Chat with your notes and files. Watch your knowledge organize itself into a living galaxy. All on your machine. No cloud required.
Own Wiki is a desktop app that turns everything you know — your notes, your documents, your conversations — into a private, searchable, self-organizing knowledge base, powered by a local LLM.
It's inspired by Andrej Karpathy's "LLM wiki" idea and tools like Notion AI / Qyntra — but it runs entirely on your computer. Your data never leaves the machine. No subscription, no API key required, no telemetry.
Ask it a question and it answers from your knowledge, with citations. Drop in a PDF and it reads it. Give a research topic to the agent swarm and it investigates in parallel and writes you a cited brief. And the whole web of what you know renders as an interactive 3D galaxy you can fly through.
TL;DR — It's a ChatGPT-for-your-own-stuff that you actually own.
| 🧠 RAG chat over your vault | Streaming answers grounded in your notes + files, with inline [n] citations. |
| 🌌 Knowledge galaxy | A 3D force-graph of every note and how they connect — hover to highlight, click to fly in. |
| 🤖 Agent swarm | Spawn parallel research agents that search the web, ground their findings, and synthesize a cited wiki page. |
| 🗂️ Desktop auto-index | Reads your Desktop / Documents / Downloads automatically — ask about a file with zero manual import. |
| 💾 Long-term memory | Extracts atomic facts after each chat, so it remembers you across conversations. |
| ⚡ int8 vector store | Custom quantized embeddings — ~4× smaller on disk, <1% cosine error, zero native deps. |
| 🔌 Multi-provider | Ollama by default; optional fallback to Groq / Gemini / OpenRouter with a pasted key. |
| 🎨 Premium UI | Light + dark themes, animated background, resizable panes, command palette (⌘K). |
 Ask your vault — clean, fast chat home. |
 Galaxy — your knowledge as a living graph. |
 Wiki — auto-curated pages with backlinks. |
 Agents — parallel swarm research, live. |
Own Wiki is one Electron app that boots a Next.js server on 127.0.0.1 and talks to a local Ollama runtime. Everything — vault, vectors, memory — lives in a per-user folder under %APPDATA%.
flowchart TB
subgraph Desktop["🖥️ Electron Desktop App"]
UI["React UI<br/>Chat · Galaxy · Wiki · Files · Agents"]
MAIN["Electron main<br/>single-instance · free-port · per-user data"]
end
subgraph Server["⚙️ Next.js server (127.0.0.1:3789)"]
API["API routes<br/>/chat /swarm /ingest /desktop-scan ..."]
AGENTS["Agent registry<br/>browser · ingest · synthesize · desktop"]
RAG["RAG engine"]
end
subgraph Data["🗄️ Per-user data (%APPDATA%/Own Wiki)"]
VEC["Vector store<br/>int8 quantized"]
VAULT["Markdown vault<br/>pages + memory + log"]
end
subgraph AI["🧠 Local AI"]
OLLAMA["Ollama<br/>llama3.2:3b + nomic-embed-text"]
end
CLOUD["☁️ Optional fallback<br/>Groq · Gemini · OpenRouter"]
WEB["🌐 Web<br/>DuckDuckGo · Wikipedia"]
FS["📁 Desktop / Documents / Downloads"]
MAIN --> Server
UI <--> API
API --> AGENTS
API --> RAG
RAG --> VEC
RAG --> OLLAMA
AGENTS --> OLLAMA
AGENTS --> WEB
AGENTS --> VAULT
API -.fallback.-> CLOUD
VEC <--> VAULT
AGENTS --> FS --> VEC
sequenceDiagram
participant U as You
participant C as /api/chat
participant V as Vector store
participant O as Ollama
U->>C: question
C->>V: embed query, cosine search (top-10)
V-->>C: relevant chunks (notes, files, pages)
C->>C: + recent memory + conversation synopsis + app facts
C->>O: grounded prompt (streamed)
O-->>U: cited answer, token by token
flowchart LR
T["Topic"] --> A1["Agent: overview"]
T --> A2["Agent: techniques"]
T --> A3["Agent: applications"]
A1 --> S["Synthesizer"]
A2 --> S
A3 --> S
S --> P["Cited wiki page<br/>+ galaxy node"]
Each browser agent runs deterministic search → fetch → grounded-summarise (SSRF-guarded), so even a small local model produces accurate, cited output. The synthesizer always renders each agent's grounded answer, then layers on cross-references and key insights.
| Layer | Choice | Why |
|---|---|---|
| Shell | Electron | One installable .exe; native filesystem access for desktop indexing. |
| Front + back | Next.js 16 (App Router) + React 19 | One framework for UI and API routes; ships as a standalone server. |
| Inference | Ollama (llama3.2:3b) |
Free, local, private. No key, no cloud, works offline. |
| Embeddings | nomic-embed-text (768-dim) | Strong local embeddings for RAG. |
| Vector store | Custom, int8 quantized | ~4× smaller on disk, <1% cosine error, pure-JS (no native deps to break the portable build). |
| Graph | react-force-graph-3d + three.js | GPU-accelerated 3D galaxy with bloom. |
| State | Zustand | Tiny, fast, persisted. |
| Fallback AI | Groq / Gemini / OpenRouter | Optional one-paste upgrade to bigger models. |
- Download
OwnWiki-1.0.0-portable.zipfrom the latest release. - Unzip anywhere and double-click
Own Wiki.exe. - (Recommended) Install Ollama and pull the models for full local power:
ollama pull llama3.2:3b ollama pull nomic-embed-text
No Ollama? The app still opens — add a free Groq/Gemini key in Settings and it works via the cloud fallback.
git clone https://github.com/vaibhav4046/mnemosyne.git
cd mnemosyne
npm install
npm run dev # http://localhost:3000
# build the desktop app
npm run build:exe # -> dist-electron/💬 Chat — Ask anything about your notes and files. Toggle RAG (retrieve from your vault) and Memory (remember facts) per thread. Drop or paste a PDF/DOCX/image right into the composer to discuss it. Every thread is saved in the sidebar; export any thread to .md / .json.
🌌 Galaxy — Your whole vault as a 3D graph. Hover a node to light up its neighbours, click to fly in and focus, double-click to open the page. Filter by tag, search nodes, toggle glow / labels / auto-rotate.
📚 Wiki — Browse the auto-curated pages the AI writes from your sources. Backlinks, table of contents, reading time. Create or edit pages by hand too. Export to MD / PDF / CSV / DOCX.
🗂️ Files — Browse your Desktop / Documents / Downloads in-app. They're auto-indexed on launch, so chat can already answer about them — no manual import. Click any file to preview, or ingest a whole folder.
🤖 Agents — Type a topic and launch a swarm: parallel agents research different angles, then synthesize one cited wiki page that drops straight into your galaxy. Watch each agent's progress live.
🔌 MCP & Settings — Connect MCP tool servers (command-allowlisted). In Settings, pick your provider/model, paste optional cloud keys (masked, stored locally), switch theme.
Shortcuts:
⌘/Ctrl + Kcommand palette ·⌘/Ctrl + 1–7jump between views.
- Local-first. Vault, vectors, and memory live only in
%APPDATA%/Own Wiki. Nothing is uploaded unless you enable a cloud provider. - Per-user data. Whoever installs it sees only their own profile.
- Hardened APIs. SSRF guard on all outbound fetches, CSRF middleware, MCP command allowlist, path-traversal + realpath checks on file access, strict CSP.
- Keys stay secret. Cloud keys are masked in the UI, stored locally, never logged or sent to the browser.
- Real-time file-watching for instant re-index
- 2D/3D galaxy toggle + timeline view
- Per-vault encryption at rest
- macOS + Linux builds
src/
app/api/ chat · swarm · ingest · desktop-scan · reindex-pages · wiki · files · providers ...
lib/
vector.ts int8 quantized vector store
wiki.ts markdown vault engine
providers.ts multi-provider LLM layer
agents/ browser · ingest · synthesize · desktop · registry
components/ chat · graph (galaxy) · wiki · files · agents panels
electron/ main.cjs (desktop shell)
scripts/ brutal-qa · provider-qa · ui-walk · repack
Built by Vaibhav Lalwani · Powered by local AI · Made to be owned.
If this is useful, a ⭐ goes a long way.