Temporal prediction is essential in many scientific and engineering problems, particularly where low power online inference is required. Reservoir computing is attractive for these tasks because a fixed recurrent dynamical system maps streaming inputs into a high dimensional state that can be decoded by a simple readout. Liquid state machines instantiate this idea with recurrent spiking neurons, yielding sparse event driven states that are well suited to neuromorphic and digital hardware. However, the common stepwise protocol trains the readout from an instantaneous reservoir state, which can be information limited when task relevant evidence is distributed across multiple time steps, especially for spike based representations where any single step may contain only a few binary events. In this study, we improve temporal prediction by providing the readout with an explicit finite temporal context. We concatenate a sliding window of recent reservoir spike states and train either ridge regression on the resulting vector or a transformer readout that attends over the same windowed sequence. On the Nonlinear AutoRegressive Moving Average (NARMA5) benchmark, increasing the context length improves both readouts. Under the standard stepwise protocol, ridge regression has a median test NRMSE of 2.106, which decreases as more temporal context is provided and reaches 0.317 at its best, an 85% reduction. The transformer readout shows the same trend, decreasing from 0.807 to a best value of 0.179, a 78% reduction. The same qualitative trend is further observed on NARMA10 and Mackey–Glass prediction tasks.These results demonstrate that exposing recent reservoir evolution at the decoder can substantially reduce the limitations of stepwise decoding while retaining the standard reservoir computing protocol