Reproduce “Structured Episodic Event Memory” on OpenResearch

About this paper

Zhengxuan Lu, Dongfang Li, Yukun Shi, Beilun Wang, Longyue Wang, Baotian Hu

Researchers from Southeast University, Harbin Institute of Technology (Shenzhen), and Alibaba Group developed Structured Episodic Event Memory (SEEM), a hierarchical memory framework for Large Language Models. This system combines dynamic episodic event frames with a static knowledge graph, integrating them via a hybrid retrieval process with reverse provenance expansion to enhance coherent context construction for long-term interactions, demonstrating up to a 4.4% absolute improvement in accuracy on the LongMemEval benchmark compared to prior memory frameworks.

Key insights

• A hierarchical memory structure that synergizes dynamic, event-centric episodic memory with static, relational graph memory provides superior coherence and logical consistency for LLMs in long-term interactions.
• Provenance pointers, especially when aggregated through associative fusion and utilized by Reverse Provenance Expansion (RPE), are crucial for reconstructing complete narrative contexts from initially scattered evidence.
• Structuring memory through event frames with detailed semantic roles allows for deeper social, causal, and temporal reasoning, mitigating the limitations of flatter memory structures.
• The SEEM framework proves to be model-agnostic and parameter-efficient, significantly enhancing reasoning performance even with smaller LLM backbones due to improved information density and narrative coherence.

Summary by alphaXiv.