SGMem: Sentence Graph Memory for Long-Term Conversational Agents

2025-09-25

AI Summary

In long-context dialogues, large language models (LLMs) face severe memory management challenges due to limited context windows; existing fact extraction or summarization methods fail to support multi-granular information organization and retrieval. To address this, we propose SGMem—a sentence-level graph-structured memory framework that unifies raw dialogue utterances and generative memories (e.g., summaries, facts, insights) into a heterogeneous sentence graph spanning three hierarchical levels: cross-turn, cross-round, and cross-session. This enables fine-grained semantic association and efficient retrieval. SGMem integrates graph-structured modeling, dynamic memory chunking, and generative context enhancement to significantly improve long-range information utilization. Evaluated on LongMemEval and LoCoMo benchmarks, SGMem achieves superior question-answering accuracy over strong baselines, demonstrating its effectiveness and novelty in organizing and retrieving memory for extended dialogues.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Long-term conversational agents require effective memory management to handle dialogue histories that exceed the context window of large language models (LLMs). Existing methods based on fact extraction or summarization reduce redundancy but struggle to organize and retrieve relevant information across different granularities of dialogue and generated memory. We introduce SGMem (Sentence Graph Memory), which represents dialogue as sentence-level graphs within chunked units, capturing associations across turn-, round-, and session-level contexts. By combining retrieved raw dialogue with generated memory such as summaries, facts and insights, SGMem supplies LLMs with coherent and relevant context for response generation. Experiments on LongMemEval and LoCoMo show that SGMem consistently improves accuracy and outperforms strong baselines in long-term conversational question answering.

Problem

• Managing long dialogue histories exceeding LLM context windows
• Organizing and retrieving information across dialogue granularities
• Providing coherent context for conversational response generation

Innovation

• Sentence-level graph representation for dialogue memory
• Combining retrieved raw dialogue with generated memory
• Multi-granularity context capture across turn-round-session levels