Muse Wonder’s Active Memory and Active Focus for agent-style work
Meta AI Research ● Covered by 10 sources
Muse just launched Muse Wonder, its new agent model. It’s built to remember context and stay on task, which is the whole trick with AI agents.
Based on reporting by Meta AI Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Muse has a new pitch for what AI should do next: not just answer, but follow through. The company introduced Muse Wonder and Muse Wonder Full as its most capable models yet, aimed at personal and professional agentic work. Muse says they set a new high point for tasks like coding, memory, and long-running workflows.
The two headline features are Active Memory and Active Focus. In Muse’s telling, Active Memory means the model can carry important context across conversations and pull it back when it matters. Active Focus is the part that keeps it working through blockers, trying another approach, asking for clarification when needed, and checking in before it takes major actions.
That sounds abstract until you look at the use cases Muse keeps coming back to: legal work, finance, health, calendars, messages, private documents, and codebases that don’t politely end after one prompt. The company says Muse Wonder and Muse Wonder Full perform well on benchmark suites for personal and professional agents, memory, and coding, and it highlights tests like FocusBench and SWE-Marathon for sustained effort on long projects.
Muse is also pushing the familiar AI argument that bigger base models help in stranger situations. It says Muse Wonder Full does better on unfamiliar tasks than its in-distribution scores alone would suggest, and that matters because personal agents have to handle messy combinations of goals, tools, and constraints. The models were trained at Muse’s Prometheus data center, and Muse Wonder is available now through the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. Muse says Muse Wonder will soon power experiences in Muse, starting with subscribers, and eventually support artifact generation and more intelligence-heavy processes.
My take — AI-written commentary, not fact-checked reporting
This is the right bet, and also the obvious one: if an agent can’t remember, focus, and keep its hands off your private stuff, it’s just a chatty liability. The industry keeps selling “personal superintelligence” while quietly rediscovering that reliability is the hard part. Fancy demos are cheap; not being weird with someone’s calendar is the real benchmark.
Read more about this at: Meta AI Research
