We design and implement a temperature-based input mechanism for molecular reservoir computing. Using temperature allows us to interact with the system while keeping it chemically closed, a crucial step to use the reservoir computing approach with standard laboratory equipment. We implement the reservoir with a robust molecular oscillator, subjecting it to sudden temperature variations and monitoring its response with fluorescent reporters. We then train in-silico neural networks on the fluorescence traces to predict the inputted temperature profiles. We reach an average of 87% accuracy for a single layer and 91% for two layers, showing the potential of such reservoir.