Meta 为 Muse 和 Meta AI 推出 Active Memory 与 Active Focus…
Current Muse Scores 47% on Meta’s Personal Superintelligence Index
Meta AI Research 于 2026 年 11 月 6 日宣布，通过 Muse Wonder 训练的专用模型已部署到 Muse 和 Meta AI，当前 Muse 在 Meta 个人超级智能指数上得分 47%，上一版为 38%。两项更新为 Active Memory 和 Active Focus：前者让助手保留并在后续对话中调用相关上下文，Meta 称关键信息遗忘减少 2.4 倍，Active Memory Eval 中正确使用关键上下文的任务占比从 76% 升至 90%；后者在任务遇到意外阻碍时维持目标推进，Muse 在出现意外问题的日常任务上完成率提升 67%。
推荐理由
AI 评分 65
原文给出 Muse 与 Meta AI 新增 Active Memory 和 Active Focus 后的具体评测数字，可据此判断助手在记忆与任务连续性上的变化。
主题
正文
Meta AI Research announced on November 6, 2026 that it is rolling out specialized versions of the models behind Muse and Meta AI, trained using Muse Wonder, with the current Muse model scoring 47% on its Personal Superintelligence Index versus 38% for the previous model.
The update brings Active Memory and Active Focus to both products. The models being deployed are specialized versions of Meta’s existing systems, improved through Muse Wonder, rather than a rollout of the Wonder series itself. That approach extends capabilities introduced with Wonder to more people through the assistants’ existing models.
Active Memory Selects Relevant Context
Active Memory lets the assistants retain important information and retrieve it when it becomes relevant to a later conversation. Training covers the difference between temporary and permanent information, reasoning over time, and recognizing when details remain relevant or have become stale. This makes memory a problem of selection and timing: the model must identify which earlier context applies now and which has gone stale.
For the improved Muse models, Meta reports a 2.4x reduction in forgetting information critical to users’ goals. Its Active Memory Eval shows the current model correctly using critical context in 90% of tasks, compared with 76% for the previous model. Those figures describe separate parts of the problem: retaining information needed for a goal and applying that context correctly during a task.
Active Focus Keeps Tasks Moving
Active Focus is intended to keep Muse moving toward a goal when an unexpected challenge interrupts the work. Meta trained the models on varied bottlenecks designed to reflect real-world conditions so they can persist, recover, or pursue alternatives for the user. In Muse, Meta reports a 67% improvement in completing everyday tasks where unexpected issues arise.
The specialized models also address how an agent works with someone while a task is underway. They are designed to multitask, choose between independent progress and live collaboration, and incorporate steering or clarifications without becoming confused or losing the task. That links persistence to responsiveness: an agent can continue pursuing the goal while still changing course when the user intervenes.
In Meta AI, the same improvements are intended to make the assistant a better thinking partner and collaborator. It can divide complex problems into clear steps, ask sharper follow-up questions, and keep pace when a user changes direction mid-conversation.
For users, the rollout’s practical test is continuity: whether useful context survives changing circumstances, work recovers from obstacles, and corrections can be incorporated without losing the goal. Those behaviors connect Meta’s reported results to the everyday collaboration it intends to support in Muse and Meta AI.