Meta Muse Models Bring Active Memory to All Meta AI Users
Meta's new Muse Wonder-trained models reach every Muse and Meta AI user, and Meta says Muse now forgets critical details 2.4 times less often than before.
Hakkı ALKANPublished 3 min read
Meta has started rolling out new Meta Muse models, trained with its Muse Wonder frontier model, to everyone using the Muse agent and the Meta AI assistant. The November 6 update brings two Wonder features, Active Memory and Active Focus, to free users, and Meta says Muse now forgets critical information 2.4 times less often.
The company announced the change in a Meta AI Research blog post on November 6. Muse Wonder, launched on October 12, stays reserved for harder tasks by paying subscribers. The new models are "specialized" versions of Meta's existing models, trained using Wonder. Meta has not detailed its method, but the approach brings some frontier behavior into models that are cheaper to run for a mass audience.
What Do the New Meta Muse Models Add?
Active Memory is meant to help the assistant remember what matters from earlier conversations and notice when an old detail becomes relevant again. Meta says it trained the models to tell temporary information from lasting facts and to spot details that have gone stale. On Meta's own Active Memory test, the current Muse model used critical context correctly in 90% of tasks, up from 76% for the previous model. That is where the 2.4x figure comes from: the failure rate dropped from 24% to 10%.
Active Focus is about persistence. When a task hits an unexpected problem, the model is supposed to retry, recover or find another route instead of giving up. Meta reports a 67% improvement in finishing everyday Muse tasks where something goes wrong, though it did not share the starting completion rate.
What Changes for Meta AI Users?
Meta AI is the assistant built into WhatsApp, Instagram, Facebook and Messenger, so this model swap reaches a very large audience at once. Meta did not publish separate numbers for Meta AI. Instead, it says the assistant now breaks complex problems into clear steps, asks sharper follow-up questions and keeps up when users change direction in the middle of a chat.
Mark Zuckerberg wrote on X that the upgrade brings Active Memory and Active Focus "to everyone." According to The New Stack, the models started shipping to all Muse and Meta AI users the same day. The update follows the recent Meta Muse voice rollout with expressive avatars in the US.
| Metric (Meta's own tests) | Previous Muse model | Current Muse model |
|---|---|---|
| Active Memory eval, critical context used correctly | 76% | 90% |
| Personal Superintelligence Index | 38% | 47% |
| Everyday tasks completed after unexpected issues | Baseline not disclosed | 67% improvement |
How Does This Compare With Rival AI Agents?
OpenAI and Anthropic are also pushing personal agents, but The New Stack notes that OpenAI's Dots and Anthropic's Duet are limited to their highest paid tiers. Anthropic is also adding personal context to its chatbot, as seen with Claude Personal Finance, which connects bank accounts through Plaid. Meta's angle is reach: it is putting agent-style memory and persistence into free apps that billions of people already use, rather than keeping them behind a top-tier plan.
Our Take
The gains look meaningful, but every number here comes from Meta's own benchmarks, with no task counts or independent testing. Better memory also means the assistant keeps more of your personal details, so users should check what Meta AI remembers and how to clear it. The bigger signal is the method: like the faster model progress METR measured at Anthropic, labs now use their frontier models to train the everyday ones, and that cycle is speeding up.