Naive RAG beats complex agentic memory on 6 of 7 benchmarks, XJTU study finds
XianJiaotongUniversity · hf · 2026-10-23
A Xi'an Jiaotong University team's RAW-RAG study challenges the consensus that complex agentic memory systems are necessary. Naive RAG with dense retrieval and reranking over raw conversation history ranks first in 6 of 7 benchmark–model settings across LoCoMo, LongMemEval, MEME, and MemoryAgentBench, beating the strongest agentic memory system by 0.13–5.53 points. Evidence analysis shows higher coverage of necessary evidence offsets lower utilization, and even full access to transformed memory stores leaves an accuracy gap versus raw evidence—preserving raw information matters more than elaborate memory transformations.
More from coding & agent
- Dev shares two AGENTS.md lines to make AI-generated UIs look less like AI — pjausovec · 2026-10-23
- Meta, Google, OpenAI and xAI all converge on cloud computers for personal agents — sujingshen · 2026-10-23
- diffy semantic diff app to auto-discover Git worktrees in next release — seatedro · 2026-10-23
- How a 5-person AI film crew produced an Odyssey pilot scene in Imagine — chaitu · 2026-10-23
- "One prompt built it all!" Frontier dev: we refactor like this several times a day — fkasummer · 2026-10-23
- Study: LLMs chime in at unwelcome moments 46.4% of the time in group chats — windx0303 · 2026-10-23