As artificial intelligence, the Internet of Things, and intelligent sensing technologies advance at a breakneck pace, there's a growing need for edge devices to seamlessly handle an enormous volume of temporal data streaming in from sensors, including those for dynamic vision and sound. Physical Reservoir Computing (PRC) harnesses the inherent nonlinear response and short-term memory capabilities of devices, offering a groundbreaking hardware approach for low-power, near-sensor intelligent computation. Nevertheless, conventional PRC approaches typically rely on segmented input, resetting the device state after each sample. When it comes to processing genuine, uninterrupted event streams, devices with multi-timescale memory persistently build up residual states from prior inputs. This accumulation results in a drift in the working baseline, ultimately compromising the stability of long-term inference.