Developer Oguz Bilgic has released Agent Kernel, a minimal system that uses three markdown files to give AI coding agents persistent memory across sessions. The open-source project, available on GitHub with an MIT license, has garnered 130 stars since its release.
The system works with existing AI coding agents including OpenCode, Claude Code, Codex, Cursor, and Windsurf without requiring additional frameworks or databases. Users clone the repository, start their preferred AI agent, and the agent automatically reads the kernel files to understand how to maintain state.
Agent Kernel uses a specific file structure: AGENTS.md serves as the generic kernel that agents read as project instructions, IDENTITY.md stores the agent’s identity (maintained by the agent), and KNOWLEDGE.md indexes knowledge files. The system creates two types of memory storage—a knowledge directory for current facts that agents update when reality changes, and a notes directory for append-only session logs that preserve the narrative of what happened each day.
The approach leverages existing agent behavior where AI coding tools already read instruction files like AGENTS.md, CLAUDE.md, or .cursorrules. “This kernel uses that mechanism to teach the agent how to remember,” Bilgic explains in the repository documentation. “The agent doesn’t need a database, a vector store, or a custom framework.”
Each agent instance operates as its own git repository, allowing users to run multiple specialized agents simultaneously. The documentation suggests creating separate agents for different domains: “You can have a homelab agent, an investing agent, a health agent — all running the same OS.” Users can create additional agents by cloning the repository into new directories and initializing them with different AI coding tools.
