The advancement of technology results in the escalating challenges posed by rapid urbanization and population growth in smart cities. Air pollution is not an exception to those challenges. It is proven that increased PM2.5 for air pollution has a significant impact on human health. This paper is based on development of a novel model using LSTM with connected memory for prediction of PM2.5. Initially, the memory module of the LSTM recurrent neural network is enhanced by consolidating the forget gate and input gate into a single modified gate. A sigmoid layer is incorporated to regulate the information update process. Moreover, information is added from the internal cell state to enhance the memory. This results in accurate and fast prediction of PM2.5 with reduced computational cost and high accuracy for recurrent dataset. The training data for the city Patna of the country India is retrieved from environmental pollution control board, India, from the year 2019–2023. The performance analysis shows that the proposed model outperforms the existing models with SVM, LSTM, RNN, and LSTM-wf in terms of MAE, RMSE, and R-squared. The proposed model demonstrates superior performance over compatible baselines such as LSTM-wf, achieving higher efficiency in terms of mean absolute error (MAE) and root mean square error (RMSE) by 18.1% and 9.10%, respectively, for 24-h forecasts. Moreover, the R-square value shows significant improvement as it is increased to 0.96. These results underscore the proposed model’s ability to deliver highly accurate predictions, surpassing other baseline models while also reducing computational costs.