Tag
This paper introduces Geometry Guided Evidence Preserving Memory (GEM), a training-free compressor for long-horizon agents that optimizes for preserved task evidence over geometric coverage, reducing token usage while maintaining performance.