Meta Muse Wonder bets on memory, not just benchmarks

What this is about

Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026. They are designed not merely to answer isolated questions, but to pursue personal and professional tasks over longer periods. Meta names two capabilities: Active Memory preserves important information from earlier conversations, while Active Focus is intended to try alternative routes when obstacles arise and ask before important actions.

This matters to real users because digital assistants often fail at transitions. They forget earlier requirements, stop after an error, or carry a request too far. Muse Wonder targets that gap directly. At the same time, Meta is initially making access a paid offering, while key safety details are due only in a later report.

What Muse Wonder actually does

Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is initially limited to Muse Code subscribers. Muse Wonder is also planned to power the Muse app later, beginning with paying users.

According to Meta, Active Memory means that the model retains relevant information across conversations and retrieves it when a related task appears. Active Focus describes its ability to keep working toward a difficult goal, try other approaches, recover from errors, and ask questions when something is unclear. The model accepts text, images, video, and audio. Independent measurements from Artificial Analysis report a one million token context window for both variants.

Meta positions the models for coding and work in law, finance, and health. Some performance claims come from internal tests. Based on Artificial Analysis, The Decoder reports 56 points for Muse Wonder Full and 52 for Muse Wonder on the Intelligence Index. On Terminal Bench 4.0, the result rose from 33 percent for Muse Spark 1.3 to 57 percent.

Why it matters

The practical competition is shifting from the best single answer to reliable work across many steps. An assistant that reviews a lease, considers appointments, and asks the right follow-up question later needs useful memory and explicit boundaries for actions. Meta wants to bring that combination into its products.

The price shows that stronger agent performance is not automatically cheap. CNBC reports Muse app subscriptions at 20 or 100 US dollars per month. Muse Code plans are listed at 50 or 200 US dollars. The Decoder puts Muse Wonder at 4.52 US dollars per task in the Artificial Analysis test, compared with 1.60 US dollars for its Muse Spark 1.3 predecessor.

Privacy also becomes more important. A personal assistant may see messages, calendars, and documents. Meta promises safeguards against unauthorized disclosure and harmful use. However, a full report with additional safety evaluations had only been announced and was not available at launch.

In plain language

Imagine a travel companion who knows not only today's train connection, but also remembers that you dislike tight transfers and will need a wheelchair space tomorrow. Active Memory would be the notebook. Active Focus would be the ability to search for another route after a cancellation, while still asking before buying a paid ticket.

A good notebook does not guarantee a good decision. If an earlier detail is wrong, outdated, or too private, memory can carry that problem into several later tasks. Controls, deletion options, and visible confirmations therefore matter as much as model performance.

A practical example

A small agency asks Muse Wonder to revise a product page. The model receives 40 requirements, five existing files, and a rule that no purchase may happen without approval. After the first test, an interface fails. Active Focus is supposed to keep it from simply stopping by examining the error, choosing a second approach, and testing again.

A week later, a similar task arrives. Active Memory could recall the preferred tone, target audience, and approval boundary. Before any paid API call, the system would still need to ask again. The value therefore comes not from blindly continuing, but from combining memory, persistence, and clear human decision points.

Scope and limits

First, strong benchmark scores do not prove reliable everyday work. Meta publishes its own evaluations, while independent testing shows progress but also a significant factual knowledge weakness compared with several rivals.

Second, the price is rising. According to Artificial Analysis, a Muse Wonder task costs almost three times as much as one using Muse Spark 1.3. For frequent or long workflows, the extra benefit must justify that increase.

Third, safety and privacy questions remain open. Meta only announced its more detailed additional safety report. Anyone processing sensitive health information, contracts, or internal messages should therefore avoid deploying the model without access boundaries, audit logs, and human approvals.

SEO & GEO keywords

Meta Muse Wonder, Muse Wonder Full, Active Memory, Active Focus, personal AI agents, Muse Code, Meta Model API, AI memory, agent security, AI model pricing