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

Researchers introduce PEAM, a parametric memory framework for AI agents in Minecraft that consolidates learned skills directly into model parameters rather than relying on retrieval-based memory. The system uses a mixture-of-experts architecture with contrastive learning to internalize both successful and failed experiences, achieving better long-horizon task performance while avoiding catastrophic forgetting.

PEAM represents a fundamental shift in how embodied AI agents store and execute learned knowledge. Rather than maintaining separate retrieval databases that slow inference and consume memory at runtime, the framework internalizes skills as permanent parameter modifications, creating a more efficient and scalable approach to agent learning. This addresses a critical bottleneck in current embodied AI systems where memory retrieval becomes computationally expensive as experience accumulates.

The research builds on growing recognition that large language models can function as both reasoners and skill executors. By pairing a slow deliberative LLM for planning with a fast parametric module for execution, PEAM optimizes the speed-reasoning tradeoff inherent in agent design. The innovation of treating failures as primary training signals—rather than merely learning from successes—mirrors human learning and enables more robust skill development.

The scale-free self-triggered consolidation mechanism holds particular significance for practical deployment. Previous parametric learning approaches required manual tuning of when and what to consolidate, creating friction in real-world applications. By automating this decision through a parameterization-worthiness score that transfers across different task distributions, PEAM eliminates a major implementation barrier.

For the broader embodied AI landscape, these advances suggest that parameter-resident memory could eventually replace hybrid retrieval-parametric systems as computational efficiency becomes increasingly critical. The Minecraft experiments validate the approach's generalization potential, though real-world robotics applications remain the ultimate test. The framework's ability to continuously improve without manual intervention positions it as a stepping stone toward more autonomous, self-directed AI agents.

• →PEAM consolidates agent memories into model parameters rather than retrieval databases, improving inference speed and efficiency.
• →The mixture-of-experts LoRA architecture with isolated adapters enables continual learning without catastrophic forgetting of previous skills.
• →Failure-correction trajectory pairs treated as first-class training signals enable agents to learn from mistakes as effectively as successes.
• →Scale-free self-triggered consolidation automates when to internalize experience, eliminating task-specific manual tuning requirements.
• →Experiments demonstrate improved long-horizon task performance and better parametric efficiency compared to retrieval-based embodied agent alternatives.