Meta upgrades the models behind Muse with better memory and persistence, scoring 47% on its own index
Specialized models trained with Muse Wonder add Active Memory and Active Focus. Meta reports 2.4x less forgetting of critical details and 67% better task completion when things go wrong.
DevIsDead Newsroom
Meta has rolled out new versions of the models that power Muse and Meta AI, trained using its top Muse Wonder models, according to Unite.AI, whose article was written by an AI research agent and reviewed by its editors. On Meta's own Personal Superintelligence Index, the current Muse model scores 47%, up from 38% for the previous one.
What's shipping isn't Wonder itself but tuned-up versions of the models Meta already runs, improved with Wonder's help. Meta said in October that Wonder would first power subscribers' Muse experiences and then be used to build specialized models for everyone; this is that second step.
Active Memory
The models are trained to hold on to important information and bring it back when it becomes relevant later, telling temporary details from permanent ones and recognizing when something has gone stale. Meta reports 2.4 times less forgetting of information critical to a user's goals, and its Active Memory Eval shows the model using critical context correctly in 90% of tasks, up from 76%.
Active Focus
This is about not giving up. Meta trained the models on a range of realistic obstacles so they can persist, recover or try alternatives. In Muse, it reports a 67% improvement in finishing everyday tasks when something unexpected happens. The models are also meant to juggle tasks, decide when to work independently versus with you, and take corrections mid-task without losing the thread. In Meta AI, the pitch is a better thinking partner that breaks problems into steps and asks sharper follow-ups.
Benchmarks vs. vibes
Every number here is Meta measuring Meta, on benchmarks Meta designed, including a "Personal Superintelligence Index" whose name does a lot of marketing work. The improvements are plausible and the targets are the right ones: memory and recovery are exactly where today's agents feel dumb. But the real test is whether your Muse remembers the thing you told it last week, and doesn't fall over the first time a website changes.
This lands the same week as Muse video calls and Spaces. Meta is clearly sprinting on agents.
Scoreboard update: AI +1. Meta says its agent now forgets less and gives up less.
Stay undead.
Comments 0