Researchers tested whether complex agentic memory systems are needed for long-term conversation understanding by comparing them with Raw-RAG, a simple baseline using dense retrieval and reranking over raw conversation history. Across four benchmarks (LoCoMo, LongMemEval, MEME, and MemoryAgentBench), Raw-RAG outperformed the strongest agentic memory system in six of seven benchmark-model settings, beating it by 0.13 to 5.53 points. The authors attribute this to Raw-RAG retrieving more of the evidence needed to answer questions, finding that even full access to transformed memory stores did not match the accuracy of raw evidence. The results suggest naive RAG should be treated as an essential baseline when evaluating new memory systems.
