Brief

A developer has created MemoryOS (MOS), a self-hosted microservice designed to manage long-term memory for large language models. The system utilizes Node.js for its backend, PostgreSQL with the pgvector extension for storing embeddings, and a separate Python service for local embedding generation. MOS incorporates a custom ranking algorithm that combines vector similarity with an importance score, includes memory expiration features, and offers basic prompt compression to reduce token usage. AI

Researchers have introduced S3MEM, a novel memory framework designed to improve long-horizon interactive question answering for AI agents. Traditional methods struggle with large trajectory histories, often retrieving incomplete evidence. S3MEM addresses this by structuring memory units and employing anchor-sensitive retrieval, creating a more efficient interface for inference. Evaluations on multiple environments show S3MEM consistently outperforms standard RAG and matches or exceeds other advanced memory systems while using significantly fewer tokens. AI