Plain retrieval over raw chat history beats many purpose-built AI memory systems—and costs much less.
A new paper finds that plain retrieval-augmented generation (RAG) over raw conversation history beats many dedicated agent-memory systems on LongMemEval, PersonaMem and MINT-Eval, at much lower cost.
RAG means retrieving relevant information and supplying it to the AI when it answers. Here, that information comes from the conversation history.
The practical question is whether a dedicated memory layer earns its extra complexity. The reported result across these benchmarks gives practitioners a concrete reason to include retrieval over raw conversation history in that evaluation.
Before building a dedicated memory layer, benchmark plain retrieval over raw conversation history as the baseline on your own representative conversations. Compare answer quality and cost, then use those results to decide whether the dedicated layer is worth adding.