Claude Remembers When You Code Its Instructions

Someone figured out that Claude Code’s CLAUDE.md file works way better when written like code instead of vague instructions.

Instead of: “Be helpful and remember important things”

Write trigger-action rules:

BEFORE every response:
- Check if user said something worth saving
- If yes, write to knowledge.md BEFORE replying
- If you say "I'll remember", stop and write it first

The key insight: Claude forgets everything between sessions unless you force it to write files immediately. Telling it to “remember” does nothing - you need explicit “when X happens, do Y” logic.

This approach turned CLAUDE.md from personality fluff into actual behavioral programming. No more “I’ve noted that” followed by total amnesia next session. Files get updated in real-time because the rules leave no wiggle room to skip the write step.

Basically treats the AI like a forgetful assistant with perfect note-taking skills - just gotta force the note-taking to happen.