PRAXIS-MD: a long-term brain-enhanced agent framework for end-to-end molecular dynamics research

Molecular dynamics (MD) simulation is widely used to investigate biomolecular stability, conformational transitions, molecular recognition and membrane transport, yet practical MD research still relies on many experience-driven decisions that are difficult to standardize and reproduce. Here, we present PRAXIS-MD, a case-guided problem-solving protocol that converts a natural-language biological question into an auditable MD workflow. The protocol couples an external long-term brain, comprising case memory, reusable skills and negative cases, with a brain-gated retrieve-plan-execute loop. This design grounds structure selection, system preparation, simulation planning, trajectory analysis and mechanistic interpretation in reusable prior experience rather than one-shot prompting. We distilled nine representative MD studies into reusable workflow knowledge and evaluated the protocol on nine unseen benchmark tasks under bare, memory-only, skill-only, rule-only and full PRAXIS-MD settings. Full PRAXIS-MD achieved the highest benchmark score, and integration with five additional LLM hosts improved performance across all hosts. In an independent membrane lipid transport case, PRAXIS-MD organized a coherent flippase/scramblase workflow without a predefined template. These results suggest that structured case learning can improve the reproducibility, transferability and biological interpretability of automated MD workflows.