Meta takes on ChatGPT with long-term-memory AI models
Meta has unveiled its first frontier models, Muse Wonder and Muse Wonder Full. Their headline feature is 'active memory', the ability to remember what users tell them over long periods.
Could an AI remember what you told it days, weeks or even months later? That is the promise of Meta’s new models. Meta Superintelligence Labs has unveiled Muse Wonder and Muse Wonder Full, the company’s first frontier models.
What ‘active memory’ means
Alongside complex tasks such as programming, Meta has focused on remembering information users share over long periods and combining different pieces of information to act on them. It calls this capability ‘active memory’. Meta says Muse Wonder outscored rival models Fable, Opus and Astra on MemoryBench, a test of an AI’s ability to keep learning and retain information.
Pricing and availability
- On Meta’s model API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens.
- Muse Wonder Full is available only to Muse Code customers, not through the API.
- Both models will be built into Meta’s Muse AI agent.
The backstory
These are Meta’s first frontier models since Scale AI founder Alexandr Wang reorganised its AI operations last August. They were trained at Meta’s Prometheus data centre in Ohio, which came online earlier this year.
The idea that an assistant becomes more personal and useful the more it remembers has pushed big tech companies into a memory race. But keeping user information for long periods also raises sharper questions about privacy.
Comments