Definition
Plain language
When something you tell an AI assistant quietly makes an earlier fact about you no longer true, without anyone ever saying so directly.
As stated in the literature
In long-term memory benchmarks like STALE, a later observation that invalidates an earlier stored belief under common-sense world knowledge, with no utterance explicitly retracting the old belief; split into co-referential (same-attribute overwrite) and propagated (cascade through a logical dependency) types.
Why it matters: Without catching these silent updates, an assistant keeps acting on facts about you that have quietly become false.
For example, if you tell an assistant you just moved to Tokyo, it should quietly stop believing you live in your old city even though you never said you left.
Heard on the show
“They use it as the opening illustration of what they call implicit conflict, and that little example carries almost the entire argument of the paper.”Episode 031 — When Your AI Assistant Won't Let Go of Old Facts About You