Meta says Muse and Meta AI now forget less and recover better when tasks go wrong
Meta used its Muse Wonder models to train new versions of the everyday models behind Muse and Meta AI, adding Active Memory and Active Focus.

Meta says its AI assistants just got better at remembering and at finishing tasks. A few weeks ago, Meta introduced Muse Wonder, its most capable models so far. Now it has used Muse Wonder to train new versions of the everyday models behind Muse and Meta AI. Meta is rolling them out to people who use both assistants.
The first upgrade is called Active Memory. It helps the assistant remember what matters and notice when something from an earlier chat becomes useful again. Meta says it trained the models to tell temporary details from lasting ones, and to tell what is still true from what is out of date. In Meta's tests, the new Muse models forget important details 2.4 times less often. On Meta's memory test, they used the key details correctly in 90% of tasks, up from 76%.
The second upgrade is called Active Focus. It helps Muse keep working toward a goal when something unexpected goes wrong. The models can try harder, recover, or find another way. Meta reports a 67% improvement in finishing everyday tasks where problems pop up.
Meta also says the models are better at juggling tasks. They know when to push ahead on their own and when to work with you live. They also handle changes and clarifications in the middle of a task without getting confused.
In Meta AI, the same upgrade helps the assistant break hard problems into clear steps, ask sharper follow-up questions, and keep up when you change direction.
Open it and check the claim yourself. That is the point.