Raw-RAG beats complex agentic memory systems on 6 of 7 long-term memory benchmarks
omarsar0 · x · 2026-10-23
A new paper from XJTU, SJTU, Zhejiang and CUHK shows plain RAG over raw chat history beats complex agentic memory systems on most long-term memory benchmarks.
- Results: Raw-RAG (dense retrieval + reranking over raw conversation) ranks first in 6 of 7 benchmark-model settings across LoCoMo, LongMemEval, MEME and MemoryAgentBench, ahead of the strongest agentic system by 0.13–5.53 points.
- Efficiency: BEAM-1M ingestion takes 0.3 hours vs 3.2–27 hours for Mem0, MemOS and Memora.
- Why: raw history preserves evidence a question needs, including relations and dependency paths; rewriting into memory entries drops them. Even full access to transformed memory stores leaves an accuracy gap.
Conclusion: complex agentic memory is over-kill for these tasks; naive RAG is an essential baseline.
Related event: Study: Plain RAG Beats Complex Agentic Memory on Most Benchmarks(2 posts)→
More from coding & agent
- Figma MCP Server converts Figma designs into pixel-perfect React code — modelcontextprotocol · 2026-10-23
- Gemini 4 Argon builds a 3D Gothic cathedral from one image via Blender MCP — GoogleDeepMind · 2026-10-23
- Reddit thread asks: which tasks do your agents fail at, forcing you to pay humans? — Aggressive-Buyer8583 · 2026-10-23
- claim MCP server string-matches citations against sources and flags fabricated quotes — BondableColt18 · 2026-10-23
- AI agents take macOS kernel vulns to root: an XNU exploit dev case study — moyix · 2026-10-23
- Wavelength SDK hits mainnet, letting humans and AI agents send bitcoin with a few API calls — RSync25 · 2026-10-23