Graft – semantic memory for AI agents, without the LLM (github.com)

Graft introduces a novel approach to improving memory retention in AI agents by providing a persistent graph memory system that operates locally without dependency on cloud services or extensive configuration. With its single binary and SQLite database, Graft allows AI systems like Claude Code, ChatGPT, and Codex to efficiently save and retrieve learned information across different sessions and contexts in mere milliseconds. This local-first model ensures user privacy and ease of migration, enabling smoother transitions between machines and environments. This innovation is particularly significant for the AI/ML community as it addresses a common limitation where AI agents forget earlier interactions and context when a session ends. Graft's architecture emphasizes a verified semantic cache that reduces unnecessary calls to large language models (LLMs), thereby cutting costs and improving response times. By linking memories through a graph structure using keywords and semantic edges, Graft ensures more accurate memory retrieval, solving the problem of knowledge loss during AI interactions and enhancing the overall efficiency of AI applications.