Is it better to use Claude Projects with CLAUDE.md files for a specific project, housing project-specific skills within an overall CLAUDE.md, rather than re-teaching context every time?

One approach shared: use a project and drag-and-drop relevant files into it rather than re-explaining context each time, since even with memory features Claude can act inconsistently and not truly retain context despite claiming to. A practical workaround is to explicitly ask Claude to pull up all prior chats related to a topic (e.g., "pull up all my chats about 7216, give me a summary of what I've taught") and have it spend a few minutes researching and telling you where you left off. However, the email/chat search feature was described as not reliable — it sometimes pulls in the wrong emails or chats even when searching for something specific.