Economic time series forecasting represents one of the most significant challenges in economics, directly impacting economic policy decisions. This research investigated the applicability of Long Short-Term Memory (LSTM) recurrent neural networks for forecasting Brazilian inflation, measured by the IPCA, from January 2015 to March 2025. The primary objective was to evaluate the predictive capacity of LSTM-based models by comparing different variable configurations and time horizons. Four distinct LSTM models were developed: Model 1 used 22 predictor variables, including 12 IPCA lags and macroeconomic variables (IGP, Selic, unemployment, GDP); Model 2 employed 16 variables, combining extensive IPCA lags with selected economic variables; Model 3 implemented a univariate architecture with 3 IPCA lags; and Model 4 used exclusively 12 consecutive IPCA lags. Different hyperparameters were tested via grid search, such that each of the four models was tested with 84 distinct combinations, totaling 336 different configurations. Variables were temporally standardized, and the data were split into training and testing sets at an 80/20 ratio, maintaining chronological order. The results demonstrated that Model 4, utilizing exclusively 12 IPCA lags in a univariate architecture, achieved the best overall performance with the lowest MAE (0.1834) and MSE (0.0669) values. This finding reinforces the importance of inflationary memory and seasonal patterns in Brazilian price dynamics, providing evidence that inflation persistence can be captured by machine learning models, making them a viable tool to complement traditional econometric models in the formulation of monetary policy