Long-horizon agents accumulate growing interaction histories, which increase context size and inference cost. We find that geometric redundancy alone is insufficient for safe compression. Although histories exhibit strong low-dimensional structure, similar global geometry can contain vastly different amounts of task evidence.
With identical retained block counts, evidence-aware selection raises the next‑action Top 3 retention from 0.31 to 0.69 while centroid similarity stays at 0.98. Controlled replacement experiments further show that action‑related information can be substantially altered while global geometric measures remain nearly unchanged.
Motivated by the gap between geometry and evidence, we introduce Geometry Guided Evidence Preserving Memory (GEM), a training‑free compressor that protects task and execution evidence before using geometric residuals to achieve coverage. GEM reduces mean combined token usage from 2.69M to 2.11M per task, a 21.4% reduction, while maintaining comparable task reward.
Our results indicate that efficient agent history compression should optimize for preserved task evidence rather than geometric coverage alone.
Review