Chapter 4 · Part 2
How Muse remembers you
One of Meta's headline claims for Muse is that it "remembers what matters to a person" and can "act on details that person only mentioned once." Their example is turning a recipe reel you saved on Instagram into a grocery list while remembering your dietary restrictions.
The surprising part is that the model can't remember anything at all. A language model is stateless. Each call starts fresh and knows only what's in that call's prompt. (That's also why long chats get expensive: the app has to resend the whole history every turn. See the context tax.)
So where does the memory live? Scroll to follow one fact from Monday to Thursday.
Monday, in passing: 'I'm vegetarian, btw.' The agent writes that fact to a memory file.
Memory is notes, not neurons
An agent's memory is a surprisingly low-tech idea. It's text the agent writes down and reads back later, and it involves two steps:
- Writing. During or after a conversation, the agent decides what's worth keeping. "Vegetarian" is worth keeping. "Asked about the weather at 3 pm" is not. Those facts get saved to long-term storage outside the model.
- Recalling. When a new task comes in, the agent searches that storage for anything relevant and pastes the best matches into the prompt, where the model can actually see them. This is the same retrieval trick behind RAG systems (Build a RAG Chatbot explains it in depth).
The model's knowledge stays frozen, and the prompt changes to include what it needs. A good memory system is really a good search system, because the hard part is finding the three relevant notes among thousands without flooding the context with noise.
You can read your own file
Meta's engineering post describes Muse's memory as files that live in your personal VM (more on that in Chapter 7), and says "you can inspect, edit and download these files freely, including Muse's memory about you." That's a useful design choice, because memory can be wrong. If the agent decided you're vegetarian after you ordered one salad, you can open the file and fix it.
Anything that writes to memory can shape the agent's future behavior. If a malicious web page persuaded an agent to "remember" a fake instruction, that instruction would show up in every later task. It's another reason Part 3's defenses matter, and a reason to look over what your agent has saved now and then.
Good memory is proactive
With memory, an agent can go beyond answering to anticipating. If it remembers Sam's birthday is October 3, it can suggest a gift in late September without being asked. Meta highlights exactly this: Muse can "make suggestions unprompted." Doing that needs one more capability, though: the agent has to be running when you're not talking to it. That's the next chapter.