Meta's Muse agent and Meta AI assistant can now carry details across conversations and recover when a task stalls, after the company pushed the upgrade to all users on November 6, 2026. Before this rollout, the two features, Active Memory and Active Focus, were limited to Muse Wonder, the company's top model.
Mark Zuckerberg announced the change on X that day, writing about "bringing Active Memory and Active Focus from Muse Wonder to everyone." Meta's AI Research blog carried the same announcement. The move takes capabilities from a flagship model and makes them part of the default experience in an assistant that competes with ChatGPT and Claude in the US consumer market.
What Meta AI Active Memory and Active Focus do
Active Memory carries details from one conversation into the next and judges whether older information still matters. Active Focus helps an agent recover or find another route when a task runs into an obstacle, according to Datastudios' summary of the rollout.
Meta describes both as upgrades to the models behind Muse and Meta AI, trained with the Muse Wonder frontier model. This is a product update to existing assistants, not a new model family. For a typical user, the change will show up as a reply that recalls something said earlier, or a task that picks up after a failed step instead of stopping. How often that happens is the open question, and Meta has not explained it.
The easy way to overstate this is to assume it changes how good the answers are, and Meta has not shown that in everyday use. What the change does is move memory and recovery from the flagship model into the default product. That narrows the gap between Meta's top model and the assistant most people already use. That is an inference from the rollout, not a claim Meta has made.
Which users get the update, and what is still unclear
Meta says the new models reach everyone using the Muse agent and the Meta AI assistant. The eligibility picture is less tidy than that phrasing suggests. One report says free users get the features, while other accounts say "everyone" without naming plan tiers. Meta has not published a rollout schedule, a regional breakdown or account-tier details.
Muse Wonder itself was not opened up. It launched October 12 and remains reserved for harder tasks by paying subscribers, according to ShiftDelete. The update follows a recent Muse voice rollout with expressive avatars in the US, a sign that Meta is widening the Muse line in stages rather than in one launch.
One report says Muse has passed 10 million setups in North America, according to RuntimeWire. That figure is not independently confirmed, and it does not say how many of those setups are on free plans.
Meta's numbers and what remains unverified
Meta's headline claim is that Muse models are 2.4 times less likely to forget critical information. Zuckerberg's post also cites a 67% improvement in self-correction and task completion. The 67% figure has no disclosed baseline, so readers cannot tell what it is measured against.
The internal evaluation is more specific. AIWeekly's report quotes Meta's internal evaluation as saying the new Muse models "correctly use critical context in 90% of tasks, up from 76% for the previous model." Separate reporting from AI TLDR puts Muse at 47% on Meta's Personal Superintelligence Index, up from 38%.
| Metric | Before | After | Basis |
|---|---|---|---|
| Critical context used correctly | 76% | 90% | Meta internal evaluation |
| Personal Superintelligence Index | 38% | 47% | Meta's index, as reported |
| Forgetting critical information | Not disclosed | 2.4 times less likely | Meta claim |
| Self-correction and task completion | Not disclosed | 67% improvement | Zuckerberg's X post |
No independent benchmark or third-party evaluation has been published. Until outside researchers run comparable tests, these results are the company's own account. They are worth tracking, but they are not verified performance data.
How to check and limit Meta AI memory today
Meta has not announced new user controls for the stronger memory features, so do not assume you can opt out by default. These steps reduce your exposure while the settings remain unclear:
- Open Meta AI or the Muse app and search your settings for "memory," "personalization" or "history." Options may differ by app version and account, so an absent option does not confirm that nothing is stored.
- If a memory option exists, review what it has saved and delete entries you no longer want kept.
- Keep account numbers, passwords, medical details and other sensitive financial or personal information out of any assistant that keeps context across sessions.
- Check important answers against a primary source before acting on them, since Meta has not said how often the memory features trigger.
Developers are not directly affected by this consumer rollout. Meta has not said whether the Active Memory models are available through its API.
What Meta has not explained about privacy and data use
Meta has not reported whether memory retention changes its privacy settings or data-use terms. Those answers matter more to most users than the benchmark numbers, because a memory feature by design keeps more of what you share.
Questions about what chatbots keep and say are already in court. A man is suing OpenAI and Sam Altman, alleging that ChatGPT helped his wife plan a poisoning, according to USA 24's report. That case involves a different company, but it shows how much attention chatbot conversations now draw.
USA 24 has reported separately on Meta's publisher AI terms and the Norway and UK probes, a different thread from this product update. You can read that coverage in our earlier story on Meta's publisher terms.
What to watch next
Three signals matter most: whether Meta publishes plan eligibility and a regional schedule, whether it adds deletion controls for stored memories, and whether the Active Memory models reach its API. Independent testing would also help, since the performance story currently rests on Meta's own numbers. Until those answers arrive, the most practical test is the one every user already has: how the assistant handles your next multi-step task and what it remembers afterward.
