Does asking an AI (like Claude Code) to list and analyze its own mistakes make it start making more errors?

I've been testing a workflow with Claude Code where I added a rule in CLAUDE.md : read a FAILURES.md file to avoid past mistakes, and add new errors to that file when they happen. Strangely, it felt like the model started making more bugs and errors after doing this. It feels like a paradox. By filling the context with words like "error," "bug," and "failure," does it push the AI to write bad code instead of good code? Has anyone else tried a "failure log" like this and noticed it makes the AI perform worse?