Shodh-Memory
A local, offline memory engine for AI agents and robots, written as a single Rust binary that stores and recalls facts without calling any language model.
Shodh-Memory is a persistent memory system for AI agents and robots that makes no LLM calls to store or recall information. It builds a knowledge graph from observations using local, on-device models: MiniLM embeddings for semantic search, a small cross-encoder for reranking, and a GLiNER-based model for named-entity and relation extraction, all run through ONNX Runtime.
Memories that are used often become easier to find, unused context fades over time, and recalling one item can surface related ones. It serves chat agents over MCP and HTTP and robots over Zenoh or ROS2, needs no external databases, API keys or cloud services, and ships as a roughly 17MB binary.
The project positions itself against cloud-dependent memory systems such as mem0, Cognee and Zep, which need external LLM and vector-database calls to store and recall memories.