OpenVLHarness brings persistent multimodal memory to visual AI agents

The OpenVLHarness paper (arXiv:2610.26860) replaces raw image and JSON dumps in agent context with lightweight artifact handles stored locally, while math and geometry run in a sandboxed Python runtime. Across 22 benchmarks this yielded 9.7–12.6 point accuracy gains and a 63% cut in model-token consumption.

• Artifacts stay outside context; the model receives only artifact_id handles
• Calculations and geometry run in an isolated Python sandbox
• 9.7–12.6 point accuracy gains across 22 benchmarks
• Model-token consumption reduced by 63%