News

Releases & Models

CONFIRMED: Meta Muse Wonder bets on agents that remember and persist

Meta launches Muse Wonder and Wonder Full with memory across conversations and a focus on completing tasks. Access starts through the API and Muse Code; savings and safe autonomy still need proof in real work.

Ilustração editorial de uma estrutura ciano conectando documentos, calendário e código, representando memória e foco de agentes de IA.

CONFIRMED

Meta launched Muse Wonder and Muse Wonder Full on October 12, artificial intelligence models designed for agents that must remember context and keep working when something goes wrong. The significant change is less the promise of better answers than an attempt to reduce two everyday automation costs: explaining everything again and rescuing a task abandoned at its first obstacle. That is the launch proposition, not savings already demonstrated in Brazilian businesses.

The rollout differs by product

Muse Wonder is available through the Meta Model API, an application programming interface, and Muse Code, its software development tool. Wonder Full is available to Muse Code subscribers. Meta says Wonder will also power experiences in its personal agent Muse soon, starting with subscribers. Launching the model therefore does not mean immediate app access for everyone or automatic availability in Brazil. The release expands the family built on Muse Spark and puts Meta in competition for professional agents, not just conversations.

Memory and persistence must help the work

Meta calls its ability to carry important information across conversations and retrieve what matters for the current task Active Memory. Active Focus is its name for persistence: trying other approaches, adapting plans, working through obstacles and asking for clarification when necessary. The company also says the agent should check with users before important actions. At a small agency, a possible application would be keeping campaign constraints intact over several revisions. At a service business, it could track documents and outstanding issues without requiring the case to be reintroduced in every session. These are potential use cases, not tests conducted by MaxAssistant.

The right calculation is per completed task

According to CNBC, future access through Muse will be tied to monthly subscriptions of $20 or $100. The outlet reports Muse Code plans of $50 or $200 for access to Wonder Full. Those are subscription prices, not API rates. Alexandr Wang told the broadcaster the model would be priced meaningfully below leading offerings from OpenAI and Anthropic, but that statement is not enough to establish the cheapest alternative. For Brazilian small and medium businesses, the potential saving lies in reducing rework and human intervention. Integration, calls, retries, review and failures must all be counted before comparing suppliers.

A persistent agent can persist in a mistake

Meta presents frontier results in coding, memory and specialized tasks, but its own tests do not replace independent evaluation in a company's actual workflow. WIRED highlights that benchmarks do not guarantee real performance and that Meta did not publicly detail specific incidents involving these agents during testing. The lab says it strengthened privacy protections and resistance to malicious instructions, with additional safety evaluations planned. None of that removes the need to restrict permissions, log decisions and require approval for spending or sensitive actions. Persistent memory also requires deciding which information may be stored and how outdated information can be corrected.

MaxAssistant's take

This is worth testing in a reversible workflow, such as reviewing documents or building a prototype, comparing correct completion, time and human intervention against the current solution. It is not a reason to migrate production on the strength of a personal superintelligence slogan. The meaningful change is treating continuity and recovery as central model capabilities. If they work beyond the demonstration, the benefit will be less repetitive supervision. If they merely extend bad attempts, the result will be a busier agent, not a more efficient business.

Sources

Meta: Muse Wonder and Muse Wonder Full announcement; Meta AI Research: memory, focus and model evaluations; CNBC: paid access and interview with Alexandr Wang; WIRED: competitive context and safety limitations.