Definition
Plain language
An AI assistant's ability to tell, when asked directly, whether something it remembers about you is still true.
As stated in the literature
In the STALE benchmark, the task of directly determining whether a stored memory remains valid given later information; distinguished from premise resistance and downstream policy adaptation, which models often fail even when state resolution succeeds.
Why it matters: It tests whether an assistant knows when its stored facts have gone stale, a basic requirement for memory that stays trustworthy over time.
For example, asked 'do I still live in Chicago?', the assistant correctly recognizes from a later message that you've since moved away.
Heard on the show
“" They call this state resolution.”Episode 031 — When Your AI Assistant Won't Let Go of Old Facts About You