E-mem: Multi-agent based Episodic Context Reconstruction for LLM Agent Memory

The article presents E-mem, a novel framework for episodic context reconstruction in Large Language Model (LLM) agents, designed to enhance System~2 reasoning by preserving contextual integrity. E-mem utilizes a heterogeneous hierarchical architecture with multiple assistant agents managing uncompressed memory contexts, leading to improved local reasoning capabilities. Evaluations on the LoCoMo benchmark show that E-mem achieves an F1 score exceeding 54%, outperforming the previous state-of-the-art GAM by 7.75% while also reducing token costs by over 70%, making it significant for practitioners focusing on efficient memory utilization in LLMs.