Joint Agent Memory and Exploration Learning via Novelty Signals

Researchers introduced JAMEL, a framework that jointly trains agent memory and exploration policy using novelty-driven interaction.

It uses persistent novelty signals such as code coverage in GUI tasks to supervise memory without manual labels.

Experiments showed stronger exploration, better transfer to unseen environments, and lower token usage than open baselines.

The approach helps agents generalize more effectively in open-ended environments while reducing compute costs.