Show HN: Mimirs – persistent local memory for AI coding agents (MCP) (github.com)

A new project named "Mimirs" has been announced, aimed at enhancing the functionality of AI coding agents by providing a persistent local memory system. The project, inspired by the Norse god of wisdom, streamlines the setup process to a single command, allowing coding agents to better retain context across sessions. This eliminates the cumbersome need to constantly guess filenames or sift through irrelevant files. On a real project, Mimirs reduced token usage during operations from approximately 380K to just 91K, resulting in a significant speed improvement of about 76%, showcasing its efficiency in accessing relevant information quickly. For the AI/ML community, Mimirs represents a major step forward in local AI capabilities, combining features such as AST-aware chunking and hybrid search techniques that utilize both vector similarity and traditional BM25 scoring. The tool is designed to facilitate enhanced code understanding while ensuring user privacy as it operates entirely on local machines without the need for external cloud services. With integration possibilities across multiple development environments, including Claude Code and GitHub Copilot, Mimirs aims to enhance coding efficiency and reduce costs associated with token-heavy models, marking a promising evolution in the realm of AI-driven code assistance.