How should we evaluate memory for AI agents?