Retrieval-Augmented Decision Transformer: External Memory for In-Context RL

Overview

This work explores in-context learning for reinforcement learning through a retrieval-augmented decision transformer. It considers how agents can infer a new task from a small set of relevant examples without retraining from scratch.

Why it matters

By introducing retrieval and external memory, the proposed approach gives a model access to useful prior experiences at inference time. The paper points toward RL systems that can adapt more effectively to new tasks while preserving the value of accumulated experience.