Meta's Routed CL matches agent memory to problem clusters, beating shared history

Meta researchers report that AI agents learn more from past attempts when each new problem is matched to notes from similar problems.

The method, called Routed CL, uses an LLM router to pick or create a memory cluster before each run, drawing on five memory types such as raw past runs or distilled lessons.

Across 398 mixed problems and 4 models, Routed CL beat a single shared history by 15% to 30% of the maximum possible score and cut estimated API cost by 65% to 72%, according to the paper as reported by Rohan Paul on X. The shared history itself barely beat no memory, with a mean gain of 0.015 out of 1.