New framework beats RAG by 56% on ALFWorld. Uses two-phase retrieval with learned Q-values instead of passive semantic matching. No fine-tuning needed.