Higher-order interactions shape complex neural dynamics but are hard to model. Here, authors use a generalization of the maximum entropy principle to introduce a family of curved neural networks, revealing explosive phaseions and enhanced memory via a self-regulating retrieval mechanism.