and Active Focus, Muse evaluation figures)
- Meta AI Research: “Muse Wonder: A Frontier Model For Personal & Professional AI Agents,” October 12, 2026 (Active Memory definition, plan for specialized models)
- Mark Zuckerberg on X, November 6, 2026
- AI at Meta on X, November 6, 2026 (what changes in Meta AI)
- WhatsApp Help Center: “About Meta AI Memory”
- WhatsApp Help Center: “About Meta AI”
- WhatsApp Help Center: “About Incognito chats with Meta AI”
- Meta Help Center: “How to manage your Muse data”
- Meta Newsroom: “Building Toward a Smarter, More Personalized Assistant,” January 27, 2025
- Meta Newsroom: “Improving Your Recommendations on Our Apps With AI at Meta,” October 1, 2025 (1 billion monthly users, ads from December 16, 2025)
- TechCrunch: “Meta plans to sell targeted ads based on data in your AI chats,” October 1, 2025
- WIRED: “What Is the Meta AI Button in WhatsApp, and How Do I Remove It?” April 4, 2025 (reset commands)
- WebProNews: “Meta Puts Muse Wonder’s Memory and Focus Skills Into Muse and Meta AI,” November 7, 2026
- AI at Meta on YouTube: “Introducing Muse Wonder & Muse Wonder Full: frontier models for agentic tasks,” October 12, 2026
All sources accessed November 9, 2026.
Last updated: November 9, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
