Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- Meta is bringing specialized versions trained using Muse Wonder into the existing Muse and Meta AI products; the announcement does not promise simultaneous availability to every user, country, or account.
- Active Memory distinguishes temporary from lasting information and judges whether earlier context is still relevant or stale. Meta reports 2.4x less forgetting of critical goal information in improved Muse models.
- Muse's Active Memory Eval is 90% versus 76% for the previous model, while its Personal Superintelligence Index is 47% versus 38%. These are different measures and cannot be used to derive the 2.4x claim.
- The 67% Active Focus result concerns Muse's improvement on everyday tasks with unexpected issues, not a general self-correction rate or Meta AI success rate. Meta has not announced countries, plans, prices, or account-level rollout dates.
What is rolling out
On November 6, 2026, Meta announced that specialized model versions trained using Muse Wonder are rolling out to people who use Muse and Meta AI. These versions improve the models behind the existing products. Meta says they are rolling out; that wording does not mean every account has already received the update.
The announcement concerns the models behind existing products. It does not publish a new API model ID, endpoint, SDK migration, or downloadable weights. It also leaves countries, plans, prices, and account-level arrival dates unspecified; those access and usage conditions still need confirmation.
Active Memory judges what information still matters
Active Memory helps Muse and Meta AI remember important information and recognize when earlier details become relevant in a later conversation. Meta says the models learn to distinguish temporary from lasting information, reason across time, and identify what remains relevant versus what has become stale, so they can build on prior conversations more effectively.
Meta reports a “2.4x reduction” in forgetting critical information needed to help users reach their goals in improved Muse models. The article does not provide a full methodology, so the figure should not be generalized into a performance guarantee for every task.
Two other Muse results are 90% versus 76% on the Active Memory Eval, measuring tasks that correctly use critical context, and 47% versus 38% on the Personal Superintelligence Index. These are different evaluation results; the 90%/76% figures do not derive the 2.4x claim.
Active Focus is aimed at tasks that hit obstacles
Active Focus is designed to keep working toward a goal when unexpected challenges arise. Meta says the models were trained on a variety of bottlenecks intended to reflect real situations, so they can try harder, recover, or find alternatives on the user's behalf.
Meta reports a 67% improvement for Muse models in completing everyday tasks where unexpected issues arise. This is a company-reported result for a specific Muse task category. It is not a 67 percentage-point increase in a universal completion rate, a general self-correction rate, or a separate Meta AI evaluation score.
Muse and Meta AI have different product descriptions
Meta also says the improved Muse models are better multitaskers: they can decide when to advance a goal independently or collaborate with you live, and adapt more effectively to directions and clarifications during a task. The announcement gives no separate numerical score for these behaviors.
For Meta AI, Meta describes a better thinking partner and collaborator that can break complex problems into clear steps, ask sharper follow-up questions, and keep up when you change direction. That is a qualitative product description; the 90%/76%, 47%/38%, 2.4x, and 67% results above are reported for Muse models.
Observe rollout in your own account carefully
The rollout announcement does not list countries, plan eligibility, prices, account-level arrival dates, or availability on glasses. Availability for your product, region, and account requires confirmation from the capabilities actually shown and any further official guidance.
If the relevant capability appears in your product, observe it with a hypothetical task containing no real personal or business information. For example, ask for a fictional project summary limited to three points, with no facts absent from the prompt, then request a follow-up.
Define success beforehand as keeping that constraint without inventing details, and record the product and date. This small example can help you observe your own task; it does not reproduce Meta's evaluation, and no model call was performed for this article.
Sources
Frequently Asked Questions
Does Active Memory permanently save every conversation?
The announcement does not specify retention periods, user controls, or that every conversation is kept. It only describes the model's ability to distinguish temporary from lasting information and relevant from stale context; it is not a promise to remember everything permanently.
Did Meta publish sample sizes or independent replications for these evaluations?
The official article does not provide sample sizes, a full testing protocol, or independent replications. Treat the 2.4x, 90%/76%, 47%/38%, and 67% figures as results reported by Meta, not as independent test results.