Admittedly not in node, but I've found that writing end-to-end tests but substituting different 'units' for in-memory or not depending on the test marries the best of both worlds. In-memory tests tend to focus on behaviour.

In the rush to shoe-horn tests into a taxonomy, or 'integration' vs 'acceptances' vs 'unit' we often loose sight of the actual value proposition. Assisting design and building confidence to make changes.