Meta Muse Wonder Explained: What Active Memory and Active Focus Actually Do
Meta launched Muse Wonder and Muse Wonder Full as frontier models for agents. Here is what is verified, what is a company claim, and who can use them today.
In this article 8 sections
Muse Wonder is Meta's new family of frontier models for personal and professional AI agents, launched October 12, 2026 as Muse Wonder and the larger Muse Wonder Full. Two features define the pitch: Active Memory, which carries context across conversations, and Active Focus, which keeps a model working through blockers. Most performance claims remain Meta's own.
Protocol Wire covers frontier model releases and the agent tooling around them, translating launch posts into what they change for engineering teams. This explainer is for developers, product leads, and technical buyers deciding whether Muse Wonder deserves a trial. It separates what is documented from what is still a claim.
What Did Meta Actually Launch?
Meta launched two models. Muse Wonder is available through the Meta Model API and Muse Code, and Muse Wonder Full is available to Muse Code subscribers. According to Meta's own announcement, Muse Wonder will also soon power experiences in Muse, Meta's personal agent, rolling out first to subscribers.
Both follow Muse Spark, Meta's first model family. Fortune reports that the new models are larger and slower than Spark, and that Muse Wonder was trained at Meta's gigawatt-scale Prometheus data center. Meta says the work began about fourteen months ago.
The company pitches the models for domain tasks in law, finance, and health, such as understanding a lease or a medical document. It also pitches them for coding projects that range from small apps to complex software, and for working directly in browsers and across apps.
How Do Active Memory and Active Focus Work?
Active Memory keeps important facts about you across conversations and pulls the relevant ones into the task at hand. Active Focus keeps the model going on a hard goal by trying other approaches, working past blockers, adapting to new information, and checking in before important actions.
Memory is the more familiar idea. Chat products already save preferences, and developers already build retrieval layers. What Meta claims is tighter integration, where the model decides what to save and recall without a separate system. The trade-off is privacy: a model with deeper access to messages, calendars, and private documents needs strong controls over what it keeps.
Focus targets a known weakness. Agents often stall after the first failed attempt or drift from the goal. Meta says it evaluates persistence with FocusBench and SWE-Marathon, a coding benchmark built around long projects. Meta's post names those benchmarks but gives no scores in its text, so there is nothing to verify yet.
How Does Muse Wonder Stack Up Against Rivals?
Independent data is limited to one index. On the Artificial Analysis Intelligence Index v4.3 at maximum reasoning effort, Muse Wonder scores 52 and Wonder Full scores 56, compared with 48 for Muse Spark 1.3, as TestingCatalog summarized. A social media post quoted there puts Wonder Full two points behind Claude Opus 5.5 on that index.
This table sets out what each source supports, as of October 22, 2026.
| Claim | Source | Status |
|---|---|---|
| Wonder Full scores 56 on the Intelligence Index | Artificial Analysis, via TestingCatalog | Independent score |
| Compares with GPT-6 Astra and Claude Opus 5.5 | Meta, via Fortune | Company claim |
| State-of-the-art on agent and coding benchmarks | Meta | No scores in text |
| Strongest safeguards against CBRN and cyber misuse | Meta | Report pending |
The honest summary is that Wonder Full sits near the top tier, not clearly above it. A two-point gap on one index is within the range where task choice matters more than the ranking. Run your own tests on your own workload before you decide.
Who Is Muse Wonder Built For?
Fortune's read is that Meta is concentrating on consumer platforms, unlike rivals that target enterprise work. It notes the agents suit tasks like handling an insurance claim or building a small business's mobile app more than expense reports or contract review. That fits Meta's position, since its users and its data live in consumer apps.
Business reasons sit behind the product. Digital advertising accounts for roughly 98 percent of Meta's $201 billion in annual revenue, according to Fortune. A paid agent could diversify that, though it would take many subscribers to matter. Deutsche Bank's Benjamin Black suggested a personal AI agent could reach up to 8 percent of revenue by 2030, about $36 billion. Meta's capital expenditure is about $145 billion, so the payoff question remains open.
For developers, the entry points are Muse Code and the Model API. Muse Code subscriptions have been reported at $50 or $200 a month, but that figure comes from a single news report and Meta's post lists no prices, so confirm it before you budget.
What Should Teams Verify Before Trying It?
Start with data handling. Ask what Active Memory stores, where it is stored, how a user deletes it, and whether it is used for training or ads. A memory feature that works well in a demo can become a compliance problem in a regulated team.
Then test persistence on real work. Give the model a task with a likely first-attempt failure, such as a flaky integration test or an API with unclear docs, and watch whether it recovers or loops. Compare against another model on the same task, and record cost and time. Our look at Shopify's AI agent checkout shows how quickly agent claims meet real-world edge cases.
Next, check the commercial terms. Pricing for Wonder Full is tied to a subscription, which affects team use. When you compare models on price, remember that cheaper audio and video processing, as covered in our analysis of omni-modal model pricing, can change what you build more than what you save. The same caution applies here.
Finally, wait for the safety report if your use involves private data. Meta calls Muse Wonder its most aligned model to date, trained for robustness against adversarial instructions and careful handling of private information. That is a promising claim, and it is exactly the kind that needs a published report behind it.
How Does Muse Wonder Relate to Muse Spark Open?
They are separate releases on separate tracks. Meta introduced Muse Spark in April as its first model family, then shipped Muse Spark Open on October 15, three days after Wonder. Spark Open is an open-weight model with about 816 billion total parameters and a context window of up to one million tokens, released under Meta's own Muse Model License.
Wonder is the opposite approach. It is a hosted frontier model, reachable through Meta's API and subscription tools, with no downloadable weights. Meta is running two strategies at once: a closed flagship for agents and an open-weight model that developers can host themselves.
The licensing difference matters for buyers. Commentators are clear that Spark Open is open-weight, not open source, and secondary reports say the license asks larger companies to sign a separate agreement before offering the model as a service. Read the license text before building a product on it. Wonder, by contrast, comes with ordinary API terms. Legal and procurement teams should still review both before any production use, and record which version of each model was tested.
For a team choosing between them, the question is control versus capability. Self-hosting Spark Open gives you data control and predictable cost at scale, but it needs serious hardware. Wonder offers a stronger agent stack with no infrastructure to run, in exchange for sending data to Meta.
Where Does This Fit in the Agent Market?
Muse Wonder arrives in a crowded field. Salesforce's agent platform, covered in our report on the Claudeforce open beta, targets enterprise workflows, while Meta aims at people and small businesses. The two do not compete head to head yet, though both depend on the same underlying question: will users trust an agent with real actions?
That trust question will decide adoption more than benchmark gaps. Features like Active Focus, which checks in before important actions, are a design answer to it. Whether users find those check-ins reassuring or annoying will show up in retention numbers long before it shows up in a leaderboard.
Protocol Wire's editors will revisit the models once Meta publishes its safety report and benchmark scores, and once independent testers have run them on long coding tasks.
Conclusion: Promising, Unproven, Worth a Small Trial
Muse Wonder looks like a serious entry. Wonder Full sits near the top on the one independent index available, and the agent features target real weaknesses. The safety report and benchmark scores have yet to land.
Keep notes as you go, because model behavior can change after updates. Your next step is a small, bounded trial. Pick two tasks, one coding and one personal-assistant style, run them against your current model, and compare results, cost, and data handling before you commit anything bigger.
