Memory-Augmented Reasoning Optimization for Long-Horizon Language Model Tasks. JAAIR 2026, 5 (1).