Recursively summarizing enables long-term dialogue memory in large language models

Qingyue Wang, Yanhe Fu, Yanan Cao, Shuai Wang, Zhiliang Tian, Liang Ding

Abstract: Highlights•Proposed method generates real-time summaries to improve LLMs’ memory handling.•Approach integrates historical data, enabling coherent responses in long- term dialogues.•Approach complements existing LLM techniques, enhancing performance across various models.•Simple and flexible method for improving LLM response consistency over multiple sessions.