everyone's screenshotting the same @karpathy line this week: stop using @claudeai to write code — point it at a folder, build yourself a second brain. the folder is the easy part. i had that in a day. it did almost nothing for two weeks. what makes it compound is the part nobody quotes: what gets written back, and when. → after every task, the agent appends ONE line to memory.md — a decision it made and why. not a summary of what happened. a decision. → anything it had to work out twice becomes a skill file. once it's a file, it's never re-derived again. → raw material — pdfs, transcripts, threads — goes in a dated inbox folder. the agent reads it on the next run and files what actually matters. → memory.md gets read FIRST, every single run, before it touches the task. a memory file nobody opens is just a file. → one weekly pass: delete anything in there that turned out wrong. a stale fact is worse than no fact. week 1 this feels like pure overhead. month 2 the agent stops asking you things it already learned, and the gap between "new chat" and "knows my whole setup" quietly disappears. the model stays swappable — claude, gpt, whatever wins next month. the memory and the git history of how it grew are yours. bookmark this. the "point it at a folder" posts will keep coming. this is the part they skip. #AI #ClaudeCode #AIagents