Meta brings Active Memory and Active Focus to all Meta AI and Muse users
Meta has started giving every user of Muse and Meta AI new specialized models trained with Muse Wonder, the frontier model it introduced on October 12, 2026. The announcement, published by Meta Superintelligence Labs on November 6, 2026, carries two capabilities that until now were tied to Wonder itself: Active Memory, which carries a user's important details across conversations, and Active Focus, which keeps an agent working toward a goal when a task runs into unexpected problems.
The company backs the rollout with four figures, all from its own evaluations. On its Personal Superintelligence Index, the current Muse model scores 47%, against 38% for the model it replaces. On an internal Active Memory evaluation, the model uses critical context correctly in 90% of tasks, up from 76%. Inside Muse, Meta reports a 2.4x reduction in how often the assistant forgets information it needs to help users reach their goals, and a 67% improvement in completing everyday tasks where something unexpected goes wrong.
Muse Wonder and Muse Wonder Full themselves do not move to the free tier. Muse Wonder remains available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, and Muse Wonder Full stays inside the $50 and $200 monthly Muse Code plans. What reaches everyone is a set of smaller, specialized versions of Meta's existing Muse models, trained using Wonder so that they inherit its memory and persistence behavior.
··········
WHAT META ANNOUNCED FOR MUSE AND META AI ON NOVEMBER 6.
The November 6 post describes the release as bringing key frontier capabilities to everyone through specialized models. Meta does not give these models public names or version numbers, and it does not say how large they are or how their training differs from the Muse Spark models that powered Muse at its September launch. The post states only that they are specialized versions of existing models, trained with Muse Wonder, and that they are available to people who use Muse and Meta AI.
The description of Active Memory in the new post goes slightly further than the October launch material. Meta says the models are trained to distinguish temporary information from permanent information and to recognize when something a user said earlier has become stale. In practice, a model trained this way should treat a one-off detail, such as a hotel booked for a single trip, differently from a lasting preference, such as a dietary restriction, and should stop relying on a fact once the user has said it no longer applies. Meta also says the models learn to recognize when earlier information becomes relevant to the current request, which is the part of memory that users notice most directly.
Active Focus is described in terms of recovery. When a task hits an obstacle, the model is meant to try harder, recover from the error or find an alternative route, instead of stopping and handing the problem back to the user. In the October announcement Meta added that Wonder asks for clarification when needed and checks in before acting on anything important, and those same expectations frame the behavior of the new models.
Meta separates what the change means for its two consumer products. For Muse, the personal agent that can act across apps and services, the company says the models handle several tasks at once, judge when to pursue a goal independently and when to work live with the user, and adapt to corrections in the middle of a task. For Meta AI, the conversational assistant, Meta says the models break complex problems into clear steps, ask sharper follow-up questions and keep up when the user changes direction during a conversation.
··········
MUSE AND META AI BENCHMARK RESULTS BEFORE AND AFTER THE UPDATE.
Meta published two before-and-after comparisons and two relative improvements. The table below collects them as Meta reported them.
........
Measure (Meta internal) | Previous model | Current model | Change |
Personal Superintelligence Index | 38% | 47% | +9 points |
Active Memory Eval, critical context used correctly | 76% | 90% | +14 points |
Forgetting critical information in Muse | Baseline | 2.4x less often | Relative only |
Completing everyday tasks with unexpected issues | Baseline | 67% better | Relative only |
........
The two point-based results can be checked against each other. On the Active Memory Eval, the share of tasks where the model failed to use critical context fell from 24% to 10%, and 24 divided by 10 gives exactly 2.4. That matches the 2.4x reduction in forgetting that Meta reports for Muse, although the post presents the two figures separately and does not say whether they come from the same measurement. The Personal Superintelligence Index gain of nine points corresponds to a relative improvement of about 24% over the previous score.
The limits of these numbers are clear from what Meta leaves out. The post contains no methodology for the Personal Superintelligence Index or the Active Memory Eval, no task counts, and no description of how the baseline for the 67% Active Focus figure was measured. Both benchmarks are Meta's own, and no third-party evaluation of the specialized models has been published. The figures describe Muse; Meta gives no separate scores for Meta AI, even though the same models are rolling out there.
··········
[[ADPLUS_300x250_1]]
WHO GETS ACTIVE MEMORY AND ACTIVE FOCUS, AND ON WHICH META PLAN.
The update changes which Meta products carry Wonder-derived behavior, while the pricing of Wonder itself stays where it was in October. The table summarizes access across Meta's consumer and developer offerings, using Meta's official pages.
........
Product or plan | Model | Price | Active Memory and Active Focus |
Muse and Meta AI, all users | Wonder-trained specialized models | Free | Yes, from November 6 |
Muse subscribers | Muse Wonder (announced as coming first to subscribers) | Subscription, prices not published in the announcements | Yes |
Meta Model API | Muse Wonder | $4 input, $0.20 cached input, $20 output per 1M tokens | Yes |
Muse Code Power Usage | Muse Wonder Full | $50 a month | Yes |
Muse Code Maximum Usage | Muse Wonder Full | $200 a month | Yes |
........
The Muse Wonder model page lists a context window of 1 million tokens for both Wonder and Wonder Full. The Power Usage plan of Muse Code includes 20 times the usage of the entry Everyday Usage plan, while Maximum Usage includes 80 times that usage. Meta's November 6 post does not state the regions where the new models are live. Muse launched in the United States in September on iOS, Android, the web and WhatsApp, and Meta AI is available more widely, so the regional reach of the update depends on where each product already operates. Meta has also not given a rollout schedule, and some users may see the change later than others.
For readers comparing Meta's free and paid tiers, our guide to Meta AI limits and free access covers the earlier structure of Meta's plans, and our report on the launch of Muse Wonder and Muse Wonder Full covers the October announcement in detail.
··········
WHAT CHANGES FOR PEOPLE WHO USE MUSE AND META AI EVERY DAY.
The most visible difference should be continuity. A model with stronger cross-conversation memory is expected to remember stated preferences, ongoing projects and recurring details without the user repeating them, and to bring them up when they are relevant. Because the models are trained to treat some information as temporary, users should also see fewer cases where an outdated detail resurfaces in a later answer. Meta's own figures put the remaining miss rate on its memory evaluation at about one task in ten, so the assistant will still overlook context in some exchanges.
Active Focus mainly affects multi-step requests. In Muse, which can book, buy, send and organize on a user's behalf, a model that recovers from errors and tries alternatives is more likely to finish a task such as rescheduling a reservation when the first option fails. Muse already shows an audit trail of past and planned actions and asks for permission before sensitive steps such as sending an email or making a purchase, and those safeguards remain the main checkpoint between a persistent agent and an unwanted action.
Stronger memory makes the existing privacy controls more relevant. When Meta launched Muse in September, it said users can tell Muse to forget specific things it has learned, can opt out of having their interactions used to train Meta's AI models, and can change app access or disconnect a connected service at any time. Meta also said that Muse does not share conversations or virtual machine data with its advertising systems. The November 6 post does not announce new memory controls, so these existing settings are the tools users have to review and limit what the new models retain. Readers who want background on how Muse handles connected apps, payments and its safety agent can find it in our report on the Muse launch.
Developers are not directly affected by the consumer rollout. The specialized models are not listed as separate options in the Meta Model API, and Muse Wonder keeps the same per-token prices it had at launch. A developer who wants Wonder's memory and persistence behavior in an application still needs to call Muse Wonder through the API or work inside a Muse Code subscription.
··········
MUSE WONDER-TRAINED MODELS: CONFIRMED FACTS AND OPEN QUESTIONS.
Meta has confirmed that specialized models trained with Muse Wonder are rolling out to all Muse and Meta AI users, that they carry Active Memory and Active Focus, and that on its internal tests they raise the Personal Superintelligence Index from 38% to 47% and correct use of critical context from 76% to 90%. It has also confirmed a 2.4x reduction in forgotten critical information and a 67% improvement on everyday tasks with unexpected issues inside Muse. Muse Wonder and Muse Wonder Full keep their October pricing and plan placement.
Several points remain undisclosed. Meta has not named the specialized models, described their size, published the methodology or task sets behind its two internal benchmarks, or released figures for Meta AI specifically. It has not published a regional rollout schedule or announced new user controls for the stronger memory features. The safety report that Meta promised alongside Muse Wonder in October would help readers judge how the memory behavior handles private information, and its publication is the next document to watch for an independent check on these claims.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
