A paper on arXiv proposes AgentMemRec, a way for AI agents to choose which stored memories to bring into a limited context window. It treats each user request like a recommendation query and each memory as an item, then uses ranking and slate-building methods to select memories without calling an LLM.
The authors report that AgentMemRec beat several memory systems by 10 to 31 percentage points on multi-session questions in LongMemEval. It matched an LLM reranker that makes 64 calls per query while costing 3.2 times less to serve, and also outperformed Mem0 and Zep on LoCoMo without LLM calls when storing memories. These are results reported in the paper; independent validation is not described in the supplied material.
Why it matters
For teams building assistants that remember users across sessions, the method could reduce the cost of choosing relevant memories, though the reported benchmark results do not establish how it performs in deployed products.
