🧠 Is agent memory really a better deal than plain RAG? Title: Memory Tax: Dissecting the Net Value of Agent Memory URL: ❓ What is the "net value" of memory? 💡 Compared with naive RAG, queries memory newly solves are its "benefit," and queries RAG solved but memory dropped are its "tax." The paper measures the balance between the two. ❓ What did they find? 💡 Across 10 memory systems and 3 benchmarks, the tax often outweighed the benefit. On LongMemEval-s, with RAG at 90.0, MemoryOS scored 80.0, MAGMA 76.0, and OpenViking 63.0. ❓ Why does the tax arise? 💡 Three reasons: ・Facts get lost during construction (38.0% of facts had no memory unit in OpenViking) ・Query-evidence similarity drifts after construction, making evidence harder to retrieve ・Swapping the backbone between construction and answering swings scores more than for RAG ❓ So how should memory be designed? 💡 The paper suggests also retrieving from raw history and reranking jointly with memory, and storing past queries and retrieval paths as recursive memory. #AIAgents# #rag#