zjunlp/LightMem

What it solves

LightMem is a lightweight memory management framework designed to give Large Language Models (LLMs) and AI agents long-term memory capabilities. It addresses the challenge of efficiently storing, retrieving, and updating information from long conversations or interactions without consuming excessive computational resources.

How it works

LightMem uses a modular architecture that separates memory management into pluggable components. It processes information through a pipeline involving pre-compression (using tools like LLMLingua-2), topic segmentation, and embedding-based retrieval. It supports both online and offline memory updates and can be integrated with various LLM backends (including OpenAI, DeepSeek, Ollama, and vLLM) and vector databases (such as Qdrant and FAISS).

Who it’s for

This framework is intended for developers building intelligent applications, such as travel planning agents or code assistants, that require a persistent, efficient memory layer to maintain context across long-term interactions.

Highlights

• Modular Design: Supports custom storage engines and retrieval strategies.
• Broad Compatibility: Works with cloud APIs and local models via Ollama, vLLM, and Transformers.
• Efficient Processing: Includes pre-compression and topic segmentation to reduce token usage.
• MCP Server Support: Compatible with the Model Context Protocol for easier integration into AI ecosystems.
• Diverse Memory Methods: Includes specialized variants like FluxMem (graph-based), EM²Mem (multimodal), and StructMem (hierarchical).