Every AI workflow eventually hits the same wall: the model is brilliant, and it remembers nothing about your project. Four approaches compete to fix that. Each is good at its job — and only one of them was designed to be the place where project knowledge lives. Here is an honest comparison.
| CLAUDE.md / AGENTS.md | Built-in AI memory | Skills | KaKeKiKoKu Studio | |
|---|---|---|---|---|
| Where it lives | In your repo — the right instinct. | In your machine and your AI tool. | With the skill, per user or per repo. | In your repo, beside the files it describes. |
| Structure | One flat document. | Invisible — you can’t inspect or organize it. | Procedure files with instructions. | Bounded blocks with identity, connected by explicit relations. |
| Who updates it | You, by hand. Which means: rarely. | The tool, silently, by its own criteria. | You, when you edit the skill. | The AI, as it works — KaKeKiKoKu makes capture part of the job. |
| When the project changes | Goes stale in silence. Nobody verifies it. | Doesn’t notice — it remembers conversations, not files. | Not its job — skills describe procedures, not state. | Drift is detected in real time; the AI is told to reconcile. |
| Portability | In the repo, but fragmented per vendor: CLAUDE.md here, AGENTS.md there. | The least portable: doesn’t travel with the repo, doesn’t transfer to another AI. | Travels with the skill, not with the project. | One knowledge base in the repo, served to any MCP client. |
| Token behavior | Injected whole, every session — all or nothing. | Opaque — you don’t control what gets loaded. | Loaded when triggered — efficient, for procedures. | Map first, detail on demand — progressive by design. |
| What it captures | Whatever you remembered to write down. | Whatever the tool decided to keep. | How to do things. | What would otherwise be lost: decisions, constraints, rejected alternatives, traps — captured as the work happens. |
| Human oversight | Whatever your text editor shows you. | A settings screen, at best. | Text files in a folder. | A dedicated local app: review every block, follow the relation graph, watch sync health live. |
If your project’s whole truth fits comfortably in one page — stable conventions, small scope — a context file is genuinely enough. And KaKeKiKoKu coexists with it happily: keep your startup rules there, let KaKeKiKoKu hold the knowledge.
It’s about you, not your project: preferences, tone, recurring instructions. Useful — and it was never meant to hold a project’s decision history. The trouble starts when it’s asked to.
The right tool for reusable procedures: deployment routines, review checklists, generation patterns. A skill tells your AI how to do something. It was never meant to know what is true about this project. The two compose beautifully.
A fair question — wiring a coding agent to an Obsidian vault became a popular pattern in 2026, and Obsidian’s official CLI and agent skills made it genuinely easier. We admire Obsidian: local-first, plain files, no lock-in — the same values KaKeKiKoKu is built on. The difference is the center of gravity. A vault is your knowledge base: notes, ideas, a second brain that lives apart from any repo. KaKeKiKoKu is the project’s knowledge base: it lives in the repo, beside the files it describes, and travels with them.
That difference shows up the day the honeymoon ends. Markdown notes cannot be queried structurally; wikilinks draw a beautiful graph but are not machine-navigable relations; nothing bounds what gets loaded into the context window; and nothing checks the notes against the code they describe. KaKeKiKoKu keeps what makes a vault lovable — plain, readable, git-friendly text files — and adds the engine a vault lacks: a live index, bounded blocks with identity, explicit relations, progressive reading, and drift detection. Keep your vault for your life. Give your projects their own memory.
When don’t you need KaKeKiKoKu? Throwaway scripts. One-off experiments. A project whose whole truth fits in one page and barely changes. KaKeKiKoKu earns its keep when knowledge accumulates faster than you can re-explain it — the moment you catch yourself telling your AI the same decision for the third time, come back.