Meta Muse Wonder adds Active Memory to keep AI agents on task

Muse Wonder is a family of two large language models (the kind of AI behind chatbots like ChatGPT) designed for agentic work, meaning tasks where the AI takes many steps on its own rather than answering one question and stopping. Meta’s own examples include building a working app, reading through a rental lease and handling an insurance claim.

The models come from Meta Superintelligence Labs, the unit Mark Zuckerberg set up after Meta fell behind OpenAI, Google and Anthropic. Meta paid $14.3 billion for Scale AI and hired its chief executive, Alexandr Wang, to run the lab as Chief AI Officer. The first model from that effort, Muse Spark, arrived in April. Muse Wonder followed about six weeks after the Muse Spark 1.3 update, and Wang said it was developed under the codename “watermelon.”

Training took place at Meta’s Prometheus data center in central Ohio. Meta has not said much more about how the models were built or what data went into them.

Active Memory and Active Focus are two abilities meant to stop an AI agent from forgetting what you asked or giving up when a task gets messy. They target a real weakness: on a big research job or a large software project, a model that drifts off course can make expensive mistakes that pile on top of each other.

• Active Memory lets the model carry the details that matter to you from one conversation to the next, then pull back the relevant ones when a new task needs them.
• Active Focus keeps the model working toward the stated goal when it hits a snag. It can try another approach, recover from errors, ask you to clarify, and check in with you before doing anything important.

That last part is the one to watch. Several recent incidents involving AI agents from other labs going off script have shown how easily a powerful model can stray when given open-ended instructions. Meta has not said whether Muse Wonder agents misbehaved in its own testing, and it says a fuller safety evaluation report is still to come.

Last month I wrote about how we can build a positive and safe future for everyone: https://t.co/1KtONlqBs5

Every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to ensure that happens.

The…

— Mark Zuckerberg (@finkd) September 15, 2026

On the Artificial Analysis Intelligence Index, an independent combined score of many tests, Muse Wonder Full scored 56. That ties it with Claude Sonnet 5.5 and places it second only to Claude Opus 5.5. The standard Muse Wonder scored 52, one point behind OpenAI’s GPT-6 Astra at 53.

Artificial Analysis said the bigger model placed in the top three on six of its ten tests and did especially well on agentic knowledge work and scientific reasoning. Meta also claims wins over GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on some narrower tests, such as working with legal documents and building software.

Real-world coding tells a more modest story. A few days after launch, the popular AI code editor Cursor added Muse Wonder. On CursorBench 4.0, a test of long coding jobs drawn from real software work, it scored 46.1% in its High setting and up to 50.4% in Max. Claude Opus 5.5 leads that test at 57.8%, so Meta’s model lands in the middle of the pack there.

Developers can use Muse Wonder now through the Meta Model API and through Muse Code, Meta’s AI coding tool. Through the API, every million tokens you send costs $4 and every million the model writes back costs $20 (a token is a small chunk of text, roughly part of a word), with cheaper rates for cached inputs. The model can read up to 1 million tokens at once and accepts text, images, video and audio.

Muse Wonder Full is more limited. It is only available to Muse Code subscribers on the $50 a month or $200 a month plans, and Meta has not announced API access for it.

Unlike Meta’s old Llama models, neither version can be downloaded and run on your own computer. Meta has released a small open-weight model called Muse Glimmer and says larger downloadable ones will come later, but Muse Wonder is not one of them. That also leaves open questions in Europe about publishing model details under the EU AI Act, and Meta has not said whether it will ever release the weights.

For most people, Muse Wonder will show up inside the Muse app rather than as a separate product. Meta says it will power Muse experiences soon, but only for paying subscribers, with plans reported at $20 or $100 a month. Muse itself has already passed 5 million app downloads.

An assistant that remembers you across chats is handy, and it is also a privacy question. If you give an agent access to your messages, calendar or documents, it can carry that information forward. Meta says the models are trained not to share private information without permission and to refuse harmful requests, such as help building weapons or hacking.

One setting deserves a check. Muse Code subscriptions reportedly default to a Contributor mode that lets Meta use what you submit to improve future models. You can switch it off. The standard API reportedly does not use submitted content for training.

If you plan to try it, start with low-stakes tasks, read the permissions screen before connecting accounts, and double-check anything the agent produces before you rely on it.