Meta is upgrading the brains behind Muse and Meta AI. From 6 November, both products run on new specialized models trained using Muse Wonder, Meta’s frontier model family, bringing two capabilities called Active Memory and Active Focus to everyday users. Active Memory is meant to help the assistants remember what matters across conversations, while Active Focus is designed to keep them working toward a goal when things go wrong. Meta says the upgraded Muse models forget critical information 2.4 times less often and finish everyday tasks 67 percent better when unexpected problems appear.
What Active Memory Does
Active Memory is Meta’s answer to one of the complaints about chatbots: they forget. The company says the new models are trained to tell temporary details apart from lasting ones, judge when earlier information becomes relevant again, and drop what has gone stale. Meta’s own evaluation puts that skill at 90 percent, up from 76 percent for the previous model.
The idea is that Muse should carry context across conversations instead of starting fresh each time. A trip plan from last week or a half-finished project should still be there when it matters, without the user repeating it. That is the promise; whether it holds outside Meta’s own tests remains to be seen.
What Active Focus Adds
Active Focus tackles the other half of the problem: getting stuck. Meta says the models are built to keep pushing toward a goal when a task hits a snag, trying harder, recovering from setbacks, or finding another route instead of giving up. Meta reports a 67 percent improvement in finishing everyday tasks when unexpected issues arise.
The same models also draw on Muse Wonder’s ability to juggle several tasks and judge when to work alone or check in with the user. In Meta AI, the changes aim at breaking complex problems into steps, asking sharper follow-up questions, and adapting when the user changes direction. Most users should see the upgrade without changing any settings.
The Numbers, With a Caveat
Meta’s headline figures are striking. On its Personal Superintelligence Index, the current Muse model scores 47 percent, up from 38 percent before. The company also reports a 2.4 times reduction in forgetting information critical to a user’s goals. But these are Meta’s own evaluations, not independent audits, and no model card or outside benchmark has been published.
That distinction matters. The specialized models are not Muse Wonder itself, which remains Meta’s frontier model for harder work and paid plans. This is a consumer-app update that brings some of Wonder’s training into the models most people actually use, and it arrives the same day Meta launched Spaces, its customizable mini-apps for the Muse app.
Our Read
Memory separates a useful assistant from a forgetful one, and Meta knows it. Muse passed 10 million users in two months, and the fastest way to keep them is to make the assistant seem like it knows them. A model that remembers a project from Tuesday and picks it up on Friday is more useful than one that starts over.
The comparison that matters is with rivals. ChatGPT, Claude, and Gemini all use memory to keep projects coherent, and each has stumbled when old context resurfaces at the wrong moment. Meta’s bet is that training the skill into the model produces fewer mistakes. The 90 percent figure is Meta’s own, so the verdict waits for real use.
Why It Matters
If it works, this upgrade makes Muse feel less like a chatbot and more like an assistant that knows you. Memory turns one-off chats into ongoing work: a gift list that stays current, a project that survives the weekend, a plan that adapts when circumstances change. The caveat matters too — remembering the wrong thing is worse than forgetting.
The rollout lands in a busy week for Muse. Meta launched Spaces, the app’s customizable mini-apps, on the same day, and rolled out real-time voice chats with an expressive avatar to all users in the United States the day before. Muse also reached Mexico this week, its latest market expansion since launching in September.
Background
Muse launched in September 2026 as Meta’s standalone AI app, built around the Muse Wonder model family that Meta positions as its path to personal superintelligence. The app passed 10 million users within two months of launch, and Meta has since expanded it to Mexico, added real-time voice chat in the United States, and launched Spaces alongside these specialized models.
Source: Alexandr Wang’s announcement and Unite.AI’s full report
0 Comments