Ambisonics is a format for representing 3D sound fields, widely used in audio applications such as immersive spatial sound capturing and virtual reality. Accurately capturing Ambisonics signals with small or irregular microphone arrays is challenging, and deep neural networks (DNNs) have recently been used for this task. However, existing approaches often do not perform well in reverberant environments or poorly preserve the required inter-channel dependencies. In this work, we adapt a long short-term memory-based DNN, originally proposed for neural directional filtering, and apply it to first-order Ambisonics encoding. To address the limitations of previous approaches, we introduce a loss function motivated by the inter-channel structure of Ambisonics signals. Compared with an existing U-Net-based approach, the proposed method achieves a more accurate inter-channel coherence and lower magnitude error, with a slight tradeoff in magnitude error induced by the coherence-aware loss.