You should consider testing our RemembrallMCP (note I'm the author). It substantially cuts token usage and you get better results in codebases. https://github.com/roboticforce/remembrallmcp