What should an AI remember from an old conversation? Muse puts memory to the test

The new Muse and Meta AI models are trained to retrieve information that remains relevant and keep working when unexpected obstacles arise.

Picking up a project several days after discussing it means retrieving the right instructions and accounting for what has changed in the meantime. This is one of the capabilities Meta has been working on in the new models it is rolling out to Muse and Meta AI. The company used its Muse Wonder model family to train specialized versions of the models already powering its two assistants.

When Muse Wonder was introduced in October, the plan was to make it available in Muse first to subscribers. The November 6 rollout expands access to two capabilities developed from this model family. Active Memory helps the assistant determine which information is worth retaining. Active Focus trains it to keep working when a step fails or the user's request changes. A Preference That Lasts, an Instruction That Expires

When planning a new trip, a travel preference mentioned last month may still be useful. The dates of that previous trip, however, are unlikely to apply. Active Memory is designed to make that distinction. To evaluate this capability, the Muse Wonder model card describes a test combining lengthy conversations, exchanges unrelated to the current request, and information corrected along the way.

In Muse's published evaluation, the new models correctly use essential information in 90% of tasks, compared with 76% previously. The share of tasks where this context is mishandled therefore drops from 24% to 10%. That ratio is the basis for Meta's claim that the models are 2.4 times less likely to forget critical information. Keeping Going When Something Goes Wrong

A changing constraint can make some of the work already completed irrelevant. In Muse, Active Focus is designed to help the assistant resume a task using the updated instructions, find an alternative approach, or ask for clarification. Meta reports a 67% relative improvement in completing everyday tasks involving unexpected obstacles. For Meta AI, the company primarily describes improvements during conversations, such as asking more precise follow-up questions or adjusting how a problem is broken down as the request evolves.

Muse's score on the Personal Superintelligence Index also rises from 38% to 47%. This internal benchmark combines seven evaluations, including web research exercises, professional tasks, and FocusBench. The latter measures whether an agent can overcome multiple obstacles and adapt to changes throughout a single task.