Meta unveils Muse Wonder and Muse Wonder Full, new agent-focused AI models with active memory and focus

Meta on October 12 introduced Muse Wonder and Muse Wonder Full, two new AI models it describes as the most capable in its Muse line yet. The company positions the models as “frontier” systems for agentic tasks, meaning software that can interpret information, work through multi-step problems and, in some settings, carry actions forward over time.

In a post on its newsroom, Meta said it began an overhaul of its AI efforts about 14 months ago and has moved from rebuilding foundations to pushing toward what it calls frontier performance. The Muse Wonder series is described as built on groundwork Meta says it laid with an earlier model, Muse Spark, and is aimed at both personal and professional use cases.

Meta said the models are designed to handle tasks that require reading what is in front of the system and working through domain-specific problems. Examples given included navigating an insurance claim, understanding a lease, and making sense of someone’s finances. The company also said Muse Wonder models are built for expert domains including law, finance and health, alongside what it called frontier coding performance for developers building applications and games.

A central element of the announcement is that Meta is adding two capabilities intended to improve how agents operate over time. Meta said the models have Active Memory, which it described as the ability to carry what matters across conversations and bring back details relevant to the task at hand. It also said the models have Active Focus, intended to help a system continue pursuing a goal when challenges arise, try alternative approaches, recover from errors, and ask for clarification when needed.

Active Focus is framed as a reliability and follow-through mechanism. Meta said the models can “check in” before completing anything important, rather than immediately taking further steps after encountering uncertainty. The intent, as presented by Meta, is to make agentic workflows less brittle when an instruction is unclear or when intermediate attempts fail.

Meta also discussed privacy, safety and security work around the agent use case. It said Muse Wonder was trained to avoid sharing private information without permission, and to refuse requests that would support harmful acts. The company further said it worked to improve security and strengthen defenses against cyber attacks, and it highlighted that developers and security teams could use models like Muse Wonder to find flaws in their own code and strengthen systems.

On governance and process, Meta said it trains, evaluates and releases its most capable AI models in line with standards set out in its Meta Superintelligence Scaling Framework. The company said a full report with additional safety evaluations would be available “soon,” and it pointed readers to its research blog for more detail on how the models were built, how they were tested, and what they can do.

The practical business implications are largely about how software agents might change workflows. If Active Memory and Active Focus perform as intended, customer support, operations, and knowledge-heavy roles could see new tooling that keeps track of context across threads and continues working through multi-step tasks. That, in turn, could raise the bar for rivals competing to sell agentic AI experiences, both to consumers and to enterprise users who want reliability and controls.

There is also a security angle to the rollout. Meta’s emphasis on training to reduce unsafe behavior and to help defenders test code suggests it expects the models to be used in both offensive and defensive testing contexts, a pattern increasingly common across the AI industry. Still, Meta did not provide measurable performance benchmarks, deployment timelines, model availability details, pricing, or specific platform integrations in the announcement itself.

What to watch next includes the “soon” safety evaluation report Meta referenced, any technical or benchmark disclosures on performance and limitations, and whether Meta publishes additional guidance on how Active Memory and Active Focus are constrained in real deployments. Investors and customers will likely look for clarity on release channels, access for developers, and how Meta will handle privacy and permission controls as these agents are used across messages, calendars and documents.