Xuan He and ten co-authors evaluate dedicated memory systems for long-horizon agents across multiple benchmarks and model backbones.
The paper uses naive RAG over raw interaction history as its reference point. Its abstract says that many dedicated memory systems underperform that baseline in many cases. The authors call the gains from memory design “memory benefit” and the losses relative to raw-history retrieval “memory tax.” Reorganizing facts or distilling preferences may help a task, but can also remove or obscure evidence.
The authors say they will open-source code and results through the project page. The abstract does not provide per-system scores, identify which memory architectures fail in which settings, or establish that raw-history RAG is superior for every agent task or deployment.