Episodic memory dramatically improves robot vision-language-action AI performance

A new AI system called MemBodied uses a fixed-size episodic memory buffer to enhance vision-language-action models. The system achieved a mean task success rate of 50%, representing a sevenfold improvement over stateless baseline models. This performance increase was also validated on three real-world robot tasks, where success rates jumped eightfold. The memory buffer also significantly reduced inference latency and GPU memory usage. The research suggests lightweight episodic memory should become a new standard, challenging the reliance on recurrent hidden states alone.