Meta's AI now remembers past chats and overcomes task roadblocks
What's the story
Meta has announced an upgrade to the models powering its Muse agent and Meta AI. The new capabilities, dubbed "Active Memory" and "Active Focus," are borrowed from the company's frontier Muse Wonder model. The update promises to make AI assistants more efficient by remembering important details from past conversations and continuing tasks even when they encounter obstacles.
Enhanced performance
Upgrades make AI assistants more efficient
Mark Zuckerberg took to X to announce the upgrade, highlighting that the models are "2.4x less likely to forget critical info" and "67% better at self correcting and completing everyday tasks."
The specialized models will start rolling out to Muse and Meta AI users from Friday.
For more complex tasks, subscribers can still use Muse Wonder itself.
Memory enhancement
Active Memory helps AI recall useful info from past chats
The first major feature of the update, Active Memory, enables AI assistants to remember useful information from previous conversations.
This way, they don't treat every chat as a fresh start.
Meta says its models can now differentiate between ephemeral details and those that are relevant over time. This could be especially useful for users working on long-term projects like trip planning.
Task completion
Active Focus helps in overcoming obstacles while completing tasks
The second capability, Active Focus, is all about getting things done even when obstacles arise.
If a website doesn't load or a booking process gets stuck, the AI is trained to try another approach instead of giving up.
Meta claims its models show a 67% improvement in completing everyday tasks when faced with unexpected problems.
They also know when to work independently and when to ask users for help while adapting mid-task corrections.
Training strategy
Meta's larger strategy to enhance AI performance
The update also reflects a larger strategy from Meta.
Instead of giving every user access to its most expensive frontier model, the company is using Muse Wonder to train specialized models for everyday tasks.
These models may not need all the knowledge of a frontier AI but must be reliable at remembering context, following instructions, and completing tasks.