Artifact fragment reassembly is a crucial component in cultural heritage preservation, but traditional manual methods highly rely on specialized domain knowledge and consume considerable time. While recent advances in computer-aided methods have shown promise, limitations such as insufficient utilization of multi-modal prior knowledge and weak scalability caused by strong rule constraints still cannot be ignored. For these challenges, this paper proposes AFR-Agent, an artifact fragment auxiliary reassembly framework based on multi-modal feature fusion representation. During training, AFR-Agent initially derives vector representation for multi-level visual feature and structured prior knowledge, followed by realizing multi-modal semantic fusion through encoder-based contrastive learning. In inference process, AFR-Agent first performs similarity search in metric space to preliminarily filter high-confidence candidate set. Then, a text-image LLM agent conducts secondary inferences, thus obtaining the final auxiliary reassembly result. Experimental validation on real-world dataset of 486 pottery fragments demonstrates that AFR-Agent achieves non-iterative precision of 71.8%. The code of AFR-Agent is available on GitHub.