- Meta launched Muse Wonder and Muse Wonder Full, agent models designed to remember personal context across conversations and keep working on long tasks.
- Meta's benchmark and safety claims are its own, and it says a full safety report will be available soon.
- Before any rollout, ask what data an agent can reach, which actions need approval, and how its handling of private data was tested.
An assistant that remembers you is useful for one reason: it knows things about you. It knows your messages, your calendar and your private documents. That is also why it needs closer scrutiny than a chatbot. A model that holds your context and acts on it is closer to a new employee with a set of keys than to a search box.
That is the idea behind Meta's latest launch. On October 12, Meta introduced Muse Wonder and Muse Wonder Full. It calls them its most capable models for personal and professional agents. The lesson for leaders is that the main question about these agents is not how smart they are. It is how much access they are given, and who has verified how they use it.
What Meta released
Meta says Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. Meta also says Muse Wonder will soon power its Muse agent, starting with subscribers. The Verge reports that Meta has given no specifics or timeline for that rollout.
Meta says the models compete with the industry's best. The Verge reports that Meta compared them with OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Claude Fable 5.1, using results from internal testing. The Verge adds that it is unclear how they compare with Google's Gemini 4 Argon.
How the two headline features work
Meta describes two ideas. The first is Active Memory. In Meta's words, this means carrying important context about you across conversations and retrieving the details relevant to the task at hand. In plain terms, the model keeps a profile of you and looks things up in it when needed.
The second is Active Focus. Meta says this is how the model keeps working toward a hard goal. It tries other approaches, works through blockers and adapts to new information. Meta says it also asks for clarification when needed and checks in before important actions. Meta measures this with two tests it names, FocusBench and SWE-Marathon, which covers long-running software projects.
Meta also reports a test of how well skills carry over. It compared its earlier Muse Spark 1.3 with Muse Wonder Full on tasks the models were trained for and on tasks they were not. Meta says Wonder Full does better on unfamiliar tasks than the earlier trend would predict. It credits larger base models. This matters for agents, Meta says, because they face combinations of goals and tools nobody can predict.
Where the risk sits
Meta is direct about the exposure. It says a personal agent takes actions on a user's behalf and might handle messages, calendars and private documents. It says that level of access demands discretion with private information and resistance to misuse. Meta says it trained the models to lead on both, including robustness to adversarial instructions.
That is the right bar. But the evidence for it is not yet public. Meta says a full report with additional safety evaluations will be available soon. The performance comparisons come from Meta's own evaluations, as the Verge notes. Neither point proves the claims wrong. They mean buyers cannot yet check them independently.
Memory adds a second concern. A chatbot that forgets you at the end of each session holds little. An agent that retains context across conversations builds a store of personal detail over time. Meta's post does not say how that store is protected, who can see it or how it can be cleared. Those are questions for any team that wants to deploy it.
What leaders should ask
Nothing here requires a decision today. Muse Wonder is offered through Meta's API and to Muse Code subscribers, and the Muse rollout has no date. But teams will be asked about agents like it soon, so prepare the questions now.
First, what can the agent reach? List the inboxes, calendars and document stores, and give it the narrowest access that works. Second, which actions need a human to approve them? Meta says the model checks in before important actions. Ask how that is defined, and who can change it. Third, what does the vendor store about each user, and for how long? Fourth, has anyone outside the vendor tested how the agent handles hostile instructions hidden in the content it reads?
Wait for the full safety report before treating the safety claims as established. Read it for method, not only for scores.
A capable agent earns trust the way a new hire does: by starting with limited access and proving itself in the open.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Meta.
