SGMem: Sentence Graph Memory for Long-Term Conversational Agents
This addresses memory organization and retrieval for long-term conversational agents, representing an incremental improvement over existing fact extraction or summarization methods.
The paper tackles the problem of long-term conversational agents struggling with memory management when dialogue histories exceed LLM context windows, introducing SGMem which represents dialogue as sentence-level graphs to capture associations across different granularities. Experiments show SGMem consistently improves accuracy and outperforms strong baselines in long-term conversational question answering on LongMemEval and LoCoMo benchmarks.
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.