A memory claim with the denominator missing

Start with the strongest number, because it is also the least checkable. A 2.4x reduction in forgetting sounds precise until you ask the obvious question: 2.4 times less than what? If the previous Muse model forgot goal-critical information in a quarter of tasks, the new one forgets it in about one task in ten, a change users would notice within a week. If the old rate was already low, the same multiplier describes an improvement most people would never feel. Meta's post does not say. The same gap sits under the 67% task-completion improvement for Active Focus: a relative gain, published without the completion rate it improves on, the number of tasks, or the tasks themselves.

Meta's accompanying charts also report an increase from 76% to 90% in correct use of critical context on its Active Memory Eval, and from 38% to 47% on its Personal Superintelligence Index. These are Meta-defined evaluations; the available announcement does not disclose their full task sets or establish independent replication. Endpoints help, but they do not settle much. The eval is Meta's, its task list is unpublished, and the index is a composite Meta designed and named itself. A nine-point rise on a scale the author invented, measured by the author, is a statement about Meta's confidence in its own progress. It is not yet evidence a reader can weigh against a competitor, because no competitor can be scored on it.

None of this makes the underlying work implausible. Teaching a model to separate the permanent from the temporary, and the relevant from the stale, is the correct problem to work on. The technical problem is not simply retaining more conversation history. An assistant must distinguish temporary instructions from durable preferences, recognize when information has become stale and retrieve the right context for a later task. Meta says Active Memory improves those behaviors. The announcement does not establish how the deployed products store, delete or expose remembered information to users. If Meta's specialized models genuinely make those distinctions more reliably, that is worth more in daily use than another point on a coding benchmark. The gap is between a plausible direction and a proven size.

One product got numbers. The other got adjectives.

Read the announcement by product and an asymmetry appears. Every quantified claim in it, the 2.4x, the 67%, the 90%, the 47%, describes Muse. Meta AI, the assistant with by far the larger audience, receives a paragraph of intentions: it can break complex problems into clear steps, ask sharper follow-up questions, and keep up when you change direction mid-conversation. Those may all be true. They are also unfalsifiable as written, and they arrive in the same announcement as a chart full of percentages for the smaller product.

The rollout details are similarly thin. The post says the specialized models are rolling out to people who use Muse and Meta AI, and outside reporting describes the upgrade as reaching everyone, with Wonder itself staying reserved for harder subscriber workloads. The post itself names no regions, no account tiers and no completion date, and it nowhere explains the training method beyond "trained using Muse Wonder": no data description, no model sizes, no card to inspect. For a change that alters what an assistant remembers about its users, the missing documentation matters more than the missing benchmark. Memory is the feature where users most need to know what is kept, what is dropped as stale, and who decided.

The practical test is the one Meta cannot run for you. If you use Muse across sessions, the next few weeks will show whether your goals survive context changes better than they did in October. That is a real signal, and it costs nothing to collect. What the announcement has not given you is a reason to switch, or to stay, based on 47% of an index only Meta can compute.

Sources

  1. [1] Meta AI Research, "Bringing Specialized Models to Muse and Meta AI", published 6 November 2026, read in full, including its charts: the rollout, both features, the 2.4x forgetting figure, the 67% task-completion figure, the temporary-versus-permanent training description and the Meta AI qualitative claims come from the post's text; the Active Memory Eval pair (76% to 90%, tasks correctly using critical context) and the Personal Superintelligence Index pair (38% to 47%) come from the post's own chart graphics, verified first-hand in the charts' embedded data.Read source
  2. [2] Unite.AI, "Current Muse Scores 47% on Meta's Personal Superintelligence Index", 6-7 November 2026, read in full. Secondary account reporting the index figures (47% versus 38%) and the Active Memory Eval pair (90% versus 76%) from Meta's announcement materials; the post links the Meta announcement as its source.Read source
  3. [3] AlphaSignal, "Meta's Muse Wonder Cuts AI Forgetting 2.4x for Free Users", published 6 November 2026. Independent secondary table reporting the identical index and memory-eval figures from Meta's announcement.Read source
  4. [4] aiweekly.co, "Meta ships Muse Wonder-trained models to all Muse and Meta AI users", alert published 7 November 2026, summarizing The New Stack. Rollout-to-everyone framing and the Zuckerberg quotation; no load-bearing figure rests on this item.