LightMem: Lightweight LLM Agent Memory with Small Language Models
A new memory system called LightMem uses Small Language Models (SLMs) to improve LLM agent memory. It modularizes retrieval, writing, and consolidation, separating online processing from offline consolidation for efficiency. Memory is organized into short-term (STM), mid-term (MTM), and long-term storage. The system aims to reduce latency while maintaining accuracy compared to large-model-based approaches.
Key facts
- LightMem is a lightweight memory system for LLM agents.
- It uses Small Language Models (SLMs) instead of large models.
- Memory operations are modularized into retrieval, writing, and consolidation.
- Online processing is separated from offline consolidation.
- Memory is organized into short-term (STM), mid-term (MTM), and long-term storage.
- The system aims to reduce latency in long-horizon interactions.
- It addresses accuracy issues in retrieval-based external memory systems.
- The paper is available on arXiv with ID 2604.07798.
Entities
Institutions
- arXiv