![[2601.21468] MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2601.21468] MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning
Summary
MemOCR introduces a multimodal memory agent that enhances long-horizon reasoning by using layout-aware visual memory, optimizing context utilization under budget constraints.
Why It Matters
As AI systems increasingly require efficient memory management for complex reasoning tasks, MemOCR's innovative approach to visual memory could significantly improve performance in applications requiring long-context understanding, such as natural language processing and robotics.
Key Takeaways
- MemOCR utilizes a structured rich-text memory to optimize long-horizon reasoning.
- The system adapts memory allocation based on visual layout, prioritizing critical information.
- Reinforcement learning is employed to train MemOCR under varying memory budgets.
- MemOCR outperforms traditional text-based memory systems in multi-hop and single-hop reasoning tasks.
- The approach offers a promising solution for efficient context utilization in AI applications.
Computer Science > Artificial Intelligence arXiv:2601.21468 (cs) [Submitted on 29 Jan 2026 (v1), last revised 21 Feb 2026 (this version, v3)] Title:MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning Authors:Yaorui Shi, Shugui Liu, Yu Yang, Wenyu Mao, Yuxin Chen, Qi GU, Hui Su, Xunliang Cai, Xiang Wang, An Zhang View a PDF of the paper titled MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning, by Yaorui Shi and 9 other authors View PDF HTML (experimental) Abstract:Long-horizon agentic reasoning necessitates effectively compressing growing interaction histories into a limited context window. Most existing memory systems serialize history as text, where token-level cost is uniform and scales linearly with length, often spending scarce budget on low-value details. To this end, we introduce MemOCR, a multimodal memory agent that improves long-horizon reasoning under tight context budgets by allocating memory space with adaptive information density through visual layout. Concretely, MemOCR maintains a structured rich-text memory (e.g., headings, highlights) and renders it into an image that the agent consults for memory access, visually prioritizing crucial evidence while aggressively compressing auxiliary details. To ensure robustness across varying memory budgets, we train MemOCR with reinforcement learning under budget-aware objectives that expose the agent to diverse compression levels. Across long-context multi-hop and single-hop qu...