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

Abstract:

This paper introduces PEAM, a framework for Parametric Embodied Agent Memory within the Minecraft environment. PEAM shifts the paradigm of agent memory by moving from inference-time retrieval to the internalization of skills directly within model parameters, achieved through experiential learning. The architecture couples a rapid, parametric module responsible for the reflexive execution of learned skills with a slower, deliberative Large Language Model (LLM) designed for open-ended reasoning. The high-speed component utilizes a multimodal Mixture-of-Experts LoRA structure featuring physically isolated adapters for each category, which facilitates continual learning at the parameter level while effectively preventing catastrophic forgetting.

In this approach, failure is treated as a primary training signal. By utilizing a combined objective of behavioral cloning and contrastive learning, the system internalizes trajectory pairs of failure and correction. This allows the agent to grasp not only successful outcomes but also the distinctions between corrected and failed actions. To manage the consolidation process, PEAM employs two key innovations: a parameterization-worthiness score that determines which experiences merit internalization, and a scale-free, self-triggered consolidation mechanism. This mechanism decides when to internalize data without relying on task-specific, hand-tuned thresholds, thereby enabling the agent to evolve autonomously as the trigger adapts across varying task distributions without the need for re-tuning.

Empirical results from Minecraft-based experiments demonstrate that PEAM enhances performance on long-horizon tasks and reduces forgetting regarding previously consolidated skills. Furthermore, the framework improves the efficiency of parametric memory compared to retrieval-based embodied agents and other parametric memory variants.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC