This PhD thesis is part of the growing field of embedded AI for the Internet of Things (IoT), where constraints in energy, area, and connectivity require rethinking the learning mechanisms of neural networks. The goal is to design neuromorphic architectures based on 3D hybrid synapses combining FeRAM and ReRAM, within an in-memory computing framework. The objective is to enable local adaptation of the model—drawing from machine learning approaches and potentially compatible with plasticity mechanisms such as STDP, VDSP, etc.—while maintaining efficient inference adapted to naturally asynchronous information. The PhD student will develop a heterogeneous memory architecture, design an appropriate local learning protocol, and implement integrated circuit demonstrators. Experimental validation on edge-relevant tasks (e.g., sensory classification) will assess power consumption, network accuracy, and adaptability. Publications and patents are expected outcomes of the thesis.