Meta pushes Muse Wonder's memory and persistence into the models behind Muse and Meta AI

Meta is bringing two capabilities from its frontier Muse Wonder models, Active Memory and Active Focus, to the far larger audience using its Muse agent and Meta AI assistant. The company announced on 6 November that specialised versions of its existing models, trained using Wonder, are rolling out to all users at muse.ai and meta.ai. Mark Zuckerberg framed the change simply: the models for Muse and Meta AI have been upgraded to bring Active Memory and Active Focus "to everyone".

Active Memory is Meta's term for holding on to information that matters over time and recognising when earlier context becomes relevant again. The company says it trained the models to distinguish temporary from permanent information, reason across time and identify when stored context has gone stale. On its Active Memory evaluation, the share of tasks in which the model correctly used critical context rose from 76 per cent to 90 per cent, which Meta summarises as a 2.4-fold reduction in forgetting.

Active Focus addresses the other common failure of agents: giving up or drifting when a task hits an unexpected obstacle. Meta says it trained on realistic bottlenecks so that Muse tries harder, recovers or finds alternatives, and reports a 67 per cent improvement in completing everyday tasks where problems arise. On its internal Personal Superintelligence Index the current Muse model scores 47 per cent against 38 per cent for its predecessor.

Meta has not named the specialised models, disclosed their size or committed to developer API access; the announcement covers only the Muse and Meta AI products. Muse Wonder itself remains reserved for the hardest workloads. All figures are Meta's own, and Muse has passed ten million set-ups in North America.

Why it matters

The industry's agent problem is less about raw intelligence than about memory and persistence. Meta's decision to productise those two properties specifically, and to distribute them free to its entire user base, signals where the competitive frontier for consumer agents now sits.

Indicops view

Memory and persistence are exactly the properties that decide whether an agent is useful in a government service context, where a citizen's case spans weeks and processes stall on missing documents. Indian teams building assistants should evaluate candidate models on long-horizon context retention and obstacle recovery, not just single-turn accuracy, and should design explicit memory stores with retention and deletion rules rather than relying on whatever a vendor model remembers implicitly.