Paper: Naive Raw-RAG Beats Agentic Memory Systems on 6 of 7 Benchmark-Model Settings, 30.4x Faster Than EverMemOS

A new paper argues the industry is over-engineering memory for conversational agents: a dense-retrieval 'Raw-RAG' baseline over raw history ranks first in 6 of 7 benchmark-model settings on LoCoMo, LongMemEval, MEME and MemoryAgentBench, beating sophisticated agentic memory systems by 0.13–5.53 points. UPDATE and DELETE write ops account for just 1.90% of decisions on LoCoMo and 1.34% on LongMemEval, and most are duplicates or malformed; Raw-RAG also completes Add-plus-Search 5.1x faster than the next system and 30.4x faster than EverMemOS. The authors conclude memory-system complexity is not required for the evaluated conversational-memory tasks.