Day 7 receipt
Meta Muse specialized models (Active Memory and Active Focus)
Claims on the record
We check on 20 Nov 2026
- They promised
- New Muse and Meta AI models that forget 2.4x less and finish 67% more tasks when things go wrong.
The claims we’ll check
- “Today, we’re introducing specialized versions of our existing models, trained using Muse Wonder, and rolling them out to people who use Muse and Meta AI.”
- “our improved models in Muse show a 2.4x reduction in forgetting critical information necessary for helping you achieve your goals”
- “In Muse, our models show a 67% improvement in completing everyday tasks where unexpected issues arise.”
- “It can break complex problems into clear steps, ask sharper follow-up questions, and keep up when you change direction mid-conversation.”
Quoted word for word from Bringing Specialized Models to Muse and Meta AI. Rolled out, but every number is Meta's own internal test with no method published, and nobody outside has measured it.
Full evidence 4 claims · 2 user reports · 7 sources
Meta began rolling out specialized versions of the models behind its Muse agent app and Meta AI, trained using Muse Wonder, with Active Memory to recall relevant past context and Active Focus to keep going when a task hits a snag.
Launch post: Bringing Specialized Models to Muse and Meta AI
Why this stamp
Rolled out, but every number is Meta's own internal test with no method published, and nobody outside has measured it.
Meta's research post, its AI account on X and six outlets describe the same update: models behind Muse and Meta AI, trained using the larger Muse Wonder, carrying two behaviours first built into Wonder, Active Memory and Active Focus. Every number comes from Meta's own tests and compares the new Muse model only with Meta's previous one: a 47% against 38% score on its Personal Superintelligence Index, 90% against 76% on its memory evaluation, a 2.4x cut in forgetting critical information and a 67% improvement on everyday tasks that hit unexpected problems. Two outlets note Meta gives no baseline for the 67% figure, and one outlet notes there is no model card or independent benchmark. Meta also gives no figures for Meta AI and has not named the models or said how they differ from Wonder in size. The only first-hand report we found says the iPhone Meta AI app still labels its free model with an older name and shows no new memory features, which fits a staged rollout but means we cannot confirm who has the update. With almost no hands-on reports yet, we re-stamp on day 7 (13 November) and day 14. Re-checked 13 Nov: No outside measurement yet; we check again on 20 Nov.
Launch dayWhat they claimed, quoted word for word · 4
“Today, we’re introducing specialized versions of our existing models, trained using Muse Wonder, and rolling them out to people who use Muse and Meta AI.”
Outlets agree Meta is rolling these out to everyone, but Meta has not named the models or given a rollout timetable. One early user says the iPhone Meta AI app still shows an older model name and no new memory behaviour, so we cannot yet confirm who has the new models.
“our improved models in Muse show a 2.4x reduction in forgetting critical information necessary for helping you achieve your goals”
This is Meta's own evaluation, compared only with its previous Muse model. Two outlets flag that Meta gives no outside comparison, one adds that no model card or independent benchmark has been published, and nobody outside Meta has tested the memory claim.
“In Muse, our models show a 67% improvement in completing everyday tasks where unexpected issues arise.”
Meta gives no starting figure for this improvement, as two outlets point out, and no outside test has reproduced it.
“It can break complex problems into clear steps, ask sharper follow-up questions, and keep up when you change direction mid-conversation.”
Meta gives no numbers for Meta AI at all, and we found no hands-on test of this behaviour yet.
Day 7What real users report · 2
A user checking the iPhone Meta AI app hours after the announcement says it still labels the free model with an older Spark name, shows no new memory features and does not include the Muse agent, so they see no difference from what Meta announced.
137 views as of 7 Nov · Open the source
A user replying under the CEO's announcement post says Muse could not open the link to Meta's own post when they tried it. This does not test the memory claims, but it is an early sign that link handling is rough.
16 likes · 3.3K views as of 7 Nov · Open the source
OthersWhat outlets and testers say · 7
Reports the same-day Active Memory and Active Focus announcement as a short aside in its story on the new spaces feature, adding no independent test.
“Meta’s Muse AI can create custom ‘spaces’ for the version of the app you need right now” · Open the source
Notes Meta has not named the models or detailed its training method, calls the figures Meta's internal evaluation, and points out Meta shared no baseline for the 67% improvement.
“Meta’s best AI model now trains the one most of its users get” · Open the source
Restates Meta's figures and stresses these are specialized versions of existing models, not Wonder itself; says the practical test is whether context survives changing circumstances.
“Current Muse Scores 47% on Meta’s Personal Superintelligence Index” · Open the source
Stresses that every figure is Meta's own, covers Muse only and compares with Meta's previous model, not rivals; Meta gives no figures for Meta AI and does not name the models.
“Meta says updated Muse and Meta AI models, trained using its Wonder AI, remember more” · Open the source
Relays the announcement and notes Meta gives no rollout timetable and no separate numeric result for Meta AI.
“Meta Brings Active Memory and Active Focus to Muse and Meta AI” · Open the source
Says the numbers are Meta's own evaluations rather than independent audits, with no model card or outside benchmark published, so the verdict waits for real use.
“Meta’s Muse Now Forgets Less and Finishes More” · Open the source
Meta's four figures are its own, with no evaluation method, sample size or independent test published through 8 Nov.
“Muse Wonder-Trained Models Are Live: 2.4x Less Forgetting, With No Method Published” · Open the source