What changed
On November 6, Meta said it was rolling out updated models for Muse and Meta AI. It used Muse Wonder, its higher-end model series, to train these versions. One change, which Meta calls Active Memory, is meant to help the assistants tell a lasting preference from a detail that mattered only once. [1]
What it can do during a task
Meta calls the other change Active Focus. In Muse, Meta says it helps the assistant keep working when the original plan hits a snag. For Meta AI, the examples are breaking a problem into steps, asking follow-up questions, and adapting when you change direction. Meta describes these behaviors and reports its own Muse results. It gives no guarantee about what an individual chat will remember. [1]
Our take
Try it with a fictional meal plan. Say that you always want vegetarian dinners and that this week you also need meals you can make without an oven. Later, ask for next week's plan and say the oven is fixed. See whether the vegetarian preference carries over and the oven restriction drops away. An answer that gets both details right is more useful than one that simply remembers more.