Claim that “AI also ages like humans when used continuously,” with performance declining as memory accumulates across sessions

University of Texas at Austin researchers say long-running AI agents can lose performance as memories accumulate and get updated.

They identify four causes of degradation: compression, interference, revision, and maintenance.

They also released AgingBench, a benchmark to measure how AI agents age over time.

Tests across 14 models and 7 scenarios found that some agents kept behavior but lost factual accuracy.