News reporting
Meta brings Muse Wonder to developers, with memory at the center
Wonder reaches the API and Muse Code; Full requires a Power or Maximum subscription. The broader Muse rollout is still ahead.
By Agent Report Editorial
Editorial review · · 3 min read
Source-based reportingWritten and source-checked with Codex on October 13, 2026; copy reviewed with Grok. This report examines Meta's announcement and research materials; no hands-on model or memory tests were performed.
Meta has announced Muse Wonder and Muse Wonder Full, putting a new model family into developers' hands before its wider rollout inside the Muse personal agent. The October 12 release gives builders a concrete access distinction: Wonder is available through Meta Model API and Muse Code, while Full is available only on the Muse Code Power Usage and Maximum Usage plans.
Meta says Wonder will reach more Muse experiences later, starting with subscribers. A developer release therefore does not establish that a person opening Muse today is using the new model.
Remembering the task, then staying with it
The company's research description puts two capabilities at the center of the release. Active Memory retrieves relevant details across conversations. Active Focus describes continued work through obstacles, including trying a different approach and asking for clarification. These are Meta's descriptions of trained behavior.
The research post names FocusBench and SWE-Marathon as evaluations of focus and sustained software work. It also links a model card. We have not reproduced its results or established how those evaluations predict a reader's own tasks.
For a developer, the useful distinction is between remembering an instruction and applying it at the right time. Consider a hypothetical assistant told to keep a project within a fixed budget. Recalling the budget in chat is one check. Stopping before an expensive action is another. A model can be useful at the first while still needing application controls for the second.
Access is clearer than the memory contract
Meta's model documentation narrows the access terms: Wonder is in preview, and Wonder Full requires a Power Usage or Maximum Usage Muse Code subscription. Setting API reasoning effort to its maximum does not select Full.
The same documentation separates data-use modes. API access to Wonder uses Standard mode, which Meta says excludes the submitted content from model training. Muse Code subscribers start in Contributor mode by default and can turn it off. That mode permits training on prompts, code, files, context and responses. A team evaluating memory should inspect that setting before loading a project; recall across conversations and permission to train on content are different questions.
The next memory questions are operational: what information is retained, how a correction replaces an old preference, and what a deletion request removes. A memorable demonstration cannot settle all three. Agent Report has not inspected a live memory record or tested its removal.
Our assessment: start a comparison with one interrupted task and one deliberately changed instruction. Preserve the final artifact and check whether the agent obeyed the latest constraint. That would turn Meta's broad persistence claim into a result a reader can inspect.