We propose a novel long-context processing framework — MemAgent, which directly optimizes long-context tasks through end-to-end Reinforcement Learning without altering the underlying model ...