// OCR-Memory // Well this is a unique approach to store memory for long-horizon agents. Most of the agent memory systems compress trajector…

// OCR-Memory // Well this is a unique approach to store memory for long-horizon agents. Most of the agent memory systems compress trajectories into text summaries and hope the model remembers what matters. But that's where the information loss hides. Long-horizon agents need to revisit raw observations, tool calls, and procedural detail, exactly what summarization throws away. OCR-Memory takes a different route. It renders the agent's interaction history as images with indexed visual anchors, then retrieve via a locate-and-transcribe pipeline. The model scans visual memory, predicts the index of relevant regions, and the original text gets fetched verbatim from the database. No free-form generation, no hallucination. Older trajectories are stored as low-resolution thumbnails to keep visual-token cost flat, with active-recall up-sampling when a fading memory turns out to matter. SOTA on Mind2Web and AppWorld under strict context limits. Paper: https://arxiv.org/abs/2604.26622 Learn to build effective AI agents in our academy: https://academy.dair.ai/

Source: DAIR.AI (X) | 2026-04-30