Meta brings Active Memory and Active Focus to Muse and Meta AI
What changes for people who use Muse and Meta AI? Since November 6, Meta has been rolling out specialised versions of its existing models, trained with the larger Muse Wonder. They bring two features: Active Memory, which holds on to important information better, and Active Focus, which finishes a task when something unexpected goes wrong. According to Meta, Muse now forgets critical information 2.4 times less often. Meta names no countries or timeline; the numbers are Meta's own.
Muse is Meta's personal agent, Meta AI its chat assistant. Both get the new models. Why memory is so hard for such assistants is explained in memory in LLM apps.
What the two features do
- Active Memory: the models are trained to classify information as temporary or permanent, reason across time and tell current from stale information. Meta describes the goal as: "remember what matters and recognize when earlier information becomes relevant".
- Active Focus: Muse keeps working toward a goal when something gets in the way, by retrying, recovering or taking another route. "Active Focus enables Muse to continue working toward a goal, even when faced with unexpected challenges."
Meta's numbers
| Measure | Previous Muse model | New Muse model |
|---|---|---|
| Personal Superintelligence Index | 38% | 47% |
| Active Memory Eval (tasks correctly using critical context) | 76% | 90% |
Meta also reports 2.4 times less forgetting of critical information in Muse and a 67% improvement in completing everyday tasks where something unexpected happens. These are relative figures without a baseline. Meta does not explain how the Personal Superintelligence Index is calculated or which test sets were used, and the results have not been independently verified.
Meta further says Muse knows better when to keep working on its own and when to check with the user, and that Meta AI breaks complex questions into steps and asks sharper follow-up questions.
What this means
Meta is betting on the same question Anthropic wrote about this week: how an agent decides what to remember. See also our piece on agent memory and 'dreaming' at Anthropic. If you want to test such claims yourself, see evals for agents. The signal appeared in the briefing of November 7 and the Radar of November 7; more on agents in our overview agentic AI.
What we did not check
Meta names no regions, timeline, prices or plans for the rollout, nor which model versions are involved. We could not establish whether the features are available in the Netherlands or the EU. We could not open Meta's post on X directly; we used Meta's blog post.
Checked on November 8, 2026.
This article was written with AI (Claude, Anthropic) and checked against the sources below.
Sources
- Meta AI, Bringing Specialized Models to Muse and Meta AI (November 6, 2026): research.meta.ai
- Meta AI, Muse Wonder: A Frontier Model For Personal & Professional AI Agents: research.meta.ai
- Announcement on X by @AIatMeta: x.com/AIatMeta
