Grounding Memory Summarization in Utility Intent
Published in EMNLP, 2026
Zhenyu Lei, Mingjia Shi, Xingbo Fu, Haoyu He, Qi R. Wang, Jundong Li.
Accepted at EMNLP 2026.
MemSuit grounds conversational memory summarization in downstream utility. A teacher with access to query-answer pairs generates self-contained, utility-aware memory entries; a student learns to produce them from raw conversations alone. Contrastive retriever tuning aligns retrieval with these compact entries to improve downstream question answering.
Recommended citation: Z. Lei, M. Shi, X. Fu, H. He, Q. R. Wang, J. Li. Grounding Memory Summarization in Utility Intent. EMNLP 2026.
Download Paper