
Breakdown / AI assistants · Model evaluation
Muse Wonder's memory promise is about using what it remembers
Meta describes an assistant that carries context across conversations and stays with a task. A remembered preference is only the beginning.
Suppose you've told an AI assistant that you avoid early flights. A week later, you ask it to plan another trip. Remembering the preference helps. Using it while comparing routes, keeping the budget in view and asking before booking is a much bigger job.
That is a useful way to understand Meta's pitch for Muse Wonder, the model family announced on October 12 for personal and professional assistants. Meta calls its two central ideas Active Memory and Active Focus. The first is about retrieving relevant context from earlier conversations; the second is about continuing towards a goal, handling blockers and checking in before important actions. Meta's research overview
Our trip example illustrates the promise. It isn't a test we've performed in Muse Wonder.
Remembering the right thing at the right time
An assistant could correctly recall that you dislike early flights and still recommend one without noticing the conflict. It could honor that preference but lose the budget. It could also keep applying an old preference after you have changed your mind.
A useful memory test therefore needs more than a question with a remembered answer. Give the assistant a task where an earlier detail matters, then change that detail and see whether the plan changes with it. Ask it to explain which information shaped the decision. Those are tests we want to run, not results the launch establishes.
The model and the app are different questions
Meta's October 12 announcement made Wonder available through its model API and Muse Code, its coding product, with Wonder Full available to Muse Code subscribers. At that point, it described a later rollout to the personal Muse assistant, beginning with subscribers. That dated announcement does not establish which model a particular account is using now. We haven't verified account-level access for this article.
The distinction matters when comparing experiences. A model can have a capability that an app has not exposed, or that depends on the tools and information the app supplies.
Read the comparisons alongside the method
Meta's model card includes internal evaluations. Its controlled internal coding comparisons use the same setup for each model, while externally reported coding results retain their original tools and agent setups. These are different kinds of comparison. The card also acknowledges errors in multi-step reasoning and limitations in three-dimensional spatial tasks. Muse Wonder model card
We haven't independently compared Wonder with other models. For a person using an assistant, the useful question is whether it carries the right information through a real task and finishes more of the work. For a developer, it's whether that holds inside their own application, with their own tools and messy inputs.