Meta Brings Muse Wonder's Memory and Recovery Skills to All Muse and Meta AI Users

In a research blog post, the company said the new models were trained using Muse Wonder, a series it introduced a few weeks ago and describes as its most capable for agentic tasks. Mark Zuckerberg announced the rollout on X. "By adding these specialized models to Muse and Meta AI, we aim to give more people access to even more capable superintelligence," Meta wrote.

The two features address the places where consumer AI agents most often break down. An assistant that loses track of a constraint mentioned earlier in a conversation, or gives up on a task when a step fails, is of little use for the multi-step errands, planning and project work Meta is now pitching Muse for.

Active Memory is meant to help the models separate temporary details from lasting ones, reason across time and recognise when older context has gone stale or has become relevant again. Active Focus is meant to keep them working towards a goal when obstacles appear, by retrying, recovering or finding another way to finish the job.

Meta's figures compare the current Muse model with its predecessor. On what it calls the Personal Superintelligence Index, the score rose to 47% from 38%. On an Active Memory evaluation measuring whether the model correctly uses critical context, it rose to 90% from 76%. Meta said Muse is now 2.4 times less likely to forget critical information, and showed a 67% improvement in completing everyday tasks in which unexpected problems arise.

In Muse, Meta said, the models have been trained to multitask, to decide when to pursue a goal on their own and when to work live with the user, and to adjust when a user changes direction or clarifies an instruction partway through a task. For Meta AI, the company described the gains in qualitative terms only: breaking complex problems into steps, asking sharper follow-up questions and coping with changes of direction mid-conversation. It gave no numbers for Meta AI.

The claims are hard to check independently. Both the Personal Superintelligence Index and the Active Memory evaluation are internal Meta measures, and the post does not explain how they are built, what tasks they contain or how many samples were used. Meta compared the new models only with its own earlier versions and published no results against rival systems from OpenAI, Google or Anthropic. It also did not give the specialised models names.

The update is the latest in a run of Muse releases. Earlier on Friday Meta added Spaces, which lets users ask Muse to build trackers, planners and project boards linked to other apps, a day after it launched voice and video chat for the agent. Muse's push into tasks such as shopping has already brought friction: Amazon blocked Meta's Muse shopping agent from its store weeks ago.

Meta said the models are rolling out to Muse and Meta AI users but did not give a timetable, list the countries covered or say whether the rollout would be staged.