Definition
Plain language
The way an AI assistant quietly gets worse over time even when its model never changes.
As stated in the literature
Degradation of deployed agent reliability driven by changes in memory, retrieval state, and compaction rather than model weight updates; framed as a lifespan property with multiple distinguishable mechanisms.
Also called: aging
Why it matters: Most reliability work focuses on the model, but in deployed agents the surrounding memory is what quietly rots — so understanding aging is essential to keeping products usable.
For example, an assistant that worked great in week one starts giving stale recommendations a few months later because its memory store has filled up with redundant or outdated notes.
Heard on the show
“The authors give the phenomenon a name — agent aging — and they argue that it isn't one thing.”Episode 086 — Why Frozen-Weight Agents Still Get Worse Over Time