Researchers have developed AgentOCR, a framework that uses visual tokens to represent the history of an agent's interactions. This approach compresses the accumulated observation-action history into a compact image, reducing token and memory costs. The system also introduces segment optical caching, which eliminates redundant re-rendering by decomposing history into hashable segments. Additionally, AgentOCR includes agentic self-compression, where the agent actively emits a compression rate and is trained to balance task success and token efficiency. --- Why it matters: This work matters because it addresses a significant challenge in practical deployment of large language models: the rapid growth of textual histories that inflate token and memory costs. By providing a scalable solution for multi-turn rollouts, AgentOCR enables more efficient use of resources in complex AI systems. Source: https://arxiv.org/abs/2601.04786