How language agents learn to remember better by sleeping on it

Language agents running multiple tasks accumulate clutter in memory. Auto-Dreamer separates the quick recording of session experiences from slow offline consolidation across sessions, where it inspects old memories, finds patterns, and replaces bloated regions with compact summaries. Trained on ScienceWorld tasks using agent performance as reward, it beats all baselines while using 12× less memory on ScienceWorld and generalizes to ALFWorld and WebArena with 6× less memory than prior methods—no retraining needed.