Mem-π: Adaptive Memory through Learning When and What to Generate

Researchers introduced Mem-π, an adaptive memory framework for LLM agents.

It uses a separate language or vision-language model to decide when to generate guidance and what that guidance should be.

The system is trained with a decision-content decoupled reinforcement learning objective.

Mem-π outperforms retrieval-based and earlier memory baselines on web navigation, terminal tool use, and embodied text tasks.