MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning

In three linesMemOCR is a multimodal memory agent that compresses long interaction histories into structured images with adaptive information density. Trained via RL with budget-aware objectives, it outperforms text-based baselines on multi-hop and single-hop QA benchmarks under extreme context constraints.

Summary generated by Claude — human-verified