Promise Ledger: Wonder Memory Claims
A standing box for this rollout, placed here so the verdict sits next to the headline numbers. Last updated November 8, 2026, from Meta AI Research's November 6 post, read on November 8, and the November 6 to 8 coverage and checks listed in Sources.
- Live now: specialized versions of Meta's existing models, trained using Muse Wonder, rolling out to people who use Muse and Meta AI (Meta AI Research, November 6, 2026). This is not Muse Wonder itself moving to the free tier. Wonder itself stays where Meta put it on October 12, for harder work and paying access. What reached everyone is a set of smaller models taught by it.
- Features that came with them: Active Memory and Active Focus, both first named with Wonder in October.
- Figure 1, forgetting: 2.4 times less forgetting of critical information in Muse. Source: Meta's post, in its own words. Verdict: Meta-claimed. Method published: no. Independent test found: no, checked November 8.
- Figure 2, getting unstuck: 67 percent improvement in completing everyday tasks in Muse where unexpected issues arise. Source: Meta's post, in its own words. Verdict: Meta-claimed. Starting rate published: no. Method published: no. Independent test found: no, checked November 8.
- Figure 3, Active Memory evaluation: tasks correctly using critical context, 90 percent for the current Muse model against 76 percent for the previous model, in Meta's chart figures as reported by readers of the November 6 post (Unite.AI, datastudios, and sanctity.ai, all November 6 to 7). Verdict: Meta-claimed. Method published: no. Independent test found: no, checked November 8.
- Figure 4, Personal Superintelligence Index: 47 percent for the current Muse model against 38 percent for the previous model, same chart sourcing as Figure 3. Verdict: Meta-claimed, on an index Meta itself runs. What the index measures, in tasks or marking: not stated in the post text read on November 8.
- Separate figures for Meta AI: none published. All four figures above describe Muse, even though the same specialized models also roll out to Meta AI.
- Next recheck, a promise with a date: this site will recheck for a published method and for a reproducible independent test of these specialized models on November 16, 2026, and will publish the result on that date even if both are still missing. It will also recheck sooner if Meta publishes a method, a sample size, or a model name for the specialized models.
The short version: trained models rolled out, Wonder itself did not
Yes, it rolled out. That is the answer readers coming from the October coverage need first, because for three weeks the status of Wonder in the app was a watch item on this site: announced for developers on October 12, subscribers first for the app, no date. On November 6 Meta closed a different version of that question. The app did not get Wonder. It got models trained using Wonder.
Meta's post, "Bringing Specialized Models to Muse and Meta AI," is short. It says specialized versions of its existing models, trained using Muse Wonder, are rolling out to people who use Muse and Meta AI. It brings two named behaviors with them, Active Memory and Active Focus. It gives two figures in words, the 2.4 times and the 67 percent. Its charts give the other two, the 90 against 76 and the 47 against 38, as reported by the readers cited in the ledger above.
What the post does not give is the part that would let anyone else check the work: no description of the tests, no number of tasks, no number of users or runs, and no name, size, or version number for the specialized models themselves. Its charts label them only Current Muse Model and Previous Muse Model, per sanctity.ai's reading of the page. So this post prints all four figures, each one labeled, and leaves the method line blank with a date on it. That blank is not a detail. On a memory claim, the test is the story.
What Active Memory is supposed to do
Meta's description is specific enough to be testable, which is to its credit. Active Memory, in its words, enables Muse and Meta AI to remember what matters and recognize when earlier information becomes relevant, so a user can build on past conversations instead of starting over.
Under that sentence sit three trained judgments, all named in the post. The models were trained to recognize what type of information is temporary versus permanent. They were trained to reason across time. And they were trained to distinguish between what is still relevant and what has become stale.
Read those three slowly, because they are harder than plain recall. A model that remembers everything you said and uses all of it at once would be worse, not better: it would quote the address you moved out of, the budget you replaced, and the plan you cancelled, all with the same confidence as the current ones. The claim here is not storage. It is selection, over time, with updates. That is also why a simple quiz, asking Muse to repeat a fact you gave it an hour ago, would not test what Meta is claiming. The test has to plant a fact, bury it under later conversation, change it, and then ask for work that depends on using the new version and not the old one.
What Active Focus is supposed to do
Active Focus is the persistence half. Meta: it enables Muse to continue working toward a goal, even when faced with unexpected challenges. The models were trained on a variety of bottlenecks reflective of the real world, so they can try harder, recover, or find alternatives on your behalf.
The post adds two working details past that headline. The specialized models learned to be multi-taskers that know when to push on your goals independently and when to collaborate live with you. And they can better adapt to steers and clarifications in the middle of running tasks without getting confused. That second detail matters in daily use. Most long tasks do not fail at the first step. They fail at the correction, when you change a date or a budget halfway through and the agent keeps working from the first version.
For Meta AI, Meta describes the same change in plainer terms: a better thinking partner that can break complex problems into clear steps, ask sharper follow-up questions, and keep up when you change direction mid-conversation. No separate score is published for any of that. The 67 percent figure is a Muse figure.
The two abilities pull against each other in real work, and that tension is where a reader test will show something a benchmark may not. An agent trained to keep going will sometimes keep going past the point where it should have stopped and asked. An agent trained to remember will sometimes use a fact it should have let go stale. Watch for those two failures in your own tasks this week: a confident push through a blocker that needed your call, and an old detail surfacing as if it were still true. One instance proves nothing. A pattern, dated and written down, is data this site can use.
The four figures, checked one by one
Forgetting, 2.4 times less. This one is in the text of Meta's post, read firsthand on November 8: the improved models in Muse show a 2.4x reduction in forgetting critical information necessary for helping you achieve your goals. No baseline count sits next to it in that sentence. It does not say out of how many tasks, or over how long a conversation.
Getting unstuck, 67 percent better. Also in the text, firsthand: in Muse, the models show a 67 percent improvement in completing everyday tasks where unexpected issues arise. Without a starting rate, a reader cannot convert that into tasks finished. A 67 percent improvement on a low starting rate and the same improvement on a high one are very different products, and the post does not say which one shipped.
Active Memory, 90 against 76. Readers of the post's charts report that the current model correctly used critical context in 90 percent of tasks, against 76 percent for the previous model (Unite.AI, datastudios, sanctity.ai, November 6 to 7). There is an arithmetic point worth making here, and datastudios made it on November 6: if correct use rises from 76 to 90, then incorrect use falls from 24 in 100 to 10 in 100, and 24 divided by 10 is 2.4. That matches the forgetting figure exactly. Meta's post, in the text read here, does not say whether the two figures come from the same measurement, so this post notes the match and does not merge them into one claim.
Personal Superintelligence Index, 47 against 38. Same chart sourcing. This is Meta's own index, with Meta's own tasks and marking, and no outside referee named in any source checked on November 8. A nine point rise on a private scale tells a reader that Meta's newer model did better on Meta's test than Meta's older model did. It does not place Muse against any other assistant, and it should not be quoted as if it did.
Independent tests of Muse Wonder itself do exist. This site covered them in October, and a SemiAnalysis benchmark dated November 2 put Wonder Full first on cyber tasks and behind rival frontier models overall. Those tests are about Wonder, the frontier model you can reach through the API and Muse Code. They are not tests of the unnamed specialized models that reached the Muse app on November 6, and they do not verify a single figure in the ledger above. Keeping those two sets of models separate is the whole job of this post.
What did not change on November 6
Price did not. US pricing on Meta's help pages, checked in the November 7 verification for the Spaces post and unchanged in the November 8 checks, is still free with a usage limit, Power at $20 a month, and Maximum at $100 a month. The specialized models are not a new tier, and Meta names no tier as required for them. Its wording is people who use Muse and Meta AI, and Mark Zuckerberg's November 6 post, as quoted by blockchain.news and Mundo IA, frames it as bringing Active Memory and Active Focus to everyone.
The country list did not change either. Muse is in five countries on this site's dated record: the US, Canada, Japan, Singapore, and Mexico, with Mexico added on November 5. The November 6 specialized models post, in the text read on November 8, names no countries at all. A rollout sentence with no country list does not prove the new models reached all five countries on the same day, and no source checked on November 8 states a country by country rollout order. That line stays open in the ledger style this site uses for Voice, where US only was stated in words.
And Wonder's own access did not change. This post is not a pricing update and does not re-report the Wonder API figures covered in the October Wonder post, because nothing in the November 6 post changes them. Readers who need those numbers should use that post, dated, rather than a November headline that would make old figures look new.
The test that would settle it
Meta could settle this with documents: publish the tasks, the sample size, and the marking for the Active Memory evaluation and the Personal Superintelligence Index, and give a starting rate for the 67 percent figure. Until it does, the claim can still be tested from the outside, in a smaller way, if the test is fixed in advance. Here is the one this site will use, stated now so the November 16 recheck and any later run cannot quietly move the goalposts.
Take a set of long, multi-step tasks of the kind Muse is sold for: a trip plan, a research brief, a multi-part booking or purchase plan. In each task, plant three kinds of fact. One is permanent and must be used at the end. One changes halfway through, and only the new version is correct. One is temporary and should be ignored by the end.
Partway through each task, insert one unexpected blocker: a closed option, a price change, a missing file. Then score four things only. Was the permanent fact used correctly? Was the changed fact used in its new version, not its old one? Did the temporary fact stay out? And did the task finish after the blocker or stall, with a note on whether it asked for help at a point where asking was the right call?
Run that set, write down the scores with dates, and publish them next to the scores from any later model change. That will not reproduce Meta's numbers, and it should not claim to. It will answer the reader question underneath them: on a long task you actually run, does Muse now forget less, use the current version, and recover when stuck. The November 16 recheck covers Meta's method and any outside test. A reader run of the test above, if readers send results in, gets reported separately, with its task count stated, however small it is.
Quick answers
Are Muse Wonder-trained models live in Muse?
Yes, on Meta's record. Meta AI Research, November 6, 2026, says specialized versions of its existing models, trained using Muse Wonder, are rolling out to people who use Muse and Meta AI, bringing Active Memory and Active Focus with them.
Is this Muse Wonder itself becoming free?
No. Meta's post is about specialized versions of its existing models, trained using Wonder. It does not move Wonder or Wonder Full to the free tier, name the specialized models, or give their sizes, per the post text read on November 8 and the November 6 to 7 coverage.
What is Active Memory?
Meta's name for remembering what matters and recognizing when earlier information becomes relevant. The models were trained to tell temporary information from permanent, to reason across time, and to tell what is still relevant from what has become stale (Meta, November 6, 2026).
What is Active Focus?
Meta's name for continuing toward a goal when unexpected challenges arise. The models were trained on real world style bottlenecks so they can try harder, recover, or find alternatives, and to adapt to steers and clarifications mid-task without getting confused (Meta, November 6, 2026).
Does Muse really forget 2.4 times less now?
Meta says so, in those words, for Muse. No evaluation method, sample size, or baseline count was published with the figure in the post text read on November 8, and no independent test of the specialized models was found in checks through November 8. Read it as Meta-claimed, not as measured by anyone else.
What are the 90 percent and 47 percent figures?
Chart figures from Meta's November 6 post, as reported by Unite.AI, datastudios, and sanctity.ai: correct use of critical context at 90 percent for the current Muse model against 76 percent for the previous model, and a Personal Superintelligence Index score of 47 percent against 38 percent. Both are Meta's own evaluations. The index is Meta's own scale.
Are there separate figures for Meta AI?
No. All four figures describe Muse. Meta describes the Meta AI effect in words only: clearer steps, sharper follow-up questions, and keeping up when you change direction mid-conversation.
Did prices or countries change with this rollout?
No change was found. US pricing is still free, Power at $20 a month, and Maximum at $100 a month. Muse is in the US, Canada, Japan, Singapore, and Mexico. The November 6 post names no countries, so a same day rollout in all five is not established.
Frequently Asked Questions
Q: How is this different from the Wonder post this site ran in October?
A: That post, idea 120 on this site's list, covered Muse Wonder and Wonder Full as developer models: API access, Muse Code, and their pricing. This post covers a different event, the November 6 rollout of unnamed specialized models, trained using Wonder, inside the Muse and Meta AI apps. Different models, different audience, different figures.
Q: Why does this site keep saying Meta-claimed?
A: Because every figure in the ledger was produced by Meta, on tests Meta designed, with no method published that would let anyone rerun them. Meta-claimed is not a dismissal. The October Wonder figures carried the same label until outside tests arrived, and some of those outside tests partly supported Meta and partly did not. These figures get the same treatment: printed, dated, and labeled until someone who is not Meta can check them.
Q: Did explainx or anyone else find a method Meta published elsewhere?
A: No, in the record this site has. explainx, in its November 6 coverage, noted that Meta published no evaluation method for these figures. This site's November 7 and 8 sweeps, through 19:06 CST on November 8, plus a fresh November 8 check of the primary post text, found the same gap: no method, no sample size, and no starting rate for the 67 percent figure. That is a dated absence, checked November 8, not a claim that Meta will never publish one. The November 16 recheck exists for exactly that reason.
Q: Can I tell whether my account has the new models?
A: Not from anything published. Meta has given the specialized models no public names or version numbers, and no source checked on November 8 describes a label, a settings line, or a version screen in the app that would identify them. If your Muse seems to remember better this week, that is consistent with the rollout. It is not proof your account has it, and this post will not pretend a feeling is a version check.
Q: Does Active Memory mean Muse keeps everything I tell it?
A: Meta's description points the other way: the training is about telling temporary information from permanent, and what is still relevant from what has become stale. What that means for retention in practice, what is kept, for how long, and what you can delete, is not stated in the November 6 post. Those are privacy terms questions, and they stay unanswered here rather than being filled in from a launch post about model quality.
Q: What should I watch for in my own use this week?
A: Two failures, both named in the field note above: Muse using a fact you later changed, in its old version, and Muse pushing through a blocker that needed your decision instead of asking. Write down the date and the task if either happens. That note, not a general impression that it feels smarter, is what the November 16 follow-up can actually use.
Q: What is the reader question for this post?
A: On a long task you actually run, does Muse now forget less and recover when it gets stuck? Not whether the announcement sounded convincing, and not a one line chat that went well. A long task, with a fact that changed in the middle and something that went wrong. Answers with dates and task types will shape the follow-up more than any percentage in the ledger, including Meta's.
Sources
- Meta AI Research, "Bringing Specialized Models to Muse and Meta AI," research.meta.ai, November 6, 2026. Read firsthand on November 8, 2026. The rollout itself: specialized versions of existing models, trained using Muse Wonder, rolling out to people who use Muse and Meta AI. Active Memory defined, with the temporary versus permanent, reason across time, and still relevant versus stale training, and the 2.4x reduction in forgetting critical information in Muse, in the post's own words. Active Focus defined, with the real world bottlenecks training, the try harder, recover, or find alternatives wording, the multi-tasking and mid-task steer details, and the 67 percent improvement in completing everyday tasks in Muse where unexpected issues arise, in the post's own words. The Meta AI description: clearer steps, sharper follow-up questions, keeping up with a change of direction. The post text read on November 8 contains no evaluation method, sample size, model name, model size, country list, or starting rate for the 67 percent figure.
- Chart figures from the same Meta post, as reported by readers of its charts: Unite.AI, "Current Muse Scores 47% on Meta's Personal Superintelligence Index," November 6, 2026 (47 percent against 38 percent on the Personal Superintelligence Index, 90 percent against 76 percent on correct use of critical context, specialized versions rather than Wonder itself, Wonder first announced October 12); datastudios, "Meta AI and Muse get Active Memory and Active Focus for all users," November 6, 2026 (the same two chart comparisons in a table, the note that 24 misses in 100 falling to 10 gives the 2.4 figure, and the limits: no methodology, no task counts, no third-party evaluation, no separate Meta AI scores); sanctity.ai, "Meta gave every Muse and Meta AI user models that remember and keep pushing," November 6, 2026 (the four figures together, charts labeled Current and Previous Muse Model only, two figures with a named previous model baseline and two without, and no identity, size, outside evaluation, or Meta AI figures established). The chart numbers are cited at this weight: consistent across three readers of the primary, not extracted from the chart rendering in this site's own November 8 read of the post text.
- blockchain.news, November 6, 2026, and AIBriefs, November 6, 2026. The first day reports of the rollout, with blockchain.news quoting Mark Zuckerberg's November 6 post: bringing Active Memory and Active Focus from Muse Wonder to everyone, 2.4 times less likely to forget critical info, 67 percent better at self-correcting and completing everyday tasks. Quoted here as a Meta claim carried by those outlets, matching the primary post read on November 8.
- explainx, November 6, 2026. Names Active Memory and Active Focus in the rollout and notes that Meta published no evaluation method for the figures. Cited for the no method point, which the November 8 firsthand read of the primary post text confirms.
- Mundo IA, November 6, 2026, and Fresh Synapse, November 7, 2026. Independent write-ups used as cross-checks only: both give the four figures as Meta's own, both note no independent verification, no sample sizes, and no starting point for the 67 percent figure, and Fresh Synapse gives the primary post URL as published by Meta.
- Editorial checks, November 7 and 8, 2026: the November 7 verification for the Spaces post checked Meta's help pages and found US pricing unchanged at free, Power at $20 a month, and Maximum at $100 a month, and the five country list with Mexico added November 5. Fresh searches on November 8 for an independent test or evaluation of the November 6 specialized models found tests of Muse Wonder itself, including the November 2 SemiAnalysis GDPx results covered in October coverage, and no test of the specialized in-app models. No method, model card, or sample size for the specialized models was found in those checks. The day's later sweeps, through 19:06 CST on November 8, changed none of this: no new Muse launch, pricing change, outage, or country dated November 7 evening or November 8, and the Wonder figures still stood as Meta-claimed with no independent test.
Comments