Meta Muse gets Active Memory: 90% on Meta's memory eval, up from 76%
Meta stopped selling its frontier model to most people and started using it as a teacher instead. On Friday the results of that swap shipped to everyone with the app.
Meta began rolling out new models to every person using Muse and Meta AI on Friday — not Muse Wonder, its frontier model, but versions of its existing consumer models trained using Wonder, the company said in its announcement. Paying subscribers still get Wonder itself for harder tasks. What the standard tiers get instead are models carrying Wonder's two headline behaviors: Active Memory, which carries what a user said weeks ago into the conversation happening now, and Active Focus, which keeps an agent working when the task hits a wall instead of giving up on it.
The numbers Meta published are all internal, and they are large. On the company's Personal Superintelligence Index, the current Muse model scores 47% against 38% for the model it replaces. On Meta's Active Memory eval — tasks where the model has to correctly use critical context — the score moved from 76% to 90%. Meta says its Muse models are now 2.4 times less likely to forget information needed to finish a goal, and show a 67% improvement at completing everyday tasks when something unexpected goes wrong.
What those two features actually do is narrower than the framing suggests, and more useful. Active Memory is not a bigger context window; it is a sorting problem. Meta says it trained the models to tell temporary information from permanent information, to reason across time, and to work out which earlier details are still true and which have gone stale. A user who mentions a flight next Tuesday and a peanut allergy two years ago should get different treatment from a model that remembers both equally. Active Focus is the other half: Meta trained on what it calls "a variety of bottlenecks reflective of the real world," so a model can retry, route around a broken site, or find another path when the booking page fails halfway through.
Meta is also claiming behaviors that never show up in a benchmark table. The models are meant to know when to push on a task alone and when to stop and ask, and to absorb a mid-task correction without losing track of everything already done. Anyone who has watched an agent confidently complete the wrong task will recognise why that is the hard part.
The technical claim worth caring about is what Meta says it did not do. It did not shrink Wonder into a general-purpose small model. It trained the new models on the behaviors that matter specifically in a consumer app — which is a different target from "a smaller model that knows slightly less." The New Stack's write-up notes that Meta did not detail the method, and that is a fair criticism: distillation is standard practice at every lab, and "trained using Muse Wonder" is a description of the teacher, not the recipe. But the selection of behaviors is the interesting part, because it tells you what Meta thinks a personal agent is for. Not answering trivia. Remembering your life and finishing your errands.
The economics explain the whole release. Wonder costs $4 per million input tokens and $20 per million output on the Meta Model API, per The New Stack. Muse has a free tier, and Meta AI runs inside WhatsApp, Instagram, Facebook and Messenger — products with billions of users and no per-token revenue. Serving a frontier model at frontier prices to that audience is not a business; serving a specialist trained on frontier behavior is. The frontier model becomes the research department, and the product becomes the cheap model that learned from it.
That is a real strategic shift, and it is not unique to Meta. Every lab is now trying to sell a top-tier model to subscribers while serving something cheaper to everyone else. What Meta has done is make the cheaper tier visibly good at the things consumers notice — remembering, persisting, not needing to be told twice — and quiet about the things they don't test.
Which is where the counter-case starts. Every figure above is Meta's own, on Meta's own index, with no third-party evaluation attached. The 67% improvement in finishing tasks has no published baseline, so there is nothing to compare it against. The Personal Superintelligence Index is not a standard benchmark; it is a number Meta defined and Meta reports.
The one independent read we have on the Wonder family is not unkind but it is not flattering either. When Artificial Analysis scored the Wonder models in October, Wonder landed at 52 on the Intelligence Index and Wonder Full at 56, behind Claude Opus 5.5 at 58, and about 30 on AA-Omniscience against Opus's 46 — the profile of a strong agentic coder that knows comparatively little, as we covered at the time. If the specialized models inherit that shape, the failure mode is specific and plausible: an assistant that never forgets your restaurant preference but is shaky on whether the restaurant is still open. Memory is not knowledge, and consumer trust tends to break on the second one.
There is also a distribution question Meta has not answered. The company did not name the new models, which makes them hard for outside evaluators to track, and it has not said whether memory will stay free once the habit is formed. Meta's own history with consumer AI features suggests the free tier is where adoption happens and the subscription is where the useful limits land.
What to watch: whether the specialized models get names and appear on independent leaderboards, which would turn Meta's index into something checkable. Whether Meta publishes a paper on the training method, since "trained using Wonder" is currently a claim without a mechanism. And whether Muse's App Store position moves — the app is a consumer product first, and a memory upgrade that users notice is the kind of thing that shows up in rankings before it shows up in evaluations.
If your assistant forgot something important this week, what was it — and did you notice at the time? Write to us via our contact page.
