🧠 Supersede: the supersession demo

Across a long, multi-session chat, a fact about the user changes (they move city, switch jobs, get a new car). A good agent must answer with the current value, not a superseded one. Models are good at recalling what they were told and bad at dropping what is no longer true, so an assistant keeps acting on your old city, job, or car. We call correct handling of these updates supersession.

Result: GRPO training on the Supersede RL environment, where the reward is temporal fact-currency, lifts held-out supersession accuracy on LongMemEval (knowledge-update, oracle) from 9.0% (base Qwen2.5-3B-Instruct) to 16.7%.

📄 Paper · 🤖 Model · 📊 Dataset · 💻 GitHub · 🌐 Prime Intellect Hub

Browse a real episode below: read the sessions, then reveal the current answer and the superseded traps.