Raw-RAG tops 6 of 7 memory benchmark settings

The paper's target is a pattern that has become common in conversational-agent design: layering dedicated memory modules, write operations and retrieval policies on top of a model. The authors tested that assumption directly by running a deliberately plain baseline — dense retrieval over the raw conversation history, which they call Raw-RAG — against more elaborate agentic memory systems on four benchmarks.

The result is that the simple baseline came out ahead in 6 of 7 benchmark-model settings, with margins of 0.13–5.53 points. The speed comparison points the same way: Raw-RAG finished Add-plus-Search 5.1x faster than the next system and 30.4x faster than EverMemOS. The authors' conclusion is that memory-system complexity is not required for the evaluated conversational-memory tasks.

A caveat sits inside the result, though. The paper's scope is those evaluated tasks, so whether the finding carries over to other memory workloads is not established here — though the margin and the speed gap make it a result agent builders will likely want to check against their own setups.