Training-Free Lexical-Dense Fusion for Conversational-Memory Retrieval

In three linesTraining-free lexical-dense fusion study for long-term conversational memory retrieval. Score-level fusion of late-interaction dense + BM25 improves Hit@1 by +8.8 to +17.2 points across six encoders (Hit@1 0.752 with e5-large-v2). Web search cross-encoder reranker degrades results (-6.9 pp). Analysis shows division of labor: dense excels on multi-hop/temporal questions, BM25 on adversarial ones.

Summary generated by Claude — human-verified