Feature selection plays a crucial role in machine learning and data analysis, enabling efficient model training, improved accuracy, and enhanced interpretability. However, traditional methods often struggle to achieve optimal feature subsets when dealing with high-dimensional data, redundant features, or complex interactions. The Crayfish Optimization Algorithm (COA) emerges as a bio-inspired approach inspired by the foraging behavior of crayfish, offering a promising solution to these challenges. This paper presents COA, a novel metaheuristic algorithm for feature selection that mimics the crayfish's exploration–exploitation balance. COA utilizes random walks for global exploration, guided by a fitness function that evaluates feature subset relevance and model performance. Simultaneously, COA employs a unique “pheromone-like” memory mechanism to exploit promising feature neighborhoods, accelerating convergence and enhancing accuracy. We evaluate COA's performance on diverse datasets and benchmark problems, comparing it to established feature selection algorithms. Our results demonstrate that COA significantly outperforms conventional methods in terms of selection accuracy, model performance, and robustness to noise and irrelevant features. Additionally, COA exhibits faster convergence and reduced computational complexity, making it particularly suitable for large-scale datasets.