March 23, 2026
Three files and a meme war
Show HN: Agent Kernel – Three Markdown files that make any AI agent stateful
3 text files give your AI a memory — fans cheer, skeptics yell “just prompts”
TLDR: Agent Kernel promises AI memory with three text files and a Git repo, no databases needed. The crowd is split between loving the simplicity and warning it won’t scale, with skeptics calling it “just prompts” and worrying that long diaries and old notes will make agents stubborn and unreliable.
Hacker News lit up over Agent Kernel, a tiny open-source trick that claims to give your AI a memory using just three text files and a Git folder. The pitch: clone this repo, start your favorite coding bot, and it will remember who it is, what it learned, and what it did last time. No fancy databases, no special tools — just plain markdown notes and a running diary. Think “personal journal for your robot.”
Cue the split reactions. The cheer squad loves the simplicity — one fan crowed that plain files beat “vector databases” (those special search engines for AI). But the skeptics came sharpened: will this scale when the diary becomes a novel? Users warn the “context window” — the amount an AI can read at once — will choke on months of logs. Others say agents cling to old notes like an ex, stuffing in outdated workarounds the moment a stray comment surfaces. The real burn: multiple posters waved this off as “just prompts” and doubted that agents reliably follow any instruction file once things get big.
Meanwhile, veterans shared they tried fancier setups (like small databases and embeddings) and still ended up just stuffing text into context. Verdict? It’s minimalism vs. machinery — a cozy notebook for your bot, or a future headache when your AI refuses to forget last season’s lore.
Key Points
- •Agent Kernel enables stateful AI agents using three Markdown files in a Git repository.
- •Setup involves cloning the GitHub repo and starting a compatible AI coding agent (e.g., OpenCode, Claude Code, Codex, Cursor, Windsurf).
- •Memory is split into knowledge/ (mutable state) and notes/ (append-only daily logs), tracked via IDENTITY.md and KNOWLEDGE.md.
- •The system relies on agents reading instruction files (e.g., AGENTS.md/CLAUDE.md/.cursorrules) to teach memory behavior.
- •Multiple agents are supported by using separate repositories with shared kernel but different identities and knowledge; licensed under MIT.