The rapid expansion of data-driven technologies is amplifying the need for computing hardware that can process information with low latency and minimal energy consumption. In conventional systems, the physical separation between memory and processing units imposes a persistent processing latency and energy penalty, especially for workloads dominated by repeated vector–matrix operations. Neuromorphic based crossbar-arrays provide a compact route to in-memory computing by encoding synaptic weights as the conductance of crosspoint elements and leveraging parallel current summation for analog vector–matrix multiplication. Despite this promise, electrically programmed crossbar-arrays commonly encounter practical obstacles at the array level. During weight updates, unintended current pathways can introduce interference between selected and unselected cells, reducing tuning precision and complicating reliable multilevel operation. Consequently, iterative program–verify cycles are often required to converge to target weight states, increasing learning overhead and reducing system efficiency.

Here, we propose a crossbar-array that enables fully optical, bidirectional modulation of synaptic weights. A hardware artificial neural network is further presented that, for the first time, demonstrates a system-level, implementable identification system by performing vector–matrix computation in an optically programmable crossbar array. Optical stimuli can be selectively applied to target cells, minimizing disturbances in unselected cells and thereby improving the precision of weight tuning. In our flatform, each crosspoint incorporates a photoresponsive conducting polymer whose conductance can be adjusted continuously through cumulative light exposure. Rather than abrupt switching, conductance tuning is governed by exposure-dependent material changes that progressively modulate conductive paths within a distributed conduction network, enabling analog multilevel conductance control with high reliability and low array-level device-to-device variation (~5%). Because weight updates are driven solely by photonic stimuli, the update pathway is functionally decoupled from readout pathway (electrical), reducing read-disturb accumulation and improving the stability of the programmed states. To enhance scalability and learning, we implement photomask-assisted parallel programming that enables spatially selective, simultaneous weight updates across the array, improving learning speed compared with serial programming approaches. To demonstrate the system-level operational feasibility of the platform, we implemented and demonstrated a hardware fingerprint identification system that combines optically programmed weight patterns with in-memory vector–matrix operations. Our platform achieves a high identification accuracy (~90%), demonstrating the feasibility of array-level, purely optical weight programming for practical recognition tasks. Overall, these results establish a scalable strategy for optical weight control in crossbar hardware and highlight its potential to bridge proof-of-concept devices and system-level in-memory neuromorphic implementations.