Reproduce “Ever-Evolving Memory by Blending and Refining the Past” 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
Seo Hyun Kim, Keummin Ka, Yohan Jo, Seung-won Hwang, Dongha Lee, Jinyoung Yeo
Researchers from Yonsei and Seoul National Universities developed CREEM (Contextualized Refinement based Ever-Evolving Memory), a framework that equips chatbots with dynamic long-term conversational memory by blending new dialogue with past information and refining existing memories. This approach leads to more consistent and personalized interactions, achieving higher G-Eval scores for memory quality and improved response generation compared to existing baselines.
Key insights
- Explicitly blending current dialogue with a broader set of contextually searched past memories is crucial for generating coherent and 'evolved' long-term insights.
- A nuanced, categorization-based memory refinement strategy (New, Redundant, Updated) effectively maintains memory consistency while preventing excessive information loss, outperforming simple deletion.
- Treating memory construction and insight generation as an inseparable reasoning step for response generation significantly enhances the coherence and contextual awareness of chatbot outputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.