Last released Oct 5, 2026
Training-free long-context memory compression for LLM conversations with structural critical-constraint retention (100% CRR across our benchmarks).