MemReread: Enhancing Agentic Long-Context Reasoning via Memory-Guided Rereading

Researchers have introduced MemReread, a novel framework designed to improve long-context reasoning in artificial intelligence agents while avoiding the quadratic computational complexity of standard attention mechanisms. Traditional agent memory approaches often suffer from information loss during linear processing, and recent retrieval-based solutions face issues with invalid queries and evidence degradation. MemReread addresses these limitations by employing a streaming reading process that triggers question decomposition and targeted rereading only when final memory proves insufficient. This method allows for the recovery of indirect facts discarded earlier, supporting non-linear reasoning without disrupting the logical flow of document comprehension. Additionally, the framework incorporates a reinforcement learning component that dynamically adjusts the number of rereading passes based on task complexity, thereby optimizing computational overhead and enhancing length extrapolation capabilities. Extensive experiments indicate that MemReread consistently outperforms existing baseline frameworks in long-context reasoning tasks while maintaining linear time complexity relative to context length. This advancement offers a more efficient and accurate approach for AI systems handling extensive textual data.