A personal knowledge base that compiles your notes, journals, and messages into a structured, interlinked wiki — maintained entirely by an LLM agent. Inspired by Farzaa and Karpathy.
data/ ← drop your source files here
↓
ingest ← agent converts them to raw entries
↓
raw/entries/ ← one .md file per diary entry / note / message
↓
absorb ← agent reads entries and compiles wiki articles
↓
wiki/ ← your personal Wikipedia (agent owns this)
↓
wiki-ui/ ← Next.js app that renders it in a browser
The agent writes and maintains all wiki articles. You source data, ask questions, and explore.
Open this project in whichever AI tool you use — the right config file is already there:
| Tool | Config file | How to invoke operations |
|---|---|---|
| Cursor (agent mode) | .cursor/rules/wiki.mdc |
Chat: "absorb all my entries" |
| Gemini CLI | GEMINI.md |
Chat: "ingest my data" |
| OpenAI Codex | AGENTS.md |
Chat: "absorb last 30 days" |
| Claude Code | CLAUDE.md + .claude/skills/wiki/SKILL.md |
Chat: "absorb all" |
| Any other agent | AGENTS.md |
Chat: describe what you want |
No setup required — just open the project and start talking to your agent.
personal-wiki/
├── AGENTS.md # Instructions for Codex + any agent
├── CLAUDE.md # Instructions for Claude Code
├── GEMINI.md # Instructions for Gemini CLI
├── .claude/
│ └── skills/wiki/SKILL.md # Full canonical skill definition
├── .cursor/
│ └── rules/wiki.mdc # Cursor agent rules
├── data/ # Drop your source data here
├── raw/
│ └── entries/ # Auto-generated by the ingest operation
├── wiki/ # Agent-compiled knowledge base
│ ├── _index.md # Master article index
│ ├── _backlinks.json # Reverse link index
│ ├── _absorb_log.json # Tracks which entries are absorbed
│ └── (categories emerge here)
├── wiki-ui/ # Wikipedia-clone UI (Next.js)
├── Dockerfile # Production Docker build
└── .dockerignore # Files excluded from Docker builds
Put any of the following into data/:
| Format | How to export |
|---|---|
| Day One | File → Export → JSON |
| Apple Notes | Export folder as HTML or text |
| Obsidian | Copy your vault folder |
| Notion | Export as Markdown & CSV |
| iMessage | Use an app like iExporter → CSV |
| Plain text / markdown | Drop files directly |
| Images | Put them alongside notes or in data/assets/ |
Open the project in your AI tool and use natural language. All agents understand the same operations:
Ingest your data
"Ingest my data" / "Process the files in data/"
Converts everything in data/ into individual markdown entries in raw/entries/. The agent writes an ingest.py script to do this. Safe to re-run.
Compile the wiki
"Absorb all entries" / "Build the wiki" / "Compile everything" "Absorb entries from last 30 days" / "Absorb 2025"
Reads every entry and builds wiki articles. This is the core step. The agent synthesizes, connects, and writes — not just files, but understanding.
Ask questions
"What are the recurring themes in my work?" "Tell me about my relationship with [person]" "What's my philosophy on product design?" "What inspired me most in 2025?" "What was happening in my life around March 2024?"
Maintain and expand
"Clean up the wiki" — audit articles, fix structure, repair broken links "Find missing articles" — identify entities mentioned but not covered "Rebuild the index" — rebuild
_index.mdand_backlinks.json"Give me wiki stats" — article count, coverage, orphan articles
cd wiki-ui
npm run devOpen http://localhost:3000. You get:
- Main page — categories and recently updated articles
- Article pages — Wikipedia-style layout with table of contents, backlinks, related articles
- Category pages — all articles by type (people, projects, philosophies, etc.)
- All articles — A–Z index
- Search — full-text search across all articles
- Stats — most linked, longest, orphan articles
The UI reads directly from wiki/ — refresh the page to see new articles as the agent writes them.
Use Obsidian alongside your agent. Open this repo in Obsidian to browse the wiki with graph view, rendered markdown, and the Dataview plugin for dynamic queries.
Useful Obsidian plugins:
- Obsidian Web Clipper — clip web articles to markdown directly into
data/ - Marp — render wiki articles as slide decks
- Dataview — query frontmatter across all articles
File outputs back into the wiki. When you get a great answer from a query, ask the agent to save it as a new article. Your explorations compound.
Run cleanup periodically. "Clean up the wiki" uses parallel subagents to audit every article and fix structure, broken links, and stubs.
Images work. The agent can reference images from data/assets/ in wiki articles. Download images locally with the Obsidian hotkey (Settings → Hotkeys → "Download attachments for current file").
Every article in wiki/ follows this format:
---
title: Article Title
type: person | project | philosophy | pattern | era | decision | ...
created: YYYY-MM-DD
last_updated: YYYY-MM-DD
related: ["[[Other Article]]", "[[Another]]"]
sources: ["entry-id-1", "entry-id-2"]
---
# Article Title
Content organized by theme, not chronology.
## Section
[[Wikilinks]] to other articles.The agent uses [[wikilinks]] to link articles. The UI resolves them to real links.
Based on the pattern from Karpathy and Farzaa:
| Layer | What it is |
|---|---|
Raw sources (data/) |
Immutable. Agent reads, never modifies. |
Entries (raw/entries/) |
One file per logical entry. Generated by ingest. |
Wiki (wiki/) |
Agent-compiled knowledge base. Agent owns this. |
Index (wiki/_index.md) |
Catalog of all articles. Auto-maintained. |
Backlinks (wiki/_backlinks.json) |
Reverse link map. Auto-rebuilt. |
UI (wiki-ui/) |
Read-only browser for the wiki. |
The key insight: the wiki is a persistent, compounding artifact. Cross-references already exist. Contradictions are flagged. Synthesis reflects everything you've read. It gets richer with every source and every question.
The wiki-ui ships with a multi-stage Dockerfile that produces a minimal production image. It works with any Docker-based hosting platform (Dokploy, Coolify, Portainer, bare Docker).
docker build -t personal-wiki .
docker run -p 3000:3000 personal-wiki- Push your repo to GitHub (or any Git host).
- In the Dokploy dashboard, create a Project and add an Application.
- Set the source to Git and point it at your repository and branch.
- Build type: Dockerfile (auto-detected from the root
Dockerfile). - Under Domains, add your domain — Dokploy provisions TLS automatically via Traefik.
- Set the container port to 3000.
- Deploy.
Wiki articles are baked into the Docker image at build time. To update after adding new articles:
- Re-deploy from Dokploy — it rebuilds with the latest
wiki/content from your repo. - Volume mount — mount a host directory to
/app/wikiin the container for live updates without rebuilding. Upload new content via rsync or scp.