Project knowledge needs somewhere to persist between conversations. Instructions, client memory, skills, and KaKeKiKoKu serve different purposes and can work together. The useful question is which information belongs in each place, and how you will keep it accurate. KaKeKiKoKu is not another memory for one AI: it holds the project’s own knowledge, which you shape with any AI and which stays with the project.
| CLAUDE.md / AGENTS.md | Built-in AI memory | Skills | KaKeKiKoKu Studio | |
|---|---|---|---|---|
| Where it lives | In repository or user instruction files. | In storage chosen by the client. | In a user or project skill folder. | Inside the project, beside the resources it describes. |
| Structure | Can be hierarchical and scoped to folders. | Varies; some clients use an editable index and topic files. | Instructions with optional scripts and references. | Resources with UUIDs, bounded sections, and explicit relations between blocks. |
| Who updates it | You or an AI you ask to edit it. | The client or AI; human controls vary. | You or an AI maintaining the procedure. | You and your AI, with a workflow that asks for knowledge capture. |
| When files change | Review depends on your workflow. | Update and retrieval rules depend on the client. | Review the procedure when its assumptions change. | Detects out-of-date context and structural issues; meaning still needs review. |
| Portability | Plain files; discovery rules depend on the client. | Storage and export depend on the client. | Repository skills can travel with the project; execution support varies. | Wuniq files travel with the project and are served through local HTTP or MCP. |
| Context loading | Scope, hierarchy, and size limits depend on the client. | May combine startup summaries with on-demand detail. | Usually discovered and loaded when relevant. | Root context and map first; relevant resources and blocks on demand. |
| Main purpose | Stable working instructions and conventions. | Remember useful information across conversations. | Reusable procedures and tool workflows. | Decisions, constraints, intent, risks, and relationships tied to project resources. |
| Human review | Read and edit the instruction files. | Client controls or editable files, where supported. | Inspect the skill and its supporting files. | Local editor, resource navigation, relations, and sync issues. |
Keep stable working rules here, including the instruction to consult KaKeKiKoKu. They need not be one flat document: Codex combines instructions along the directory hierarchy. For a small project, well-maintained instruction files and ordinary documentation may be enough.
Check your client's actual implementation. For example, Claude Code's memory includes editable project memory files, an index, and topics read on demand. It can retain project facts as well as personal preferences. Decide whether its storage, review controls, and portability fit your work.
Use skills for reusable procedures such as reviews, deployment, or document generation. A skill stored in the repository can travel with it. It can consult KaKeKiKoKu for the project's decisions instead of embedding copies of those decisions in the procedure.
An Obsidian vault can hold project knowledge and can live in a repository. Its official CLI exposes links, backlinks, and Bases queries with structured output. It is a viable choice when your workflow centers on notes and you are comfortable defining how an agent reads and maintains them.
KaKeKiKoKu provides a specific convention for attaching context to real files and folders, plus standalone entities, bounded sections, explicit relations, and synchronization checks. Compare that workflow with the one you actually use in your vault, including any plugins or scripts. The distinction is the integrated maintenance model, not whether Markdown notes can be queried or linked.
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.