Reproduce “On Memory Construction and Retrieval for Personalized Conversational Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Xufang Luo, Hao Cheng, Dongsheng Li, Yuqing Yang, Chin-Yew Lin and 3 more
SECOM optimizes long-term conversational agents by introducing a segment-level memory construction strategy coupled with compression-based denoising. The method consistently outperforms existing memory management techniques on benchmarks like LOCOMO and Long-MT-Bench+, achieving higher GPT Scores with fewer tokens and improving retrieval accuracy.
Key insights
- Memory granularity critically impacts RAG performance; topically coherent segments achieve a better balance of relevant information and minimal irrelevance compared to turn-level or session-level units.
- Redundancy in natural language acts as noise for retrieval systems, and prompt compression serves as an effective denoising mechanism to mitigate this.
- Leveraging powerful LLMs like GPT-4 for zero-shot conversation segmentation, enhanced by a reflection mechanism for limited data, yields superior memory unit construction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.