Meta rolls out memory and task recovery to Muse users

Meta began rolling out unnamed specialized models to everyone using its Muse personal agent and Meta AI on 6 November. The models bring two features introduced with Muse Wonder: Active Memory, which carries details between conversations, and Active Focus, which is designed to help tasks continue when problems arise.

Muse has a free tier, while subscribers retain access to Wonder itself for harder tasks; the specialized models were trained using Wonder. Meta AI also operates inside WhatsApp, Instagram, Facebook and Messenger.

Muse launched in September on the smaller Muse Spark model. Meta introduced Wonder and the more capable Wonder Full on 12 October, The New Stack said.

What the models are designed to do

Meta says it trained the models to distinguish temporary information from lasting details and identify what remains relevant. It also says the models can recover or find alternatives when they encounter obstacles, handle corrections during a task and better judge when to work independently or check with the user. Meta has not detailed its training method.

In Meta’s internal evaluation, the new Muse models correctly used critical context in 90% of tasks, compared with 76% for the previous model. Meta also said completion of routine Muse tasks when unexpected problems arose improved by 67%, but did not share the baseline for that figure.

Wonder is also available through the Meta Model API and in Muse Code, Meta’s coding agent. Wonder Full remains limited to Muse Code’s two most expensive plans.