Meta's Muse Is 2.4x Less Likely to Forget, Company Claims

Meta says the assistant in its Muse app is now 2.4 times less likely to forget critical information people need to reach their goals, after the company moved Muse and Meta AI onto new models trained using Muse Wonder, its most capable model family. The claim comes entirely from Meta's own internal evaluations, and the company did not name the new models or say how quickly the rollout will reach every account.

The company described the change Friday in a Meta AI Research blog post, calling the new models specialized versions of its existing ones. Mark Zuckerberg put it more broadly in a post on X, writing that the upgrade brings "Active Memory and Active Focus from Muse Wonder to everyone." That is a marketing framing, not a technical one; Meta's own post is more cautious about naming what shipped.

What shipped, and where did it come from?

Active Memory and Active Focus debuted with Muse Wonder and Muse Wonder Full on October 12. At launch, Wonder was available through the Meta Model API and Muse Code, while Wonder Full was restricted to Muse Code subscribers. The October announcement also said Meta would later use Wonder "to develop specialized models that we can bring to everyone using Muse." Friday's release is that follow-through, roughly a month later.

In plain words: capabilities that were previously limited to developers and paying subscribers have now reached Meta's mainstream consumer products.

What does Active Memory actually change?

Active Memory is meant to carry useful details from one conversation to the next — the difference between an assistant that remembers your constraints from last week and one that starts every chat from zero. Meta said it trained the models to:

• tell temporary information from permanent information;
• reason across time;
• separate what is still relevant from what has gone stale.

On Meta's internal Active Memory evaluation, the current Muse model correctly used critical context in 90% of tasks, compared with 76% for the previous Muse model. That is a 14-point gain on a benchmark Meta itself designed and ran — no third-party evaluation is cited anywhere in the post.

The takeaway: the memory improvement is real by Meta's own yardstick, but the yardstick is Meta's.

How does Active Focus help when tasks break?

Active Focus targets a different weakness — an AI agent that stalls when something breaks mid-task. Meta said it trained the models on "a variety of bottlenecks reflective of the real world" so they can try harder, recover, or find another route.

In Muse, the company reported a 67% improvement in completing everyday tasks where unexpected issues come up. It did not give the baseline completion rate behind that number, which makes the percentage difficult to interpret. A 67% improvement from a low starting point could still leave substantial failure rates.

The takeaway: relative improvement numbers without baselines tell readers about momentum, not about how reliable the assistant is today.

How should the two assistants behave differently?

For Muse, Meta said the models learned when to push ahead on a user's goal on their own and when to work through it live with the user, and they can take corrections in the middle of a running task without getting confused. For Meta AI, the company expects the same training to help the assistant break complex problems into clear steps, ask sharper follow-up questions, and keep pace when someone changes direction mid-conversation.

Note the verb shift: Meta published benchmark figures only for Muse, and all of those figures come from its own evaluations. For Meta AI, the company is describing expectations, not measured results.

What about everything Muse remembers?

A longer memory raises a question Friday's post does not address: what does Muse keep about people, and who controls it? Last month, WebProNews reported on Muse instructions that security researcher Karan Joshi extracted and shared with Wired, which showed the agent creating "a page for every person in the user's life."

Friday's post explains how the new models judge what is still relevant, but it does not mention any new controls for users to review or delete what the assistant remembers. Better memory with no announced deletion or review tools means the profiles the agent builds could persist — or go stale — with no user visibility either way.

The takeaway: Meta shipped smarter remembering, not clearer accountability for what gets remembered.

Why this matters

The release shows Meta's strategy of using a frontier model family — Muse Wonder — as a training source for models that run across its consumer surface area, closing the gap between what it demos in research and what everyday users experience. Every number in the announcement, however, is self-reported, unbenchmarked by outsiders, and in Meta AI's case not benchmarked at all. Until independent testing or published baselines appear, the "2.4 times less likely to forget" figure should be read as a vendor's claim about its own product.