EvoEmbedding: Evolvable Representations for Long-Context Retrieval and Agentic Memory

EvoEmbedding is a new embedding model designed for long-context retrieval, introducing evolvable representations that adapt based on dynamic input sequences. It utilizes a continuously updated latent memory alongside raw content to generate context-sensitive embeddings, addressing the limitations of static models. The model is trained on the EvoTrain-180K dataset and demonstrates superior performance on long-context retrieval benchmarks, outperforming larger models like Qwen3-Embedding-8B and KaLM-Embedding-Gemma3-12B, while also enhancing agentic workflows in retrieval-augmented generation (RAG) systems.