Objective. Real-time closed-loop neural feedback control requires the analysis of action potential traces within several milliseconds after they have been recorded from the brain. The current generation of spike clustering algorithms were mostly designed for off-line use and also require a significant amount of computational resources. A new spike clustering algorithm, termed 'enhanced growing neural gas (EGNG)', was therefore developed that is computationally lightweight and memory conserving. The EGNG algorithm can adapt to changes of the electrophysiological recording environment and can classify both pre-recorded and streaming action potentials. Approach. The algorithm only uses a small number of EGNG nodes and edges to learn the neural spike distributions which eliminates the need of retaining the neural data in the system memory to conserve computational resources. Most of the computations revolve around calculating Euclidian distances, which is computationally inexpensive and can be implemented in parallel using digital circuit technology. Main results. EGNG was evaluated off-line using both synthetic and pre-recorded neural spikes. Streaming synthetic neural spikes were also used to evaluate the ability of EGNG to classify action potentials in real-time. The algorithm was also implemented in hardware with a Field Programming Gate Array (FPGA) chip, and the worst-case clustering latency was 3.10 µs, allowing a minimum of 322 580 neural spikes to be clustered per second. Significance. The EGNG algorithm provides a viable solution to classification of neural spikes in real-time and can be implemented with limited computational resources as a front-end spike clustering unit for future tethered-free and miniaturized closed-loop neural feedback systems.