---
language:
- en
- zh
license: apache-2.0
pipeline_tag: feature-extraction
tags:
- embedding
- retrieval
- long-context
- rag
- qwen
---

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

🔗 **[GitHub Repository](https://github.com/MiG-NJU/EvoEmbedding)** | 🌐 **[Project Page](https://clare-nie.github.io/EvoEmbedding/)** | 📚 **[Training Dataset](https://huggingface.co/datasets/MiG-NJU/EvoTrain-180K)** | 📄 **[Paper](https://huggingface.co/papers/2606.21649)**

**EvoEmbedding** is a novel embedding model designed for long-context and dynamic retrieval scenarios. Unlike static embedding models that chunk text in isolation, EvoEmbedding maintains a continuously updated **Latent Memory Queue**. This allows it to capture temporal dynamics and generate *context-aware, evolvable embeddings* for precise retrieval in agentic workflows and long-conversations.

## 🚀 Quick Start

### Installation

First, clone the repository and install the dependencies:

```bash
git clone https://github.com/MiG-NJU/EvoEmbedding.git
cd EvoEmbedding
conda create -n evoemb python=3.10 -y
conda activate evoemb
pip install -r requirements-evoembedding-lite.txt
```

### Usage

#### As an Embedding Model

```python
from model.client import EvoEmbeddingClient

client = EvoEmbeddingClient()

messages = [
    {"role": "user", "content": "I visited Paris in April."},
    {"role": "assistant", "content": "Noted."},
    {"role": "user", "content": "I bought a new laptop yesterday."},
    {"role": "assistant", "content": "Got it."},
    {"role": "user", "content": "Where did I travel in spring?"},
]

embeddings = client.encode_messages(messages)
```

The `messages` input preserves the original dialogue order. `encode_messages` returns normalized embeddings for the history turns and the final query.

#### As a Reranker

```python
candidates = [
    "I visited Paris in April.",
    "I bought a new laptop yesterday.",
    "The meeting was moved to Friday.",
]
query = "Where did I travel in spring?"

ranked_candidates, ranked_indices = client.rerank(
    query,
    candidates,
    top_k=1,
    return_indices=True,
)
```

The reranker takes a direct list of candidate strings and returns them in relevance order.

## 📦 Model Family

We provide EvoEmbedding in three sizes based on the Qwen architecture:

| Model | Parameters | Base Model | Hugging Face Link |
| :--- | :---: | :---: | :--- |
| **EvoEmbedding-0.8B** | 0.8B | Qwen3.5-0.8B | [MiG-NJU/EvoEmbedding-0.8B](https://huggingface.co/MiG-NJU/EvoEmbedding-0.8B) |
| **EvoEmbedding-2B** | 2B | Qwen3.5-2B | [MiG-NJU/EvoEmbedding-2B](https://huggingface.co/MiG-NJU/EvoEmbedding-2B) |
| **EvoEmbedding-4B** | 4B | Qwen3-4B | [MiG-NJU/EvoEmbedding-4B](https://huggingface.co/MiG-NJU/EvoEmbedding-4B) |

## 📚 Citation

If you find this model or our methodology useful, please cite our paper:

```bibtex
@article{nie2026evoembedding,
  title={EvoEmbedding: Evolvable Representations for Long-Context Retrieval and Agentic Memory},
  author={Nie, Chang and Fu, Chaoyou and Feng, Junlan and Shan, Caifeng},
  journal={arXiv preprint arXiv:2606.21649},
  year={2026}
}
```