Morning Edition · Friday, October 9, 2026Published at 2:20 AM EDT · New York

Agent-Controlled Forgetting Cuts a Debugging Session's Prompt Tokens From 912,000 to 232,000 Without Retraining

A separate paper finds that a widely cited gain from encoder-only healing in SVD KV-cache compression disappears once each method gets its own tuned learning rate.

Tool-using agents keep large tool outputs in context long after they have used them. Jan-Peter Franke's agent-controlled forgetting lets the acting model choose earlier tool results and replace each one with a short note at its original pos…