But, it does a good job following existing conventions in a codebase, as long as they're really consistent. So the more actively you enforce that consistency the more likely it is to do the right thing without memories or prompting.
I don't like "never do" or "always do" type rules in AGENTS.md or in memory, as it often over-interprets them and ties itself in knots trying to satisfy an impossible set of goals.