XSkill enables continual learning for multimodal agents

XSkill addresses the "stateless" bottleneck of current LLM agents by providing a structured way to remember and reuse successful strategies without retraining.

• –Dual-stream architecture decouples immediate tool selection (Experiences) from long-term task planning (Skills)
• –Parameter-free approach allows any off-the-shelf VLM to improve continuously through a closed-loop accumulation phase
• –Visually grounded retrieval ensures that the agent's "memory" is contextually relevant to the current state
• –Outperforms traditional learning-based baselines across diverse benchmarks, including web navigation and tool use
• –Represents a significant step toward autonomous agents that actually get smarter the more they work