Memory management is one of the most challenging parts of working with Claude Code; too little effort or too much, and you waste tokens and Claude gets confused.

> "We attempted to use CLAUDE.md and continue to do so. Our root-level CLAUDE.md helps communicate some of the rules of our repo, such as approaching changes via test-driven development (TDD), as well as tribal knowledge our team has internalized. However, we don’t want to overload it with information about every area of the codebase, given context window constraints and our desire to avoid confusing Claude with irrelevant details."

Having issues with Claude [dot] md seems to be a common experience, and leveraging rules and having a background agent analyze each session is a clever approach here that works well. I've found this to be incredibly helpful.