The field of AI is in constant need for more scalable solutions. One promising option are Hopfield networks, that allow for an analog electronics implementation that is faster and much more energy-efficient than modern digital hardware. Unlike traditional feedforward neural networks, these networks also incorporate feedback connections, inspired by the structure of the human brain. Nonetheless, current models are typically limited to unimodal settings, ignoring the brain's highly-multimodal nature that encompasses modalities such as vision, audio, and touch. This proposal aims to address this limitation by integrating Hopfield networks with multimodality, with the ambition of creating highly efficient, robust, and multimodal neural networks. As the paradigm behind Hopfield networks and their training procedure is completely different from standard Deep Learning technology, fundamental challenges remain in terms of training stability and architectural building blocks with suitable inductive bias towards different modalities. To tackle these challenges, this proposal draws upon recent works that have improved crucial aspects of the Hopfield network, such as expressivity and training scalability, to design advanced network architectures and training strategies, complemented with proof-of-concept implementations. The proposed research aims at unlocking the full potential of Hopfield networks, paving the way towards extremely efficient neural network systems in the future.