A new open-source project called agent-kernel offers a minimal approach to making AI coding agents stateful. The tool works by cloning a git repository and pointing any compatible agent—like OpenCode, Claude Code, or Cursor—at it. No database or complex framework required.

Instead, the system uses three simple markdown files and a git repo to create persistent memory. Agents read the kernel's instructions and automatically maintain separate identity and knowledge files. The structure includes AGENTS.md for generic kernel settings, IDENTITY.md for agent-specific details, and KNOWLEDGE.md for indexing facts stored in the knowledge/ directory.

Daily session notes go into append-only notes/ files, creating a running narrative of decisions and actions. This approach leverages existing mechanisms where AI agents already read project instructions like .cursorrules or CLAUDE.md. Multiple agents can run independently, each with its own identity and knowledge base, making it possible to have specialized agents for different domains like homelab management or investing—all using the same simple kernel system.