The authors evaluate representative agent memory systems against a naive retrieval-augmented generation baseline that searches raw conversation history directly, across several benchmarks and model backbones. They split each system’s net value into memory benefit (queries it answers that RAG misses) and memory tax (queries RAG answers that memory loses), tracing the tax to information loss during memory construction, reduced retrievability, and sensitivity to backbone changes. Naive RAG proves a strong baseline that many memory systems underperform, and simple fixes such as falling back to the source history at retrieval time reduce the tax.
