A new paper claims the fix for AI systems that misremember things about you isn't a bigger memory store - it's sorting the memories you already have.
The paper, titled "Fortunate Recall," introduces a policy layer that sorts personal facts an AI has stored about a user into eleven behavioral categories, then applies different rules to each: some facts decay over time, some get overwritten when a newer fact supersedes them, and some are tied to a specific point in time. Tested against a new 516-question benchmark called LifecycleBench, the system scored 76.9%, ahead of four existing memory tools - Mem0, A-MEM, Memory-R1, and MemoryOS - which scored between 61% and 70.5%. Compared directly against Mem0, the new approach cut the rate of invented answers from 45.1% to 22.4% while also answering more questions correctly. The results held up on a second AI model, Kimi K2.5, and a second, independently built benchmark called BEAM.
The interesting wrinkle is in the fine print: an ablation test found that generic bookkeeping metadata alone, not the elaborate category system, accounts for most of the raw accuracy gain. What the categories actually do is cut down on confident wrong answers, roughly halving the confabulation rate. That's a useful distinction for anyone evaluating a "smarter" AI memory product: getting answers right and not making things up can come from different parts of the system.
Worth noting: the headline benchmark, LifecycleBench, was built by the same team. The BEAM replication and the second-model test help, but this is one preprint, not an audited product.