Agent memory is an agent's ability to remember information about a user — preferences and decisions already made — across multiple conversations and tasks. It doesn't store the chat history word for word; instead it keeps conclusions like "the user likes concise replies" or "the last decision was option A", so they can be used directly next time without the user repeating everything. The key is the boundary: memory saves repeated confirmation, but it only remembers what it is allowed to remember.
Agent memory has three layers — don't mix them up
The top layer is the context window: the text the model can directly see in this one conversation. Close the conversation and it's gone. The middle layer is long-term memory: the agent extracts facts from conversations and writes them into a memory store, to be retrieved next time — a job, frequently used tools, preferences. The bottom layer is a memory file or user profile: some agents let users view, edit, or even import a description of themselves — essentially an instruction manual about you, written by you. The three layers cost different amounts: context is the most expensive and the easiest to forget; long-term memory is the most convenient, but a wrong retrieval can mix people up; a memory file is the most controllable, but you have to maintain it yourself.
Agent memory is not the same as RAG
Both involve looking something up and then answering, but they serve different goals. RAG attaches an external knowledge base to the model and searches documents before answering — it solves "things the model doesn't know". Agent memory solves "things the model doesn't know about you": it stores information about the user, not general knowledge. Put simply, RAG is like looking something up; memory is like recognizing someone. One agent can use both: documents via RAG, your preferences via memory.
Three common misunderstandings, cleared up first
First, deleting a conversation does not delete the memory. Many products store the chat history and the memory separately. Deleting the chat only removes the interface; the extracted memories may still be there and need to be cleared separately in the memory or privacy settings — here's how to handle it when Muse still remembers you after you delete a conversation. Second, memory does not mean the model was trained on it. What you tell an agent doesn't immediately become part of its weights; it's only recorded under your account and doesn't exist under another account. Third, more memory isn't better. Too much outdated information misleads instead — change jobs and it keeps recommending based on the old one. Cleaning up regularly matters more than accumulating endlessly. Memory's value lies not in quantity, but in accuracy and freshness.
People who work with the same agent often, on tasks that continue over time, benefit most: the format of a weekly report, usual coding conventions, regular contacts — remember once and you save the repeated explanation every time. On the other hand, for private information, shared devices, or a question asked only once in a while, leaving it off is the safer choice. The rule of thumb is simple: if you'd be happy to say it again next conversation, it doesn't need remembering; if you want the agent to act on it by itself next time, turn memory on.
