Meta Muse Wonder Wants to Remember You: Active Memory Comes to Frontier Models

Meta ships its first frontier models, and the pitch is memory

On October 12, Meta released Muse Wonder and Muse Wonder Full, the first models it describes as frontier class. The code name during development was Watermelon. The company says the pair is competitive with the best systems in the world and especially strong on the kind of work personal and professional agents actually do.

If you follow the labs, the framing is familiar and the features are not. What Meta is selling here is not a bigger slice of benchmark glory. It is a promise about continuity: an assistant that carries something about you from one conversation to the next.

That promise is worth taking seriously, because it points at a capability gap the industry has lived with for two years. Models have gotten better at almost everything except remembering. Muse Wonder is Meta's attempt to fix the one thing users complain about most.

Two capabilities do the heavy lifting

Meta calls them Active Memory and Active Focus, and the naming is unusually literal.

Active Memory keeps important details about a user across conversations and pulls them back when they become relevant to a new task. In practice, that means you stop re-explaining your stack, your preferences, or the constraints of a project every time you open a fresh thread. Alexandr Wang, who runs Meta Superintelligence Labs, put it plainly: memory matters because it means nobody has to reintroduce themselves over days, weeks, or months.

Active Focus addresses a different failure. Agents are good at starting and bad at finishing. Active Focus is designed to keep a system working toward a goal, trying alternate approaches when the first one stalls and adapting when new information arrives. It is the difference between a model that answers a question and one that grinds through a task until it is done.

Both features attack the same root problem from opposite ends. Memory gives an agent a past. Focus gives it a future. A system with neither is stuck in an eternal present, reacting to whatever you typed a second ago.

What Meta is claiming, and against whom

Meta published benchmarks showing the Wonder models matching or topping recent releases from OpenAI and Anthropic, including GPT-6 Astra and Claude Fable 5.1. WIRED framed the launch as Meta returning to the frontier after a costly rebuild of its AI organization. The models were trained at Meta's Prometheus data center, which the company describes as gigawatt scale.

Pricing is set at $4 per million input tokens and $20 per million output tokens. Muse Wonder is available through the Meta Model API and Muse Code, while the more capable Muse Wonder Full is reserved for Muse Code subscribers. Muse Wonder will also power future versions of Meta's Muse assistant, starting with paid users.

The pricing is notable for what it does not do. Meta is not trying to win on cost. It is priced roughly in line with the frontier, which signals that the company believes the capability, not the discount, is what will pull users over.

What memory looks like in a real workflow

It helps to be concrete about why Active Memory is a bigger deal than it sounds. Today, working with an agent across a long project means maintaining a mental ledger of context. You remind it of the framework, then the coding style, then the deployment target, then the fact that a particular service is deprecated. Every new session restarts that ledger from zero, and every restart is a chance for the agent to drift.

A model that holds those details changes the rhythm of the work. You spend your energy on the problem, not on re-briefing. Over a project that lasts weeks, the saved attention compounds, and the output stays more consistent because the agent is not re-deriving its assumptions every morning.

The catch is that memory is only as good as its hygiene. Old decisions become wrong decisions. A stored preference can harden into a bias. A remembered fact can be a remembered mistake. Good memory systems need to forget gracefully, and nobody has fully solved that yet.

The open-source question

Meta built much of its reputation in open models, and it is not abandoning that story. Wang pointed out that the company recently released the weights for Muse Glimmer along with the technology needed to put a Muse agent onto your own hardware. "We're still committed to open source," he said.

What Meta is not saying is whether the Wonder family itself will be opened. The silence is the interesting part. If frontier capability increasingly lives in memory and long-horizon focus, then the weights matter less than the harness and the infrastructure around them, and those are exactly the pieces a company might prefer to keep.

A step toward "personal superintelligence"

Meta's stated long-term ambition, in Mark Zuckerberg's phrase, is personal superintelligence: AI that understands an individual's goals and preferences rather than producing generic competence. The Wonder models are positioned as an early step, with more specialized models planned for broad deployment across Meta's apps.

That direction is coherent with the product it sells. Meta already owns the surfaces where hundreds of millions of people spend their attention. A model that remembers you is more valuable inside those surfaces than a model that merely scores higher on an exam. The company that knows your feed also wants to know your intentions, and memory is how it gets there.

The part that should give you pause

An agent that remembers personal information and acts on your behalf is exactly the kind of system that turns a memory feature into a security surface. What gets stored, where it is stored, who can retrieve it, and how a wrong memory propagates are not footnotes. They are the product.

Meta says it has strengthened safeguards against cyber and biological misuse and added protections for personal information, with a model card released alongside the launch and a fuller report to follow. Independent testing will decide whether the claimed advantages hold up once researchers outside the company get their hands on the systems. Until then, the benchmarks are the company's word, and memory-heavy systems are hard to evaluate from the outside because so much of their behavior depends on what they have been told to keep.

Why this release matters beyond Meta

The last two years of model competition were mostly about raw capability per dollar. Muse Wonder points at a different axis: how long a model stays useful to one person. Memory and focus are not glamorous, and they are hard to demo in a 30-second clip. They are also the things that separate a chatbot you visit from an assistant you keep.

If that axis turns out to be what users value, the competitive map changes. The labs that win will be the ones that make continuity trustworthy, not the ones that only keep climbing a leaderboard. And the labs that treat memory as a checkbox will find that a forgetful assistant is less useful and, in practice, harder to trust.