Meta introduces Muse Wonder with memory and sustained agent work

Muse Wonder reaches developers first

Meta's announcement introduces Muse Wonder and Muse Wonder Full, following the foundations laid with Muse Spark. The company positions the series for personal and professional agent tasks, including coding and work involving complex documents.

Muse Wonder is available through the Meta Model API and Muse Code. The announcement describes a forthcoming rollout in the Meta AI application, starting with subscribers. Developer availability and the planned consumer rollout are separate parts of the release.

Memory and persistence address different needs

Meta calls its cross-conversation context capability Active Memory. The stated purpose is to retain relevant details and retrieve them when needed for a later task.

Active Focus describes sustained work when a task becomes difficult. Meta says the models can try alternative approaches, recover from errors, request clarification and check with the user before important actions. Those are vendor-described behaviors, rather than a Franklin hands-on evaluation of reliability.

The company also describes training intended to avoid sharing private information without permission and to reject harmful requests. It says a fuller report with additional safety evaluations is forthcoming. The announcement does not establish that every possible action or information-sharing risk has been resolved.

Test the handoff between remembering and acting

Franklin analysis: an agent that recalls useful context and continues work could reduce repeated explanations across a project. The same workflow needs clear boundaries around what it remembers and what requires a fresh decision from the user.

A focused trial should use a familiar task, deliberately change one relevant detail, and observe whether the agent uses the current instruction. Include a point where it should pause for clarification or approval. That tests the claimed memory and persistence together, while keeping the evaluation tied to a concrete task. Await the detailed evaluation report before treating broad capability claims as a complete comparison with other models.