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

In three linesPEAM is an embodied agent memory framework in Minecraft that internalizes experience as parameters rather than inference-time retrieval. It pairs a slow LLM for reasoning with a fast parametric module (Mixture-of-Experts LoRA) learning via behavioral cloning and contrastive objectives. Failures are treated as training signals to learn corrected actions.

Summary generated by Claude — human-verified