kenkeep

A team-shared, git-native knowledge base for AI coding sessions.
Built up in your repo, reviewed and versioned like code, with no extra infrastructure to run.

How it works  ·  Installation  ·  Daily use  ·  Knowledge packs  ·  Troubleshooting

---

Coding assistants forget everything between sessions. Kenkeep salvages the gold nuggets from your past conversations and discards the rest, so the detail you explained two weeks ago is there the next time the assistant needs it.

Kenkeep is a team-shared, git-native knowledge base for AI coding assistants. Conventions, gotchas, module names, and the reasons behind decisions get captured from your sessions, curated by a human, committed to the repo, and injected back into every future session.

Overview

Why kenkeep

Built up and shared across your team One markdown file per fact, accumulated from real coding sessions and stored in your repo. It travels with git pull, so every teammate works from the same conventions instead of rediscovering them alone. The nodes/ tree is an Open Knowledge Format bundle any OKF tool can read.	Reviewed and versioned like code Nothing reaches the knowledge base without a human approving it. Every addition is an ordinary git diff you review in a commit or PR, with the full history there to blame or revert.
No extra infrastructure No daemons, services, databases, or vector stores. Kenkeep is Node and git. Nothing to provision, host, keep alive, or secure.	No API keys It runs inside the assistant you already pay for: Claude Code, Codex, Cursor, OpenCode, or Copilot. There is no separate key to obtain, store, or rotate.

How it works

• Capture is automatic. When a session ends, a hook saves the transcript.
• Curate is yours to start. Run /kk-curate and the assistant drafts one note per durable fact, then walks you through any contradiction with a note you already have.
• Review is yours to decide. Read the notes with git diff and commit the ones you want.
• Recall is automatic. Every new session starts with the root catalog and descends only into the notes the task needs, so the payload stays small as the base grows.

A note curate can't place still doesn't get stuck: it lands at the nodes/ root, and the next npx kenkeep init --upgrade files it from its own edges and tags, or removes it when it matches nothing in the tree. npx kenkeep node sweep does the same on demand between upgrades. Both leave the result uncommitted, so you accept it with git commit and reject it with git restore.

Full walkthrough: How it works.

Quick start

npx kenkeep init --harnesses claude
npx kenkeep doctor

Swap claude for codex, cursor, opencode, or copilot, or pass a comma-separated list. Per-harness details, including where GitHub Copilot CLI keeps its hooks and skills, are in Installation.

Then code as usual. When the assistant nudges you, run /kk-curate in your session. New notes land under .ai/kenkeep/nodes/. Review them with git diff and commit the ones you want to keep.

Seed from existing docs If your repo already has READMEs, ADRs, or module docs, seed the knowledge base from them. Inside a session: /kk-bootstrap The scan starts at the repo root and honors .kkignore, which init creates with gitignore syntax. Review the resulting notes with git diff and commit the ones you want.

Add knowledge manually Mid-session, mention /kk-add and the assistant records the point you just made: No, you got that wrong. This project aims to maximize code re-use, instead of duplication. Adapt and extend the abstractions to fit this use case. Also, /kk-add this.

Knowledge packs

A pack is a reviewed nodes/ tree published for a framework, platform, or shared domain. Import one and it lands as a single isolated branch in your own knowledge base:

npx kenkeep pack import e0ipso/kenkeep-pack-drupal
npx kenkeep pack import https://github.com/e0ipso/kenkeep-pack-drupal --as drupal

Import is deterministic and never calls an LLM. Colliding note ids are skipped with a warning. Full guide: Knowledge packs.

Documentation

Full documentation: https://kenkeep.canpicasoft.com

Working on the package itself? See CONTRIBUTING.md.

License

MIT