PEAM: Parametric Embodied Agent Memory through Contrastive Internalization of Experience in Minecraft

Researchers introduced PEAM, a hybrid embodied-agent framework that pairs a deliberative LLM with a fast parametric memory module.

PEAM learns from failure-correction trajectories using behavioral cloning and contrastive objectives, then self-triggers consolidation when experiences should be stored.

In Minecraft tests, it improved long-horizon task performance, reduced catastrophic forgetting, and beat retrieval-based memory methods.

The work shows embodied agents can turn experience into lasting parameterized skills instead of depending mainly on external memory.