Spectrally Partitioned All-Optical Memristors for Highly Linear Neuromorphic Vision Systems.

All-optical control of synaptic weights offers a promising route toward low-power and massively parallel neuromorphic hardware. However, existing optoelectronic memristors often require electrical assistance or exhibit nonlinear and asymmetric conductance updates, limiting their energy efficiency and scalability. Herein, we report a spectrally partitioned all-optical memristor based on a PbS/PEDOT:PSS/VOx heterostructure, in which near-infrared and visible light are coupled to opposite defect-state filling and depletion pathways. Near-infrared illumination selectively excites PbS quantum dots and drives gradual electron transfer into oxygen-vacancy-related defect states in VOx through the PEDOT:PSS-regulated interface, producing continuous conductance potentiation. By contrast, visible illumination activates VOx and promotes interfacial hole-assisted recombination, leading to reversible depletion of defect-state electrons and conductance depression. This wavelength-selective carrier-transfer process enables highly linear and symmetric all-optical weight updates, achieving a linearity of 0.9994 and an effective 8-bit conductance resolution. The electrical energy consumption per optically induced synaptic event is calculated to be as low as 0.63 fJ under an ultralow probing bias. By integrating the devices into a 32 × 32 all-optically programmed array, we further demonstrate hardware-based BloodMNIST microscopic blood-cell image classification with an accuracy of 96.5% and strong noise robustness. These results provide a viable strategy for developing high-linearity, low-power, and spectrally programmable neuromorphic hardware.