Image: Meta
Meta is rolling out specialized models for Muse and Meta AI that it says remember important information more reliably and recover better when tasks go wrong. The company trained these versions of its existing models using Muse Wonder, bringing capabilities from that more powerful model family to a wider audience.
Mark Zuckerberg announced the upgrade on X, saying Active Memory and Active Focus were coming to everyone. Meta's research post describes a rollout to people using Muse and Meta AI. These are models trained using Wonder, rather than an announcement that every user will now run Wonder itself.
Upgraded the models for Muse and Meta Al, bringing Active Memory and Active Focus from Muse Wonder to everyone. In Muse, these new specialized models are 2.4x less likely to forget critical info and are 67% better at self correcting and completing everyday tasks which means less repeating yourself and fewer dead ends.
Mark Zuckerberg November 6, 2026
More on the models: https://t.co/P2DvI43xDj
That distinction matters because an AI model is the software behind an assistant's responses and decisions. Training an existing model using a more capable one can transfer useful behaviors without putting that larger model behind every interaction. Meta says it used Wonder for this training, but its announcement does not explain the exact process or disclose the new models' sizes.
Active Memory is Meta's name for recognizing which details from earlier conversations matter now. According to the company, the models have learned to distinguish temporary information from lasting information and to tell when an old detail has become outdated. The aim is to let people continue a task without explaining their circumstances again.
Meta reports that the current Muse model correctly uses critical context in 90 percent of tasks on its Active Memory evaluation, compared with 76 percent for the previous model. Those figures mean the measured failure rate fell from 24 percent to 10 percent. That is the basis for Meta's claim of a 2.4 times reduction in forgetting, rather than a claim that memory accuracy itself increased 2.4 times.
The numbers describe Meta's own test, not a guarantee that Muse will remember every important detail in a real conversation. The announcement also does not explain changes to memory storage, retention periods or user controls, so the upgrade should not be read as a new privacy policy.
Active Focus addresses a different problem. An assistant may understand a request but stop when an unexpected obstacle interrupts the task. Meta says it trained the models on such problems so they can recover, try another approach and continue toward the user's goal.
In Muse, Meta reports a 67 percent improvement in completing everyday tasks when unexpected problems arise. Zuckerberg described the improvement as better self correction and task completion. The blog does not provide the underlying success rates for that comparison, so it does not establish that the models now complete 67 percent of all tasks.
Meta also says the specialized models better handle several tasks at once, decide when to work independently and when to collaborate with a person, and respond to corrections while work is underway. For Meta AI, it describes benefits such as breaking a problem into steps, asking better follow up questions and adapting when a conversation changes direction.
The New Stack corroborated the rollout and reported that Meta had not named the specialized models or detailed its training method. Its coverage also distinguishes the broadly distributed models from Wonder itself. The performance figures remain company claims, rather than results independently reproduced in that reporting.
At the time of this writing, Meta's announcement does not give a detailed regional rollout schedule. The practical promise is fewer repeated explanations and fewer unfinished tasks. Whether users experience those gains depends on how well Meta's reported improvements carry over from its evaluations into their own conversations and work.
Comments