Human Locomotion Mode Recognition (LMR) has the potential to be used as a control mechanism for lower-limb active prostheses. Active prostheses can assist and restore a more natural gait for amputees, but as a medical device it must minimize user risks, such as falls and trips. As such, any control system must have high accuracy and robustness, with a detailed understanding of its internal operation. Long Short-Term Memory (LSTM) machine-learning networks can perform LMR with high accuracy levels. However, the internal behavior during classification is unknown, and they struggle to generalize when presented with novel users. The target problem addressed in this paper is understanding the LSTM classification behavior for LMR. A dataset of six locomotive activities (walking, stopped, stairs and ramps) from 22 non-amputee subjects is collected, capturing both steady-state and transitions between activities in natural environments. Non-amputees are used as a substitute for amputees to provide a larger dataset. The dataset is used to analyze the internal behavior of a reduced complexity LSTM network. This analysis identifies that the model primarily classifies activity type based on data around early stance. Evaluation of generalization for unseen subjects reveals low sensitivity to hyper-parameters and over-fitting to individuals’ gait traits. Investigating the differences between individual subjects showed that gait variations between users primarily occur in early stance, potentially explaining the poor generalization. Adjustment of hyper-parameters alone could not solve this, demonstrating the need for individual personalization of models. The main achievements of the paper are (i) the better understanding of LSTM for LMR, (ii) demonstration of its low sensitivity to learning hyper-parameters when evaluating novel user generalization, and (iii) demonstration of the need for personalization of ML models to achieve acceptable accuracy.