A context engineer changes memory retrieval settings and wants to establish whether cross-session personalization improves. Which evaluation is most useful?

1. Compare versioned multi-session tasks with known preferences, checking correctness, privacy, and retrieval relevance.
2. Assume more recalled memories always improve every answer.
3. Measure only the number of memory rows stored.
4. Test only a single session containing the preference in its latest user message.