agent-kernel

agent-kernel provides a method for creating stateful AI agents using only three markdown files and a git repository. It enables agents to retain memory between sessions, take notes, and build on previous work without requiring a separate framework or database. md files as project instructions, such as OpenCode, Claude Code, Codex, and Cursor, among others.

md indexes knowledge files. Additional directories include knowledge/, which holds mutable facts about the current state of the world, and notes/, which contains append-only daily session logs. This structure allows agents to update facts as reality changes and maintain a narrative of decisions and actions taken during each session. Each agent operates in its own repository, making it possible to manage multiple agents with distinct identities and knowledge bases using the same kernel.

To use agent-kernel, users clone the repository, start their coding agent within the cloned directory, and interact with the agent, which will prompt for its identity and remember information provided. The tool does not require a database or additional infrastructure, as all state and memory are managed through the markdown files and git. For those seeking additional features such as integration with Telegram, Slack, and daemon mode, a separate runtime called kern-ai is available and built specifically for agent-kernel, supporting multiple channels and user management.

agent-kernel is positioned as a tool for developers or users who wish to create and manage AI agents with persistent memory using a lightweight, file-based approach.