Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought

Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought. By Zhengye Bao and five others, from Xi'an Jiaotong University, Shanghai Jiao Tong University, and two other institutions. Published on the twenty-second of October. Its code repository is worked out: about six hundred and fifty source files, with a readme and a licence.

Memory for large language models is widely used in long-term conversation understanding, and agentic memory systems have grown elaborate. They mutate what they store, organise it into structures, learn retrieval strategies, or apply other complex operations. This work asks whether that complexity is always necessary, and answers with a deliberately plain baseline. Raw-RAG uses dense retrieval and reranking over the raw conversation history, and simply hands the retrieved evidence to a reader LLM. The claim is that this naive approach is stronger than expected.

The comparison runs across LoCoMo, LongMemEval, MEME, and MemoryAgentBench. In six of seven benchmark-model settings, Raw-RAG ranks first among the compared memory systems, beating the strongest compared agentic memory system by between zero point one three and five point five three points.

The authors then ask why raw-history retrieval stays competitive. Their evidence analysis points to a trade. Raw-RAG covers more of the necessary evidence, and that higher coverage can offset using it less efficiently. Useful evidence turns out to include not just the answer values themselves, but the relations and dependency paths needed to derive them. And even complete access to the transformed memory stores still leaves an accuracy gap relative to the raw evidence. Transforming memory loses something that the raw history keeps.

The boundaries are stated plainly. The findings challenge the necessity of complex agentic memory for the evaluated conversation tasks, not for every use of memory. The win is also not universal within those tasks: one of the seven settings did not go to Raw-RAG, and the margins range from a fraction of a point to a few points.

What the authors ask for is modest and practical. Naive retrieval over raw history should be treated as an essential baseline, one that any more complex memory system has to beat before its complexity is justified. Preserving the raw information, rather than rewriting it, is the lesson they draw from the gap that remains even with full access to transformed stores.