One-shot gradient-free learning on frozen transformers via hidden-state episodic memory and sleep consolidation. Tommi Niemi / Rotko Networks