Definition
Plain language
A benchmark and training approach for teaching an AI assistant to overwrite an old fact in its notes when a newer one replaces it.
As stated in the literature
A method that trains, via GRPO with a reward for temporal currency, the policy of recognizing when a new value cancels a stored one and using the current value — the supersession skill — in bounded-memory agents.
Why it matters: It matters because an assistant that keeps acting on outdated facts gives wrong answers, so learning to overwrite them is essential for reliable memory.
For example, if a user first says a meeting is at 3pm and later changes it to 4pm, Supersede trains the assistant to drop the old time and use 4pm.
Heard on the show
“This is a paper called "Supersede," posted to arXiv on June 25th, 2026, by a researcher named Vedant Patel.”Episode 180 — The Bug Where Smart Assistants Read a Fact and Still Forget It