A month of my Claude Code agents writing down every correction: 247 lines, and most aren't about the model
retained text
I run about 15 Claude Code agents on my mac (family stuff, a few side projects, one reads my email). A month ago I gave all of them one rule: when I correct you, write it as one dated line in memory/lessons.md, and read that file at the start of every session. it's at 247 lines now. busiest agent has 47, a few have under 5. I thought it'd be model mistakes. it's mostly my mac lol - zsh doesn't word-split a variable, so `$CMD tick` fails with "no such file" - `find` here is bfs and rejects `-newermt '-30 minutes'`. with `2>/dev/null` on it a failed check looked like a pass, that one hurt - no PIL in python, use `sips` - tool calls die at 600s, long stuff goes to the background some are me being vague. I said "comments 5-6 times a day" and it couldn't know if I meant a count or how often. now it asks. some are it being sure without checking. told me a doc didn't exist, it was in the chat history. got a weekday wrong in a reminder, so dates are computed in code from the message time now. to keep it from rotting there's a nightly job that reads the day's chats and proposes add/update/delete on single lines, a script applies them so the model never rewrites the whole file. same correction twice means the line gets rewritten more specific. and once a week anything that shows up in 2+ agents moves to a shared file they all import, 19 rules there so far. anyone else doing this? what's on yours edit: since people asked, each agent loads only its own file + the shared one, ~3.5k tokens, cached, and it all lives in git
document fetched 2026-10-22 00:43 · canonical https://reddit.com/r/ClaudeCode/comments/1xbnwcw/a_month_of_my_claude_code_agents_writing_down