Meta paper shows routed memory beats shared history for agent learning
Splitting agent memory by problem type, rather than one shared log, delivers both better task performance and much lower token spend at the same time. Practitioners building agents that handle mixed task types can apply this by routing each new problem to a memory pool of similar past attempts. The 65% to 72% API cost reduction makes it especially relevant for anyone running agents at volume on metered model APIs.
A new Meta paper introduces Routed CL, a continual learning method for AI agents. An LLM router decides before each run whether to reuse an existing memory cluster or start a new one, with each cluster using one of five memory types such as raw past runs or distilled lessons. Across 398 mixed problems and 4 models, a single shared history barely beat having no memory, with a mean gain of 0.015 out of 1. Routed CL beat the shared-history baseline by 15% to 30% of the maximum possible score and cut estimated API cost by 65% to 72%.
New Meta paper finds that AI agents improve more from their past attempts when each new problem is matched to notes from similar problems, instead of one shared record of everything they have done.
A simple way to let agents learn from past runs is one shared history. With mixed tasks, it gets costly and mixes unrelated experience.
Routed CL adds an LLM router that reuses a memory cluster or starts a new one before each run. Each cluster uses 1 of 5 memory types, such as raw past runs or distilled lessons.
Across 398 mixed problems and 4 models, the shared history barely beat no memory, with a mean gain of 0.015 out of 1. Routed CL beat it by 15% to 30% of the maximum possible score and cut estimated API cost by 65% to 72%.
If your agents handle mixed task types, split their memory into separate pools by problem type instead of sharing one history.
