Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-horizon Agents

Researchers introduced PerMemBench, a benchmark for personalized memory in long-horizon AI agents with multi-year, multi-domain user histories.

They tested lightweight session-level gating that can skip memory operations for low-value interactions to save limited storage.

Personalized memory policies improved retention under ideal gating conditions.

However, the study found that accurate automatic gating remains unsolved, limiting practical deployment.