This paper presents a novel approach to reservoir computing (RC) using thin-film transistors (TFTs) with controllable temporal dynamics. The authors fabricated an indium gallium zinc oxide (IGZO) TFT with adjustable memory behavior (FM), which allows for tuning the temporal dynamics of the reservoir. This device was used to construct a temporal adaptive reservoir capable of extracting information from multiple timescales, improving the accuracy of human activity recognition tasks. The system's performance was enhanced by a closed-loop architecture that dynamically adjusts the reservoir hyperparameters based on real-time feedback, enabling it to adapt to diverse input signals with varying speeds. The closed-loop RC system demonstrated accurate real-time recognition of objects moving at different speeds, showcasing its potential for processing spatiotemporal signals with compound temporal characteristics. The work provides a promising solution for self-adaptively optimizing RC systems without requiring prior knowledge of input signal timescales.This paper presents a novel approach to reservoir computing (RC) using thin-film transistors (TFTs) with controllable temporal dynamics. The authors fabricated an indium gallium zinc oxide (IGZO) TFT with adjustable memory behavior (FM), which allows for tuning the temporal dynamics of the reservoir. This device was used to construct a temporal adaptive reservoir capable of extracting information from multiple timescales, improving the accuracy of human activity recognition tasks. The system's performance was enhanced by a closed-loop architecture that dynamically adjusts the reservoir hyperparameters based on real-time feedback, enabling it to adapt to diverse input signals with varying speeds. The closed-loop RC system demonstrated accurate real-time recognition of objects moving at different speeds, showcasing its potential for processing spatiotemporal signals with compound temporal characteristics. The work provides a promising solution for self-adaptively optimizing RC systems without requiring prior knowledge of input signal timescales.