ICML 2026 paper. SOTA LoCoMo 86.3%, LongMemEval 87.4%. Up to 98% fewer context tokens vs full-history baseline. Open source on GitHub.