agent-kernel

A small markdown workspace that helps an agent accumulate durable context over time.

This repository is a public reference template for people who want simple, inspectable files instead of a heavy memory framework.

At a glance

The agent maintains the workspace. Humans can browse it, but they do not need to manually maintain it for it to be useful.

The kernel

The core pattern comes from agent-kernel (h/t oguzbilgic/agent-kernel). A few markdown files make an agent stateful:

• **AGENTS.md** — the agent reads this to know how to operate. Session protocol, memory structure, rules.
• **IDENTITY.md** — who the agent is, where it runs, what it does. Customized by the human.
• **KNOWLEDGE.md** — index of everything in the knowledge layer. The agent's table of contents for cold starts.

The agent has no built-in memory between sessions. The repo is how it becomes stateful — read to remember, write so the next session knows what happened.

Memory structure

Two kinds of memory, kept separate:

• State (knowledge/) — facts about how things are right now. Mutable. Updated when reality changes.
• Narrative (notes/) — what happened, what was tried, what's still open. Append-only. Never modify a past day's entry.

Raw and compiled knowledge

Inspired by Andrej Karpathy's LLM Wiki — the idea that LLMs should compile knowledge once into structured pages rather than re-derive it from raw sources each time.