Meta's Muse Now Forgets 2.4x Less: Active Memory and Active Focus Come to Muse and Meta AI

What Meta Announced on 6 November

Meta began rolling out specialised versions of its existing models to Muse and Meta AI on 6 November 2026, trained using its Muse Wonder model. In Muse, the new models forget critical information 2.4 times less often, complete everyday tasks 67% better when unexpected problems arise, and score 47% on Meta's Personal Superintelligence Index, up from 38%. Active Memory is Meta's name for an assistant that remembers what matters from earlier conversations and recognises when that information becomes relevant again.

This is not a new Wonder model. Muse Wonder itself launched on 12 October 2026 as Meta's most capable model series, and it stays reserved for harder tasks and paying subscribers. What arrived on 6 November is different: the everyday models behind Muse and Meta AI, improved with what Wonder learned. Meta's own words are that it is "bringing some of those capabilities to even more people" — the frontier behaviour, distilled into the models that ordinary users actually talk to.

Active Memory: Remembering What Matters

Active Memory is the bigger of the two changes, and the one most people will notice first. An assistant that forgets what you told it last week is a frustrating tool. It asks the same questions again, it loses the thread of a project, and it treats every conversation as if it were the first. Active Memory is Meta's attempt to fix that: the assistant remembers what matters from earlier conversations, recognises when an old detail becomes relevant again, and builds on past conversations instead of starting over.

Illustration created for INORYVA

Active Focus: Not Giving Up When Things Go Wrong

Active Focus is the second ability, and it is about persistence rather than memory. An assistant that stops at the first obstacle is only useful for easy jobs. Active Focus is meant to keep Muse working toward a goal even when faced with unexpected challenges — to try harder, to recover from a failed step, or to find an alternative route on your behalf. Meta says it trained the models on a variety of bottlenecks that reflect the real world, so the assistant has practised the messy part of a task, not just the clean version.

Meta reports a 67% improvement in completing everyday tasks in Muse where unexpected issues arise. What Meta did not publish is the starting point: it did not say what share of such tasks the old model completed, so a reader cannot tell whether the improvement takes completion from, say, 30% to 50%, or from 60% to nearly all of them. A 67% improvement sounds large either way, but without the base rate it is impossible to judge how capable the assistant actually is at these tasks today. That gap is worth stating plainly, because it is the difference between a useful upgrade and a genuinely reliable one.

The Numbers, in One Place — and What They Do Not Prove

All of Meta's figures come from its own internal evaluations. No independent organisation has verified them, and Meta has not published a model card, parameter counts, or API pricing for the specialised models. That does not make the numbers false, but it does set the limit on what they can prove. Here is the full set, with the source of each figure stated.

Measure (Meta's own tests, Muse)	Previous model	Current model
Forgetting critical information	Baseline	2.4x reduction
Active Memory Eval — critical context used correctly	76%	90%
Everyday tasks completed where unexpected issues arise	Baseline (not published)	67% improvement
Personal Superintelligence Index	38%	47%

The Personal Superintelligence Index deserves a note of its own. It is Meta's own overall measure of how well an assistant helps with personal and professional tasks, and the current Muse model scores 47% on it against 38% for the previous model. Meta has not explained in the announcement text what the index actually measures, task by task, so the 47% figure is best read as Meta's own summary score — a direction of travel, not a benchmark a reader can compare against any rival. The same applies to the whole table: it shows that Meta's own tests improved, and it says nothing about how Muse compares with ChatGPT, Claude or Gemini on the same tasks, because no such comparison was published. Meta also did not describe user controls for this memory in the announcement — what a user can see, edit or delete — so readers should not assume controls Meta did not state.

Illustration created for INORYVA

Meta Spaces Is a Different Story

On the same day, Meta also launched Spaces in the Muse app — custom interactive tools built from chat. Spaces and the specialised models are easy to confuse because they arrived together and both involve Muse, but they are separate launches. Spaces is about building small interactive experiences from a conversation. The specialised models are about the assistant itself remembering better and persisting longer. This article covers only the models. Spaces deserves its own story, judged on its own facts, and it is not covered here.

The wider context is worth one paragraph. Meta is spending heavily to catch up in personal AI: NVIDIA $3 billion Meta investment in AI infrastructure was one of the biggest stories of last week, and Wonder is the flagship model that investment is meant to power. The specialised models are how Meta turns that flagship investment into something ordinary users can feel. Claude Duet is Anthropic's answer to the same problem — a proactive assistant for small businesses — and Gemini edit files shows Google pushing its assistant deeper into everyday documents. Every major lab is now competing on the same ground: not who has the smartest model in a lab test, but whose assistant remembers your life and finishes your tasks. Meta's 6 November update is its latest move in that race, and the 2.4x figure is the number it wants remembered. Whether users remember it will depend on whether the assistant remembers them.

Illustration created for INORYVA

Key Takeaways

• Meta rolled out specialised versions of its existing models to Muse and Meta AI on 6 November 2026, trained using Muse Wonder — this is not a new Wonder model launch (Wonder launched 12 October 2026).
• Active Memory helps Muse remember what matters across conversations; Meta reports a 2.4x reduction in forgetting critical information, with its Active Memory Eval rising from 76% to 90%.
• Active Focus keeps Muse working through unexpected problems; Meta reports a 67% improvement in completing everyday tasks where issues arise, without publishing the starting rate.
• Meta's Personal Superintelligence Index rose from 38% to 47% for the current Muse model — all figures are Meta's own internal evals, with no independent verification and no equivalent figures published for Meta AI.
• Meta Spaces, launched the same day, is a separate story and is not covered here; no model card, parameter counts or API pricing were published for the specialised models.

Conclusion

Meta's 6 November update is a quiet kind of launch with a loud claim. No new flagship model, no new price, no new app — just the everyday assistant, made to remember better and give up less easily. If the 2.4x figure holds up outside Meta's own tests, it will be the most useful upgrade Muse has had, because forgetting is the flaw users feel every day. If it does not, it will be another internal benchmark that looked better in the lab than in real life. The rollout is live, the controls for that new memory have not been fully described, and the independent verdict has not arrived yet. For now, the honest position is to welcome the improvement, watch how it behaves in real conversations, and judge it by whether it remembers what matters — and forgets what should be forgotten.

Sources

• Meta AI Research — Bringing Specialized Models to Muse and Meta AI, 6 November 2026 — https://research.meta.ai/blog/specialized-models-for-muse-and-metaai (primary source, opened and read in full at drafting)
• Unite.AI — Current Muse Scores 47% on Meta's Personal Superintelligence Index, 6 November 2026 — https://www.unite.ai/current-muse-scores-47-on-metas-personal-superintelligence-index/ (Active Memory Eval 90% vs 76%; Personal Superintelligence Index 47% vs 38%; Wonder launch 12 October 2026)
• ShiftDelete — Meta Muse Models Bring Active Memory to All Meta AI Users, 7 November 2026 — https://shiftdelete.net/en/meta-muse-models (2.4x derivation: failure rate 24% to 10%)
• AlphaSignal — Meta's Muse Wonder Cuts AI Forgetting 2.4x for Free Users, 6 November 2026 — https://alphasignal.ai/news/meta-s-muse-wonder-cuts-ai-forgetting-2-4x-for-free-users