Researchers have built an AI agent memory system that splits fast decisions from slow reasoning, and the split makes it both smarter and dramatically faster.
The system, called Jev-Mem, borrows the System-One/System-Two framework from cognitive science. A lightweight System-One controller handles the routine work: organizing new memories, routing queries, deciding how much retrieval budget to spend, walking the memory graph, scoring candidates, and knowing when to stop searching. The heavier System-Two component, built on a large language model, only gets called in for genuinely hard reasoning and final answer synthesis. On the LoCoMo benchmark, Jev-Mem scored 0.777 on an LLM-as-a-judge metric, an 11 percent improvement over the best prior system, while building memory in 158 seconds - a 6.6x speedup over the fastest rival - and answering queries in 0.93 seconds, down 36.7 percent.
The interesting part isn't the accuracy bump. It's the efficiency gain. Most agentic memory systems lean on the same expensive LLM calls for both housekeeping and reasoning, which means every memory write or read burns tokens it doesn't need to. Jev-Mem's bet is that most memory operations are mechanical, not creative, and don't need a language model second-guessing them.
That's the same logic behind caching layers and query planners in ordinary databases - it just took agentic AI a while to catch up.