Device-level stochastic switching enables tunable short-term memory in magnetic tunnel junction neurons

Short-term memory is essential for temporal information processing, yet its hardware implementation commonly requires recurrent connections or additional state storage. Here, we demonstrate a device-level short-term memory mechanism enabled by stochastic switching in magnetic tunnel junction neurons. Probabilistic RESET failures retain the neuronal state across time steps, generating an exponentially decaying memory trace whose duration is directly controlled by the RESET probability. This relationship was established theoretically and verified through device-level state-retention measurements, delayed-matching, and n-back tasks. We further incorporated stochastic-RESET neurons into an SEW-ResNet18 architecture and evaluated N-CARS classification, Time-Reversed DVS Gesture recognition, and DDD20 steering prediction. Appropriate RESET probabilities consistently outperformed the memory-free condition, while different tasks favored different memory durations. These results demonstrate tunable, compact, device-native short-term memory for efficient spatiotemporal computing.