Stanford’s Memorilla memory system feeds frozen Qwen3-8B 16 memory tokens, beating full context on FactKG: 84.33% vs 80.43%.