Meta has launched Muse Wonder and Muse Wonder Full, two AI models that its own tests put close to the best from OpenAI and Anthropic. Meta is selling them on more than raw intelligence. It built them to carry what matters to a person from one conversation to the next, and to bring back the right detail when a later task needs it.
Muse Wonder is available to developers and in Meta's coding tool. Wonder Full is only for subscribers to that tool. The Muse assistant for everyday users gets Wonder soon, subscribers first.
Memory is not new, but now it is the headline
ChatGPT and Claude have had memory for some time, and Muse already kept a memory file for each user. What has changed is where memory sits in the pitch. A company with a model near the top of the field has made remembering you the main selling point.
That makes sense once you see how the competition is going. Abilities are converging, and so are prices: Wonder costs $4 per million input tokens and $20 per million output tokens, roughly what the leading rival models cost. When ability and price barely separate the choices, the thing left to separate them is what each assistant has learned about you.
The longer you use it, the dearer it is to leave
Take the head of a school department who spends a year teaching an assistant which teachers cannot cover which periods, and how she likes letters to parents worded. If she corrected it just twice a week, that is about a hundred lessons it would need again after a switch. A rival could be 10% better at every task and she would still stay, because the first month on a new assistant would be worse than the last month on the old one.
The 2003 phone rule shows what changes this. Once a number could move between carriers, the carrier had to keep customers by being better, not by holding something that was theirs. Memory has the same shape, and I expect people to ask for the same right over it within a few years.
What the memory holds is more than what you typed
Muse is more open than most on the visible part. Your memory is a file you can read, edit and delete, and you can download your chats and files. Claude already lets you bring over context from ChatGPT, Gemini and Copilot through a copy-and-paste import.
The harder part is what the assistant concludes by itself. Muse reviews each day's conversations overnight and writes down what works on a particular user, such as which hour short nudges land best. Its help page also says that after you delete something, Muse may still remember what it learned from it. A memory like that is only partly yours to carry away.
Keep your own copy of how you work
The practical answer is to own the part you can. Keep one page in your own documents that says how you work: your suppliers, your tone, your standing rules, the things you are tired of repeating. Paste it into whichever assistant you use. Then the memory that matters most lives with you, and changing assistants costs an afternoon, not a year.