Meta Gives Muse and Meta AI Better Memory and Follow-Through, Trained by Muse Wonder

Meta has started rolling out new models to everyone who uses Muse and Meta AI. These aren't Muse Wonder itself, the frontier model Meta introduced a few weeks ago. They're specialized versions of Meta's existing models that were trained using Muse Wonder, so the everyday assistants pick up some of its best habits.

Active Memory: knowing what to keep

The headline feature is Active Memory. Meta says the models were trained to tell temporary details from lasting ones, to reason across time, and to notice when something you said earlier has gone stale or suddenly matters again. In practice that should make it easier to pick up where a past conversation left off.

Meta's own numbers: inside Muse, the new models forget critical information 2.4 times less often. On Meta's internal Active Memory eval, the current Muse model used critical context correctly in 90% of tasks, up from 76% for the previous one.

Active Focus: not giving up halfway

Active Focus is about persistence. Meta trained the models on real-world bottlenecks so they try harder, recover, or look for another route when a task hits a snag, and it reports a 67% improvement in finishing everyday tasks where something unexpected goes wrong. The models are also meant to handle being redirected mid-task without losing the thread. On Meta's Personal Superintelligence Index, the current Muse model scores 47%, against 38% for the one it replaces.

My take

Memory is the thing I notice most with assistants. If I'm tracking a season's simulcasts or a long manga run, I want it to remember what I've already watched and drop plans I've cancelled, not mix the two up. Still, every figure here comes from Meta's own evaluations, and the post doesn't share how those tests were built, so I'd treat them as a starting point until people put the update through real use.

Source: Meta AI Research