Definition
Plain language
A benchmark that tests whether an AI assistant can recall facts buried in a long past conversation.
As stated in the literature
A long-term conversational memory benchmark measuring fact recall over extended dialogue histories; representative of the recall-focused memory evaluations that conflict-detection benchmarks like STALE contrast against.
Why it matters: It reveals whether an assistant truly retains details over a long conversation instead of quietly forgetting earlier context.
For example, it might ask an assistant 'what restaurant did I say I disliked back at the start of our chat?' after many intervening messages.
Heard on the show
“There are big benchmarks for this — LoCoMo, LongMemEval.”Episode 031 — When Your AI Assistant Won't Let Go of Old Facts About You