Meta says new Muse models forget 2.4x less

Meta finally put numbers on Muse Wonder. It also wrote the test.

On Nov. 6, Meta said it is rolling out specialized versions of its existing models, trained using Muse Wonder, to people who use Muse and Meta AI. The new models bring two features Meta calls Active Memory and Active Focus to both apps.

What Meta says changed

Active Memory is meant to help the assistant remember what matters and notice when something you said earlier becomes relevant. Meta says the models are trained to tell temporary information from permanent information, reason across time and separate what is still current from what is stale.

Active Focus is about persistence. Meta says the models are trained to "try harder, recover, or find alternatives on your behalf" when a task runs into trouble.

Meta's headline figure is its own:

47% on Meta's Personal Superintelligence Index, up from 38% for the previous Muse model.

On Meta's Active Memory eval, the share of tasks that correctly used critical context rose to 90% from 76%. Meta also reports a 2.4x reduction in forgetting critical information and a 67% improvement in completing everyday tasks where unexpected issues come up.

Meta's numbers, graded on Meta's tests

When Muse Wonder launched in October, its announcement came with no benchmark scores or prices. Now there are scores. Both tests are Meta's own, and the post does not explain how the index is scored or point to an outside run of either test.

Those are claims. They are not results yet.

Still, Meta picked the right thing to fix. An assistant that keeps your trip dates straight across a week is worth more to you than one that wins a reasoning chart. AI memory is the feature that turns a chatbot into something you hand errands to.

More memory also means more of your life stored on Meta's side. This is the app that paused chat sharing in October after 6,800 shared links surfaced. Memory makes that trade bigger, in both directions.

What it means for you

Meta's post gives no rollout schedule, no regions and no plan tiers. You may already have the new models, or you may not.

If you use Muse or Meta AI, run your own test. Tell it three facts about a plan on Monday. Ask about them on Thursday. Change one in between and see whether it notices.

That is the eval that matters to you, and it costs nothing.

If you build assistants yourself, note the vocabulary. Meta is training models to treat some information as temporary and some as permanent. Expect other assistants to be judged on the same split.

What to watch

An independent benchmark run of Muse against other assistants on memory tasks, and whether Meta publishes how its Personal Superintelligence Index is built.

Until then, grade it yourself.