Has anyone found a good way to stop AI from rebuilding the same mistake twice?

I'm noticing something strange while using Cursor and Claude Code on larger projects.

The first time an AI makes a bad architectural decision, I fix it.

Three weeks later, I ask it to build something related...

...and it confidently recreates the exact same mistake because that earlier discussion is gone.

I've started keeping a small "things we've already learned" document that I point the AI to before bigger tasks.

It's helping, but it still feels like a workaround.

How are you preventing agents from repeating old mistakes on long-running projects?

Are you using CLAUDE.md, Cursor Rules, memory files, documentation, or something else?