Meta's Muse and Meta AI get better at remembering and finishing tasks

Meta says the assistants inside Muse and Meta AI now remember important details better and are less likely to give up when something goes wrong. The company is rolling out updated versions of its existing models, trained using its newer Muse Wonder series. Meta's own numbers show gains in memory and in finishing everyday tasks.

What changed

Meta introduced the Muse Wonder series a few weeks ago and calls it its most capable set of models to date. The new releases are specialized versions of older models, trained using Muse Wonder. The idea is to bring some of its abilities to more people. Two features carry over: Active Memory and Active Focus.

Memory that sorts what matters

Active Memory is meant to help the assistant remember what matters and notice when earlier information becomes useful again. Meta says the models learned to tell temporary details from permanent ones, and to spot what is still relevant versus out of date. In Muse, Meta reports a 2.4x reduction in forgetting critical information needed to help you reach your goals.

• On Meta's Active Memory test, tasks that correctly used key context rose from 76% to 90%.
• On Meta's Personal Superintelligence Index, the current Muse model scored 47%, up from 38%.

Sticking with a task

Active Focus is about persistence. Meta says it trained the models on real-world bottlenecks so they can try harder, recover, or find another route on your behalf. In Muse, the company reports a 67% improvement in completing everyday tasks where unexpected issues come up. Meta also says the models handle several jobs at once and can adjust when you clarify or redirect them mid-task.

What it means for Meta AI

Meta AI gets the same upgrades, but the focus is on conversation. Meta says it can break complex problems into clear steps, ask sharper follow-up questions, and keep up when you change direction partway through a chat.

What is not known yet

Meta did not say when each person will see the update, or whether it reaches everyone at once. It also did not explain how its tests were built. The 76%, 90%, 47% and 38% scores come from Meta's own evaluations, and the article cites no outside check. Real-world results may differ, so judge the assistants by how they handle your own tasks.