Meta’s Muse Wonder brings memory and focus to AI agents for work

Meta unveiled Muse Wonder and Muse Wonder Full on October 12, 2026, two new artificial intelligence (AI) models designed for personal and professional tasks.

The models bring together two capabilities Meta calls Active Memory and Active Focus, designed to help AI agents retain relevant information, work through complex tasks, and adapt when circumstances change.

Meta describes Muse Wonder as its most capable model series yet, with improvements in coding, professional tasks, and long-term contextual memory. The announcement also marks another step in the company’s efforts to build AI agents that can operate across applications rather than respond to individual prompts.

Memory and focus

At the centre of Muse Wonder are two capabilities that Meta believes are important for making AI agents more useful.

Active Memory allows the models to carry relevant information across conversations and retrieve it when needed. Instead of requiring users to repeatedly explain their preferences or previous instructions, an agent can draw on earlier context while handling a new task.

“One of the most critical steps towards useful personal and professional agents is reliable memory,” Meta said in its announcement.

Active Focus addresses a different challenge. It enables the models to continue working towards a goal over an extended period, try alternative approaches, work through obstacles, and adapt to new information.

Meta says the agents can also ask for clarification when instructions are unclear and seek user approval before taking important actions.

Together, these capabilities are intended to support tasks that cannot always be completed in a single interaction. An AI agent preparing a document, planning an activity, or working on a software project may need to remember earlier decisions while responding to new requirements.

This is particularly relevant for business applications, where workflows often involve multiple systems, instructions, and approval stages.

Beyond everyday tasks

Meta is positioning Muse Wonder for a wider range of professional activities, including software development, legal work, financial analysis, and tasks involving health information.

The company has tested the models on specialised evaluations that measure how well AI systems follow detailed instructions, interpret source documents, and complete work requiring domain knowledge.

These include Harvey’s Legal Agent Benchmark, which evaluates the ability to produce legal work from source materials, and Vals Finance Agent V2, which measures performance on finance-related tasks.

Meta has also assessed the models using AutomationBench-AA for business workflows, MemoryBench for contextual memory, and SWE-Marathon and DeepSWE for sustained software engineering.

According to the company, Muse Wonder and Muse Wonder Full delivered state-of-the-art results across several of these evaluations.

Independent benchmarking provides additional context. Artificial Analysis reported that Muse Wonder scored 52 on its Intelligence Index, while Muse Wonder Full scored 56, compared with 48 for the earlier Muse Spark 1.3 model.

The results indicate an improvement in overall capability, particularly for agentic tasks. However, benchmark performance does not necessarily establish how reliably these models will handle unpredictable work in live enterprise environments.

Meta has also examined how well Muse Wonder Full performs on tasks that were not specifically targeted during post-training.

The company says the model performed better on unfamiliar tasks than the trend established by Muse Spark 1.3 would predict, suggesting broader capabilities beyond the tasks emphasised during training.

Coding gains

Software development is another area where Meta is placing significant emphasis.

The company says Muse Wonder can manage complex coding tasks involving multiple files, changing requirements, and repeated testing and revision.

Such work requires more than generating code from a prompt. An AI agent needs to track the requirements of a project, understand previous changes, and continue working through problems without losing the context of what has already been completed.

Meta trained both models on diverse, long-running tasks to improve their ability to sustain progress through software development workflows.

The capabilities extend from smaller applications to more complex projects, according to the company.

For enterprise technology teams, this could support development workflows that require ongoing code changes and coordination across different tasks. However, the usefulness of these capabilities will depend on how accurately agents operate within existing development environments and how effectively their work can be reviewed.

Availability and deployment

Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is currently available to Muse Code subscribers.

The difference in availability is significant. Developers can access Muse Wonder through an API, but the more capable Full version is not being offered through the same channel at launch.

Meta plans to bring Muse Wonder into its Muse platform, initially for subscribers. The model will support more complex experiences, generate artefacts, and carry out tasks requiring additional reasoning and execution capabilities.

Over time, Meta also intends to use Muse Wonder to develop specialised models for wider access through Muse.

Both models were trained using Meta’s gigawatt-scale Prometheus data centre, reflecting the infrastructure investment behind the company’s latest AI development efforts.

The launch builds on Muse Spark and comes after Meta began restructuring its AI efforts more than a year ago.

While the earlier model focused on establishing Meta’s approach to model scaling, Muse Wonder places greater emphasis on sustained execution, contextual memory, and professional applications.

Safety and privacy

As AI agents take on more responsibility, they may also require access to sensitive information.

A personal agent could work with emails, calendars, and private documents, while a professional agent might handle business records, financial information, or internal software systems.

This makes data access and user control important considerations, particularly when agents can act across applications.

Meta says it trained Muse Wonder to handle private information carefully and resist adversarial instructions intended to manipulate its behaviour.

The company also claims that Muse Wonder includes its strongest safeguards to date against chemical, biological, radiological, and nuclear risks, as well as cyber misuse.

At the same time, it has retained advanced cybersecurity capabilities for defensive applications under its Meta Superintelligence Scaling Framework.

Meta has published initial safety evaluations covering areas such as unauthorised code execution and personal information handling. A more detailed report with additional safety assessments is expected.

For enterprise customers, these safeguards will need to translate into reliable controls over what information an AI agent can access, which actions it can perform, and when human intervention is required.

From models to agents

Muse Wonder showing a big change in AI development, where the ability to complete work over time is becoming as important as the quality of an individual response.

Meta is bringing memory, sustained execution, and professional task capabilities into the same model series, with plans to make them available through developer tools and consumer applications.

Its reported benchmark improvements suggest progress in these areas, although their practical value will become clearer as developers begin using the models in more demanding environments.

For businesses, the significance lies in whether AI agents can handle work involving multiple applications, changing instructions, and sensitive information without requiring constant supervision.

Muse Wonder brings Meta further into that market. Its success will depend not only on how much an agent can remember or how long it can work, but also on whether users and organisations can rely on the decisions it makes along the way.