On 6 November 2026, Meta said it was rolling out specialized versions of its existing models to people using Muse and Meta AI. Trained using Muse Wonder, the models bring what Meta calls Active Memory and Active Focus into those products. The announcement concerns the models behind the two services; it does not announce a separate API release for these specialized versions.
What Active Memory and Active Focus are meant to do
Active Memory is Meta’s approach to recognizing which details from earlier conversations still matter to a task and which have become stale. That distinction matters for an agent working over time: recalling an outdated instruction can be as unhelpful as forgetting a relevant one.
Active Focus is meant to help Muse continue toward a goal when a step fails or circumstances change. Meta also says the models can respond to corrections during a task and judge when to work independently or ask for input. These are descriptions of intended behavior, not guarantees for every connected app or workflow.
What Meta’s tests show
In its Active Memory evaluation, Meta reports that the current Muse model used critical context correctly on 90% of tasks, compared with 76% for the previous model. It separately reports a 67% improvement in completing everyday tasks when unexpected problems occur. The company does not give the underlying completion rates for that second figure, so it should not be read as a 67% success rate.
The announcement does not provide the task set, sample size, scoring rules or independent replication for these results. It also does not establish that the Muse figures apply unchanged to Meta AI. The numbers describe Meta’s internal evaluation, not measured reliability in a reader’s own workflow.
How to assess the update
A useful test is a reversible multistep task that includes an earlier detail which later becomes irrelevant. Change an instruction or introduce a blocked step, then check what the agent remembers, when it asks for help and whether it completes the work. Teams should repeat that test with the app connections and permission settings they intend to use.
