Meta's Muse Wonder models add active memory to curb agent drift

Meta’s Superintelligence Lab released two new models on Monday, Muse Wonder and Muse Wonder Full, which the company is calling its first frontier model release. Unlike the earlier Muse Spark, built for speed at low cost, the Wonder models are larger and aimed at harder tasks such as coding, according to Meta’s blog post announcing the launch.

A frontier model refers to a system pushing near the current limits of AI capability. In its model card, Meta says Wonder and Wonder Full score comparably to Fable 5.1 and GPT-6 Astra on benchmark tests, though the comparisons come from Meta’s own testing rather than independent review. “Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence,” the company’s blog post reads.

Meta is pricing Wonder at $4 per million input tokens and $20 per million output tokens through its Model API, putting it roughly in line with GPT-6 Sol and Opus 5.5. Wonder Full, the more capable of the pair, will not be sold through the API at all. It will only be accessible to subscribers of Muse Code, Meta’s coding product. Both models will also run underneath the Muse agent, Meta’s assistant product.

Memory as a guardrail

The editorial framing behind this release centers on guardrails for agentic systems, AI programs that take multi-step actions on a user’s behalf without constant supervision. Meta’s specific answer is something it calls “active memory,” the ability of a model to retain and reuse information a user has given it across a long stretch of time. The company points to MemoryBench, a benchmark designed to test continual learning and memory construction, where it says Muse Wonder now outscores Fable, Opus, and Astra.

“We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond,” research scientist Kavya Kopparapu said in a video accompanying the launch, according to TechCrunch. Meta has not published independent data showing how active memory translates into fewer agent errors or safer long-running behavior. The connection between better memory retention and an agent staying “on the rails” is implied by Meta’s framing rather than demonstrated with a separate safety benchmark.

The release is Meta’s first frontier-scale model since Scale AI founder Alexandr Wang took over the company’s AI efforts in a restructuring last August. In the time since, Meta shipped the lightweight Spark model and the Muse agent but had not put out a state-of-the-art release to rival the pace set by OpenAI and Anthropic. Meta says Wonder and Wonder Full were trained at its Prometheus data center in central Ohio, which came online earlier this year.

What to watch next is whether outside developers and benchmark groups replicate Meta’s claimed scores once Wonder is broadly available through the API, and whether Wonder Full’s memory gains hold up once Muse Code subscribers put it to use on extended, multi-step tasks.