Is Complex Agentic Memory Always Needed? Naive RAG Proves Stronger in Long Conversations
Original title:Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought
As agentic architectures grow increasingly intricate with specialized memory mutations and graph structures, a new study demonstrates that a simple retrieval baseline can outperform them. Termed Raw-RAG, the naive approach applies dense retrieval and reranking directly over raw interaction history. Across four conversational benchmarks—LoCoMo, LongMemEval, MEME, and MemoryAgentBench—Raw-RAG ranked first in six of seven settings, surpassing complex agentic memory frameworks by 0.13 to 5.53 points. The authors find that while raw logs contain noise, they preserve the relational context and dependency paths that structured transformations routinely discard.
Why it's worth reading
It directly challenges the industry's rush toward convoluted agentic memory systems by proving that standard retrieval over raw conversation history reliably preserves essential context that structured memory loses.