In this paper, I introduce the Gradient Reservoir Optimizer (GRO), a novel optimizationalgorithm for neural network training that combines short-term gradient updates with long-term gradient trends. GRO maintains a dynamic "reservoir" of recent gradient directions and utilizes their aggregated trends to influence parameter updates. By blending current gradients with a history-aware reservoir, GRO aims to stabilize convergence and improve robustness to noisygradients. This novel approach provides an additional mechanism to mitigate common issueslike gradient noise and plateaus in training loss. I demonstrate the theoretical underpinnings of GRO, provide its algorithmic structure, and evaluate its performance on benchmark datasets. The results show promise for GRO as a viable alternative to existing optimizers like SGD, Adam, and RMSProp. Additionally, GRO offers flexibility for tuning the influence of historical gradients, making it adaptable across a variety of tasks and architectures.