Think about an assistant’s notebook for one client. After a few meetings, the assistant writes down what keeps coming up: the client hates long emails, the owner signs off on pricing. They aren’t becoming a different person, just keeping notes so they don’t ask the same questions again. That’s project memory.
How it shows up
It doesn’t mean the model permanently learned something. The model isn’t retrained because you corrected a name or said you like a certain email style; the tool is saving a note, preference, or summary and bringing it back as context later. We have explained it this way: memory isn’t something you update every time; Claude can remember things and take notes, like “he likes to use LOL in his emails.”
Project memory shows up after repeated sessions: you correct a company name’s spelling or tell the agent a project principle, and later it uses that memory in a new session. This differs from project knowledge, the files attached to the project (the contract, the transcript, the brand guide). It also connects to a second brain, the larger place where your notes and decisions compound; project memory is smaller, scoped to the tool, and shouldn’t be the only source of truth.
Why you care
Keep memory clean. If the agent remembers something wrong, correct it; if a preference changed, update it; if it’s stale, remove it. Bad memory is worse than none because it feels helpful while quietly steering the work wrong. Project memory matters because repeated work should get easier, not restart from zero every morning.