Light Mem

Abstract

1. Introduction

LightMem: Lightweight and Efficient Memory-Augmented Generation

backbone: gpt-4o-mini, judge model: gpt-4o-mini & qwen2.5-32b-instruct

backbone: qwen3-30b-a3b-instruct-2507, judge model: gpt-4o-mini & qwen2.5-32b-instruct

2. Objective

[ICLR 2026] LightMem: Lightweight and Efficient Memory-Augmented Generation

This project demonstrates how Python can be applied to a real-world AI & Machine Learning problem.

4. Technology Stack

5. System Requirements

General requirements for this technology stack — check the README for exact versions.

• Python 3.8 or later
• pip / virtualenv for dependencies
• VS Code, PyCharm or Jupyter Notebook
• Git (to clone the repository)

6. Installation & Setup

git clone https://github.com/zjunlp/LightMem.git
cd LightMem

1. Modify the JUDGE_MODEL, LLM_MODEL, and their respective API_KEY and BASE_URL in API Configuration.
2. Download LLMLINGUA_MODEL from microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank and EMBEDDING_MODEL from sentence-transformers/all-MiniLM-L6-v2 and modify their paths in Model Paths.
3. Download the dataset from longmemeval-cleaned, and modidy the path in Data Configuration.

# Clone the repository
git clone https://github.com/zjunlp/LightMem.git
cd LightMem

# Create virtual environment
conda create -n lightmem python=3.11 -y
conda activate lightmem

# Install dependencies
unset ALL_PROXY
pip install -e .

pip install lightmem  # Coming soon

cd experiments
python run_lightmem_qwen.py

Full setup instructions are in the project README.

7. Future Enhancements

Suggested extensions you can add to make this your own project.

• Deploy the model as a web app with Streamlit, Flask or FastAPI
• Compare against an additional model and report the metric difference
• Add explainability (SHAP / Grad-CAM)

8. Viva / Review Questions

Common questions examiners ask for projects in this domain.

1. What dataset does the project use and how was it pre-processed?
2. Which algorithm / model architecture is used and why was it chosen over alternatives?
3. How are training and testing data split, and how is overfitting avoided?
4. Which evaluation metrics (accuracy, precision, recall, F1) are reported and what do they mean here?
5. How would you deploy this model for real users?

9. Source Code & License

This project is developed by zjunlp and published on GitHub under the MIT License. Please follow the license terms and credit the original author when you use or modify this code.