<p>Reservoir computers (RCs) are a class of recurrent neural networks that incorporate brain-inspired principles and provide an efficient alternative to deep learning. With fixed random internal connections and trained output weights, they simplify learning but remain sensitive to hyperparameters governing activation and connectivity. While various hyperparameters are commonly tuned, the relative balance between excitatory and inhibitory (E-I) signals—fundamental to brain function—is typically fixed in RCs. Here, we investigate tuning this balance and show that strong performance consistently arises in balanced or slightly over-inhibited regimes, not excitation-dominated ones. Further, we introduce a self-adapting mechanism that locally adjusts E-I balance to achieve target firing rates, reducing hyperparameter tuning costs and yielding up to 130% performance gains in memory capacity and time-series prediction. Incorporating heterogeneity in firing-rate targets further enhances robustness. These findings highlight dynamic adaptation as a promising design principle, improving RC performance while offering insights into neural computation.</p>