Meta AI will forget less and give up less often with new upgrade

Meta is upgrading Muse and Meta AI with capabilities borrowed from its Muse Wonder model. The new models promise better memory, fewer task failures and smarter AI assistants that can keep working when things go wrong.

In Short

• Meta brings Muse Wonder features to Muse and Meta AI
• Active Memory helps AI remember important details
• Active Focus helps Muse recover from task failures

Meta has upgraded the models powering its Muse agent and Meta AI, borrowing two capabilities from its frontier Muse Wonder model. The update promises AI assistants that can remember important details from past conversations and keep working when a task hits a roadblock. In short, Meta wants its AI to do more than answer questions and actually help users get things done.

In a blog post, Meta said the specialised models would begin rolling out to Muse and Meta AI users on Friday. Mark Zuckerberg also announced the upgrade on X, saying the models were “2.4x less likely to forget critical info” and “67 per cent better at self correcting and completing everyday tasks”. Subscribers can still use Muse Wonder itself for more demanding tasks.

Meta AI gets better memory

The first major addition what the company calls is Active Memory, which helps AI assistants recall useful information from earlier conversations instead of treating every chat as a fresh start. Meta says its models can now distinguish between temporary details and information that remains relevant over time.

This could make a difference when users are working on something over several days. For instance, an AI assistant planning a trip should remember a user's budget or preferences without requiring them to repeat everything. Meta says its upgraded Muse models show a 2.4x reduction in forgetting critical information. In its internal evaluation, the models used important context correctly in 90 per cent of tasks, compared with 76 per cent previously.

The second capability, Active Focus, is about getting things done even when things go wrong. If a website fails, a booking process gets stuck or an unexpected issue appears, the AI is trained to try another approach instead of simply giving up.

Meta says its models show a 67 per cent improvement in completing everyday tasks when unexpected problems arise. The company also says they can better decide when to work independently and when to ask users for help, while adapting to corrections mid-task.

Why Meta is training specialised AI models

The update also reflects a bigger strategy. Instead of giving every user access to its most expensive frontier model, Meta is using Muse Wonder to train specialised models for everyday tasks. These models may not need all the knowledge of a frontier AI, but they need to be reliable at remembering context, following instructions and completing tasks.

That matters as AI assistants move beyond chat windows and start handling practical work. Muse is designed to act as a personal agent, while Meta AI is available across WhatsApp, Instagram, Facebook and Messenger.

Meta says the goal is to bring “even more capable superintelligence” to more people. The real test, however, will be whether these upgrades make its assistants genuinely more useful in everyday life.

- Ends