Meta extends Wonder-trained memory and focus to all Muse users

Summary

• Meta released specialized Muse and Meta AI models trained with Muse Wonder to all users.
• Wonder Full stays limited to paid subscribers, a month after Meta announced both tiers.
• Free users now get an assistant that retains goals and recovers from unexpected obstacles.

The latest

Specialized versions of the models powering Muse and Meta AI are rolling out to all users of both assistants, Meta AI Research said on November 6, 2026. The models are trained with Muse Wonder technology, first offered to Muse subscribers, while the Wonder Full tier remains restricted to paying customers. Meta said the update extends Wonder's capabilities across its existing assistant models.

Details

• Active Memory: The first of two features lets both assistants retain important information and surface it again when it becomes relevant. According to Meta, training covered distinguishing temporary information from permanent, reasoning across time, and detecting details that have gone stale.
• Active Focus: The second feature keeps Muse on its objective when an unexpected challenge interrupts a task. Meta said it trained the models on a range of obstacles reflecting real-world conditions, so the assistant can persist, recover, or take an alternative route to the same goal.
• Multitasking design: Active Focus is also built for handling several tasks at once, choosing between independent progress and live collaboration with the user, and absorbing new user direction mid-task without losing the original objective, according to Meta.
• Meta AI side: The same improvements are aimed at making Meta AI a stronger thinking partner, Meta said: breaking complex problems into steps, asking sharper follow-up questions, and adapting when the user changes direction during a conversation.
• Benchmark figures: On Meta's Personal Superintelligence Index, the company put the current model at 47% against 38% for its predecessor. Meta also reported cutting the loss of critical information by a factor of 2.4.
• Memory evaluation: On the Active Memory Eval, Meta said the current model uses critical context correctly in 90% of tasks, compared with 76% for the previous model. All the figures released are the company's own measurements.
• Task completion: Meta reported a 67% improvement for Muse in completing everyday tasks that run into unexpected problems, the metric most directly tied to the Active Focus training.
• The timeline: Meta announced Muse Wonder on October 12, 2026, an announcement that also covered Wonder Full. Wonder's capabilities went first to Muse subscribers before this week's extension to the wider user base of both assistants.
• Not disclosed: Meta did not name rollout regions, give market-by-market timing, or state any pricing attached to the expanded models. The paid Wonder Full tier was described only as remaining subscriber-only.

Between the lines

The split matters more than the numbers. Meta is moving the memory and persistence layer, originally a Muse subscriber feature announced in October, into the free tier while keeping Wonder Full behind payment. That structure suggests the paid product is now defined by a higher capability ceiling rather than by exclusive access to the underlying training itself.

What's next

Watch for Meta to specify rollout regions and timing, for any pricing change around Wonder Full, and for independent testing against the company's reported 47% index score and 90% memory-use figure.

Source: Meta AI Research