To improve the prediction accuracy and robustness of ammonia nitrogen concentration in recirculating aquaculture systems， this study introduces a memory-augmented compensation LSTM model （MA-LSTM-C）. The model uses bidirectional LSTM and Conv1D parallel structure to extract long and short sequence features at the same time， and embedding a memory matrix to fuse historical information and current input through multi-head attention mechanism. An ammonia nitrogen concentration interval attention mechanism with a learnable mask was introduced to strengthen the feature expression in the key concentration interval of ammonia nitrogen. Finally， through the dynamic error compensation layer， the prediction results were adaptively corrected by combining the instantaneous error， the average error and the trend error. To accommodate varying sampling frequencies from different sensors， PCHIP interpolation based on causal constraints is utilized for interpolated sampling of existing samples. Experimental results show that compared with the standard LSTM model， the MSE， RMSE， MAE and MAPE of MA-LSTM-C are decreased by 76.19%， 51.09%， 58.47% and 55.42%， respectively， and the R² is increased by 12.35%. Ablation experiments further reveal that incorporating physical mechanisms enhances each evaluation metric by approximately an average of 40%. This model demonstrates superior capability in capturing sudden fluctuations in ammonia nitrogen concentration and offers an effective approach for predicting ammonia nitrogen levels in recirculating aquaculture systems.