Meta’s Muse Scores 47% on Personal Superintelligence Index

By Sophie Lindqvist

Meta AI Research said on November 6, 2026, that the current Muse model scores 47% on Meta’s Personal Superintelligence Index, compared with 38% for the previous model. The same announcement said Meta is rolling out specialized versions of the systems behind Muse and Meta AI, trained using Muse Wonder. Active Memory and Active Focus are included in that rollout.

Meta said this is not a rollout of the Wonder series itself. The company is extending capabilities introduced with Wonder through specialized versions of its existing models, which puts the research inside products people already use. That makes the deployment a test of whether benchmark gains become reliable everyday assistance.

What changed on November 6?

Meta had outlined this path on October 12, 2026. The company described Muse Wonder and Muse Wonder Full as its “most capable models to date.”

At that time, Meta said Wonder would first power Muse experiences for subscribers. It would then use those models to develop specialized versions for everyone using Muse.

The November announcement carries that plan into both Muse and Meta AI. Active Memory and Active Focus are being rolled out through specialized versions of existing systems, rather than a general release of Muse Wonder or Muse Wonder Full. The announcement separates the underlying models from the behavior being deployed.

Meta said it is improving specialized versions of its existing models through Muse Wonder. In short, Meta is shipping improvements to existing models, not a new Wonder model line.

How does Active Memory decide what to remember?

Active Memory lets an assistant “retain important information and retrieve it when it becomes relevant to a later conversation,” Meta said. The company’s training covers three parts of the problem:

• Distinguishing temporary information from permanent information.
• Reasoning over time.
• Recognizing when earlier details remain relevant or have become stale.

That training turns memory into a problem of selection and timing. According to Meta, the model must identify which earlier context applies now and which context no longer applies. The system filters what comes back rather than treating every saved detail as equally useful.

Meta reports a 2.4x reduction in forgetting information critical to users’ goals in the improved Muse models. Its Active Memory Eval also shows the current model using critical context correctly in 90% of tasks, compared with 76% for the previous model.

Meta says the two measurements cover different parts of the workflow. The 2.4x figure concerns retaining goal-critical information, while the 90% result concerns applying relevant context correctly during a task.

The distinction matters because better retention alone does not establish that a model will use what it remembers. Meta presents the figures as complementary evidence that useful memory needs both selection and correct application.

Meta’s naming keeps the measurements separate. The 47% headline comes from the Personal Superintelligence Index, while the 90% and 76% figures come from Active Memory Eval. Neither number substitutes for the 2.4x retention measure because Meta describes retention and application as separate capabilities.

In plain terms, Active Memory has to get the right detail back at the right moment.

How does Active Focus handle interruptions?

Active Focus is designed to keep Muse moving toward a goal when an unexpected challenge interrupts the work. Meta trained the models on varied bottlenecks intended to reflect real-world conditions. The company said the training teaches them to “persist, recover, or pursue alternatives for the user.”

That mechanism targets failures that appear after work has started. According to Meta, the specialized models should keep the original objective in view while looking for another way forward, rather than stopping at the first unexpected obstacle.

Meta reports a 67% improvement in completing everyday tasks where unexpected issues arise. This measure covers completion under interruption in Muse, rather than the broader Personal Superintelligence Index score.

The specialized models also change how an agent works with someone while a task remains underway. Meta says they can:

• Multitask.
• Choose between independent progress and live collaboration.
• Incorporate steering or clarifications without becoming confused or losing the task.

That links persistence with responsiveness. According to Meta, an agent can continue pursuing the goal while changing course when the user intervenes. A system that keeps working but ignores corrections has not preserved the right objective.

In Meta AI, Meta said the same improvements are intended to make the assistant a better thinking partner and collaborator. The company says it can divide complex problems into clear steps, ask sharper follow-up questions, and keep pace when a user changes direction mid-conversation.

The practical goal is to keep working without becoming rigid when the person asks for a change.

What changes for users?

For users, Meta frames the rollout’s practical test as continuity. The company identifies three linked requirements:

• Useful context survives changing circumstances.
• Work recovers from obstacles.
• Corrections join the task without making the assistant lose the goal.

Meta ties those behaviors to the reported results. The 47%, 90%, 76% and 67% figures are Meta’s own measurements, while the user-facing claim covers longer and changing interactions.

The 47% headline gives Meta a higher score, but the product change reaches further. Meta is sending specialized updates to both Muse and Meta AI, so the research moves from a model into two assistant experiences. The research stakes are practical: can the systems retain context, recover from interruptions, and accept direction without turning long-running work into disconnected prompts?

Meta is following the staged route it set out in October. Wonder began with subscriber experiences and pointed toward specialized models for everyone using Muse; the November rollout extends that route to Meta AI. This is a practical test of whether specialized versions can preserve the capabilities Meta found in its newest models.

Going forward, the update’s value will rest on whether Active Memory and Active Focus can keep existing assistants dependable through longer, messier tasks, without a full Wonder-series rollout.