Raw-RAG Leads Compared Memory Systems in Conversation Benchmark Study
An arXiv preprint introduces Raw-RAG, a baseline that uses dense retrieval and reranking over raw conversation history to supply evidence to a reader LLM. Across LoCoMo, LongMemEval, MEME, and MemoryAgentBench, the authors report that Raw-RAG ranked first among the compared memory systems in six of seven benchmark-model settings, outperforming the strongest compared agentic-memory system by 0.13 to 5.53 points. Their evidence analysis found that broader coverage of necessary information—including answer values and the relations and dependency paths needed to derive them—could offset lower evidence utilization.
The study challenges the necessity of complex agentic-memory operations for the evaluated conversation tasks and establishes naive RAG as an essential baseline. The findings are limited to a single preprint and the evaluated benchmarks, models, and compared systems.
Computer Science > Artificial Intelligence
arXiv:2610.30755 (cs)
[Submitted on 22 Oct 2026]
Title:Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought
Authors:Zhengye Bao, Shengtao Zhang, Haokai Zhang, Shiqi Lou, Muning Wen, Minnan Luo
View a PDF of the paper titled Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought, by Zhengye Bao and 5 other authors