Meta has upgraded the models behind its Muse agent and Meta AI assistant, Mark Zuckerberg posted on X on Friday, bringing two capabilities from its Muse Wonder series, Active Memory and Active Focus, “to everyone”. In Muse, he wrote, the new specialized models are “2.4x less likely to forget critical info and are 67% better at self correcting and completing everyday tasks”.
Meta’s research blog puts the current Muse model at 90% on its Active Memory eval, against 76% for the previous model. Read the other way, the share of tasks that missed critical context fell from 24% to 10%, which is where the 2.4x comes from, by our arithmetic. These are Meta’s own measurements.
Business Pill · AI Agents Explained
A short explainer on how AI agents move from answering to doing, by calling tools. It teaches the general idea only and does not discuss this story.
The key insight: As we read Meta’s blog, the change is a retraining of existing models with Muse Wonder rather than a new flagship: on Meta’s own Active Memory eval, the share of tasks that missed critical context fell from 24% to 10%, by our arithmetic, and the eval itself is not published.
The blog, dated 6 November, says Meta is “introducing specialized versions of our existing models, trained using Muse Wonder, and rolling them out to people who use Muse and Meta AI”. It describes Muse Wonder as “our most capable models to date”. The post does not name the new specialized models or give their size.
Share of tasks in Meta’s Active Memory eval where the Muse model missed critical context, previous vs current model, derived from the 76% and 90% scores in Meta’s 6 November 2026 blog. Meta’s own internal evaluation.
What Was Posted on X
Upgraded the models for Muse and Meta Al, bringing Active Memory and Active Focus from Muse Wonder to everyone. In Muse, these new specialized models are 2.4x less likely to forget critical info and are 67% better at self correcting and completing everyday tasks which means less repeating yourself and fewer dead ends.
Mark Zuckerberg announced the model upgrade for Muse and Meta AI on X on 6 November, writing that the new specialized models are 2.4x less likely to forget critical info and 67% better at self correcting and completing everyday tasks, figures this piece sets against Meta’s blog.
Meta’s Muse agents continue to get new features and integrations. Spaces promises to be a new way to interact with the AI helpers. https://spr.ly/6014BHNYXo
TechCrunch posted on X on 6 November that Meta’s Muse agents “continue to get new features and integrations” and that Spaces “promises to be a new way to interact with the AI helpers”, linking its report on the Spaces launch covered in the update box above.
How Meta Measured It
On Meta’s Personal Superintelligence Index, the current Muse model scores 47% and the previous one 38%, according to the blog’s chart. On the Active Memory eval, the measure is “Tasks Correctly Using Critical Context”: 90% for the current model and 76% for the previous one.
For Active Focus, the blog says that “In Muse, our models show a 67% improvement in completing everyday tasks where unexpected issues arise”. It gives no chart or base rate for that figure. Meta says the models were trained “on a variety of bottlenecks reflective of the real world” and to “recognize what type of information is temporary versus permanent”.
The blog does not give the size of either eval, how the tasks were built, whether outside testers checked them, or any cost or latency figures.
Update, 6 Nov, 19:05 UTC: Meta also launched spaces in Muse. In a newsroom post published at 18:00 UTC, Meta said it is introducing spaces, which it describes as โa custom experience you create inside Muse by describing your idea, like a holiday gift tracker, a daily word game, or a planning board for your next DIY projectโ. Meta says users can connect the apps they use so that spaces โkeep themselves up to date in the backgroundโ, and that โStarting today, spaces are available everywhere Muse is available.โ
The same post ends with the model rollout covered above: โWe’re also rolling out specialized versions of our models to Muse, trained with learnings from Muse Wonderโ. TechCrunch reported the launch at 18:44 UTC. Meta’s post gives no pricing, and says it is exploring ways to let users share spaces with others.
As we read the blog, Meta is pushing capabilities from its top Muse Wonder series down into the models most people use, the same pattern Zuckerberg describes as bringing them “to everyone”.
As we read the figures, the 2.4x is a ratio of miss rates, 24% versus 10%, not a 2.4x higher score; the score itself rose from 76% to 90%.
As we read the coverage, memory and goal persistence are the same capabilities that decide agent outcomes in tests such as Visa’s shopping benchmark, where Muse Wonder bought on 2 of its 8 failed runs.
Meta AI Research blog, 6 November 2026
“Today, weโre introducing specialized versions of our existing models, trained using Muse Wonder, and rolling them out to people who use Muse and Meta AI.”
Three Implications
FOR MUSE USERS Meta says Active Memory and Active Focus now reach everyone using Muse and Meta AI, not only Muse Wonder.
FOR RIVALS Meta reports 47% on its own Personal Superintelligence Index for the current Muse model, against 38% before; no outside benchmark is cited.
FOR READERS The 2.4x and 67% are Meta’s internal figures; the eval sizes and methods are not published.
The framework says: “The surface, not the model, governs autonomy.”
As we read it, Meta’s memory and focus features sit in the Muse surface as much as in the model: Meta says it trained the models to know when to push on goals independently or collaborate live. The framework predates this release and does not discuss it.
What Is Not Established
The 2.4x and 67% figures are Meta’s own, from internal evaluations the blog does not describe in detail. We have not seen independent tests of the new models, and the blog does not say whether the upgrade changes Muse’s cost to serve.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Meta says it has moved two Muse Wonder capabilities, memory and goal persistence, into the models everyone uses in Muse and Meta AI. On its own eval, misses of critical context fell from 24% to 10% of tasks; the claims are Meta’s and the test details are not public.
A note on sourcing. The model upgrade and its metrics come from Meta’s research blog of 6 November and Mark Zuckerberg’s post on X the same day; the 2.4x, 67%, 90%, 76%, 47% and 38% figures are Meta’s own internal measurements. The 24% and 10% miss rates are our arithmetic from Meta’s scores. We did not contact Meta.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
