Meta Unveils New AI, “Muse Wonder,” With Better Memory Across Conversations and Greater Persistence on Difficult Tasks

Meta announced its new AI model, “Muse Wonder,” and its higher-end version, “Muse Wonder Full,” on October 12, 2026 (local time). The models are now available through an API for developers and Muse Code, Meta’s coding assistant, and the company plans to bring them to its personal AI agent, Muse, soon. Their key selling points are the ability to remember information across conversations and see difficult tasks through to completion. The announcement raises the question of what really matters when you delegate work to AI.

AI needs the ability to remember and see things through

The question “Is this AI smart?” is starting to feel insufficient. Sending emails, booking trips, finding properties and even making inquiries—when you’re delegating tasks like these, it matters more whether the AI remembers what you discussed yesterday and keeps going when it hits a snag than how good a single response is. It’s a lot like assigning work to a new hire. Those are precisely the two abilities Meta put at the center of this announcement.

Active Memory remembers across conversations

Active Memory is the ability to retain information that matters to users across conversations and independently recall details relevant to a task. Meta says the model remembers small habits and preferences, as well as even casual exchanges from the past. Researchers say they focused on making sure it does not forget important things, even after days, weeks or months.

Meta cited MemoryBench as a measure of memory. In the company’s internal evaluations, Muse Wonder scored higher than competitors Claude Fable 5.1, Claude Opus 5.5 and GPT-6 Astra. It’s worth noting that these figures come from internal evaluations.

Active Focus helps the model recover when it gets stuck

The other feature, Active Focus, is the ability to keep working toward a goal even on difficult tasks. When it gets stuck, the model tries a different approach; when errors occur, it recovers and completes the task. Meta says it is designed to ask for clarification when needed and get the user’s approval before taking important actions.

The aim is to prevent AI from losing sight of its objective and compounding errors during complex research or large-scale software development. One way to think about it is that Meta is aiming for a subordinate with memory and persistence who also knows how to report, communicate and consult—the so-called “ho-ren-so.”