Meta's AI assistant gets better memory and more persistence

On November 6, 2026, Meta announced updated, specialized models for Muse and Meta AI. They gain two features known from Muse Wonder: Active Memory and Active Focus. The models will be rolled out to users of Muse and Meta AI.

Remembers what still applies

Active Memory is meant to let the models remember information from earlier conversations and recognize when it becomes relevant again. According to Meta, the models are trained to distinguish between temporary and permanent information. They are also meant to reason across time and tell apart what is still relevant from what is outdated.

For everyday users, this means in practice that the assistant should not need the same details explained again and again, and that it should avoid using old information that no longer fits.

Meta reports a 2.4x reduction in forgetting critical information the model needs to help the user reach their goals. On Meta's own Active Memory evaluation, the new model solved 90 percent of tasks with correct use of critical context. The previous model reached 76 percent. The share of tasks where important context was not used thus fell from 24 to 10 percent.

Keeps going when something goes wrong

Active Focus is meant to make Muse keep pursuing a goal even when unexpected obstacles arise. According to Meta, the model is trained on real-world obstacles to move forward and find alternatives.

Meta reports a 67 percent improvement in completing everyday tasks where unexpected problems occur. It is not stated how the figure was measured or what the test basis is.

On Meta's Personal Superintelligence Index, the new model scores 47 percent versus 38 percent for the previous one.

Meta also mentions improvements in multitasking, independent goal pursuit, live collaboration, adapting mid-task and breaking down complex problems.

The figures are Meta's own

All figures come from Meta's own evaluations, and no independent verification of them has been found. How much difference the update makes in daily use will therefore have to be tested by users themselves once the models have been rolled out.