CAM: A Constructivist View of Agentic Memory for LLM-Based Reading Comprehension
Rui Li, Zeyu Zhang, Xiaohe Bo, Zihang Tian, Xu Chen, Quanyu Dai, Zhenhua Dong, Ruiming Tang
TL;DR
This work tackles the challenge of long-context reading with LLMs by introducing CAM, a memory module grounded in Piaget's constructivist theory. CAM builds a hierarchical, structured memory (structured schemata) that supports flexible assimilation (overlapping, multi-abstraction contributions) and dynamic accommodation (local hierarchy updates) and uses incremental overlapping clustering for memory development plus a prune-and-grow retrieval strategy for inference. Empirical results across six long-form reading benchmarks show CAM achieves superior accuracy and efficiency compared with both unstructured and structured baselines, including batch online updates that outperform full memory reconstructions. The approach demonstrates that a principled, cognitively inspired memory design can significantly improve autonomous LLM-based reading agents with practical implications for long-form QA, summarization, and claim verification. The work also discusses limitations, potential extensions to other domains, and avenues toward trainable memory controllers to further enhance adaptability and reliability.
Abstract
Current Large Language Models (LLMs) are confronted with overwhelming information volume when comprehending long-form documents. This challenge raises the imperative of a cohesive memory module, which can elevate vanilla LLMs into autonomous reading agents. Despite the emergence of some heuristic approaches, a systematic design principle remains absent. To fill this void, we draw inspiration from Jean Piaget's Constructivist Theory, illuminating three traits of the agentic memory -- structured schemata, flexible assimilation, and dynamic accommodation. This blueprint forges a clear path toward a more robust and efficient memory system for LLM-based reading comprehension. To this end, we develop CAM, a prototype implementation of Constructivist Agentic Memory that simultaneously embodies the structurality, flexibility, and dynamicity. At its core, CAM is endowed with an incremental overlapping clustering algorithm for structured memory development, supporting both coherent hierarchical summarization and online batch integration. During inference, CAM adaptively explores the memory structure to activate query-relevant information for contextual response, akin to the human associative process. Compared to existing approaches, our design demonstrates dual advantages in both performance and efficiency across diverse long-text reading comprehension tasks, including question answering, query-based summarization, and claim verification.