[AI Minor News Flash] Can AI Have ‘Memory’ with Just 3 Markdown Files? Introducing the Minimal Agent Framework ‘Agent Kernel’

📰 News Summary

• Minimal Setup: A system has been released on GitHub that allows AI agents to have ‘memory’ using just three Markdown files (AGENTS.md, IDENTITY.md, KNOWLEDGE.md) and a Git repository.
• No Framework or DB Needed: By reading and writing files, the AI retains its role and past decisions without relying on complex databases or vector searches.
• Compatibility with Existing Tools: Works with any AI coding agents capable of reading directive files within the project, such as OpenCode, Claude Code, Cursor, and Windsurf.

💡 Key Points

• Two Types of Memory Structures: Information is organized into a ‘knowledge/’ section (variable) for managing facts, and a ‘notes/’ section (append-only) for logging daily session data.
• Self-Updating Identity: Upon first launch, the AI asks the user, “What do you want me to be?” and writes the response into IDENTITY.md to remember it.
• Complete Portability: You can create new agents simply by cloning the repository, making it easy to manage specialized agents for different purposes, like ‘investment’ or ‘health monitoring’.

🦈 Shark’s Eye View (Curator’s Perspective)

This is a brilliantly clever implementation that takes advantage of the “existing habits” of AI agents prioritizing directive files! It’s astonishing that such practical statefulness can be achieved by simply teaching the AI the “memory writing rules,” without the need for a complex vector DB. The differentiation between notes/ being append-only to maintain history and continuously updating knowledge/ for current understanding is a remarkably concrete and rational approach!

🚀 What’s Next?

We may see a trend where specialized AI capable of ‘memory’ is mass-produced without heavy agent frameworks, simply by cloning repositories. This will likely accelerate the personalization of ‘companions’ in personal local development environments.

💬 One Shark’s Take

Creating ‘memories’ without needing a database is a perfect minimalist invention for us sharks! Time to dive in and make the most of it! 🦈🔥

📚 Glossary

• Stateful: The property of saving past interactions and states, allowing for continuous processing based on that history.

• Repository: A place where files and update histories are stored. Here, Git is used to version control the AI’s ‘memories.’

• Append-only: A recording method that adds new information to the end without changing or deleting existing data, helping to prevent history tampering.

• Source: Agent Kernel – Three Markdown files that make any AI agent stateful