Meta's Muse Now Forgets 2.4x Less

Meta began rolling out new models trained with its frontier Muse Wonder system to all users of its Muse personal agent and Meta AI assistant on November 6, the company's Superintelligence Labs said in a research blog post. Meta says the upgrade makes Muse 2.4 times less likely to forget information users need to reach their goals.

Key takeaways

• Muse and Meta AI now run on unnamed "specialized" models trained using Muse Wonder, which Meta launched on October 12.
• On Meta's own tests, the new Muse model scores 47% on its Personal Superintelligence Index, up from 38%, and uses critical context correctly in 90% of tasks, up from 76%.
• Muse Wonder itself stays behind the paid API and Muse Code plans; free users get the two headline features, Active Memory and Active Focus.

What Meta changed in Muse and Meta AI

Meta describes the release as specialized versions of its existing models, trained with Muse Wonder to carry over some of its frontier behavior. The company did not name the models, disclose their size or say how quickly the rollout will reach every account.

Mark Zuckerberg summarized the update on X as bringing "Active Memory and Active Focus from Muse Wonder to everyone." Until now, both features were tied to Wonder, which is available through the Meta Model API and Muse Code, Meta's coding agent.

How Active Memory works

Active Memory is meant to carry useful details from one conversation into the next. "We've trained our models to recognize what type of information is temporary versus permanent, to reason across time, and to distinguish between what is still relevant and what has become stale," Meta wrote.

On Meta's internal Active Memory evaluation, the current Muse model used critical context correctly in 90% of tasks, compared with 76% for the previous model. Put the other way, the failure rate fell from 24% to 10%, a ratio that lines up with the 2.4-fold drop in forgetting Meta reports, although the post does not say the two figures come from the same test.

What Active Focus does when a task breaks

Active Focus targets a different weakness: agents that stall when something goes wrong. "We've trained our models on a variety of bottlenecks reflective of the real world, so that they can try harder, recover, or find alternatives on your behalf," the company said.

Meta reports a 67% improvement in completing everyday tasks in Muse where unexpected issues arise. It did not publish the baseline completion rate behind that number.

For Muse, Meta says the models also learned when to pursue a goal independently and when to work live with the user, and how to absorb corrections mid-task. In Meta AI, the company says the same training helps the assistant break problems into steps, ask sharper follow-up questions and keep up when a user changes direction.

Context: two tiers of Meta models

Meta introduced Muse Wonder and the more capable Muse Wonder Full on October 12, saying both were trained at its Prometheus data center. According to The New Stack, Wonder costs $4 per million input tokens and $20 per million output tokens on the API, while Wonder Full is limited to Muse Code's two most expensive plans.

Muse itself launched in September on the smaller Muse Spark model and includes a free tier. Training lighter models with Wonder lets Meta push some frontier behavior to that free audience without serving its largest model to everyone.

What's next

Every figure in the announcement comes from Meta's own evaluations, and the post includes no methodology, task counts or separate scores for Meta AI. No third-party testing of the new models has been published.

Stronger memory also puts more weight on privacy controls. The November 6 post does not announce new tools for reviewing or deleting what the assistant retains, and Muse has drawn scrutiny before, including over a 6.8GB filesystem dump Meta called intended behavior. Meta said in October that a full Muse Wonder safety report would be released "soon."

FAQ

Do free Meta AI users get Muse Wonder?
No. They get specialized models trained using Muse Wonder. Wonder itself remains available through the Meta Model API and Muse Code.

What is the Personal Superintelligence Index?
It is an internal Meta benchmark. The current Muse model scores 47% on it, up from 38% for the previous model. Meta has not published its methodology.

Is the 67% Active Focus gain independently verified?
No. Meta reported it from internal testing in Muse and did not disclose the baseline.

References

• Bringing Specialized Models to Muse and Meta AI — Meta Superintelligence Labs (Nov. 6, 2026)
• Muse Wonder: A Frontier Model For Personal & Professional AI Agents — Meta AI Research (Oct. 12, 2026)
• Meta's best AI model now trains the one most of its users get — The New Stack (Nov. 6, 2026)