Manages attention and evidence in long agent sessions. Use for lost instructions, dropped evidence, or large multi-agent contexts.
