Chat with your documents
Chat with your local documents. Offline, if you want.
Point a vault at a folder. Your PDFs, Word files, notes and spreadsheets enter memory on their own. Ask a question, and the answer cites the file it came from. With a local model, nothing leaves your machine.
The problem
A thousand documents you own, and no way to ask them anything.
Contracts, reports, papers, manuals, the notes of ten years. Search finds a file name. It does not answer "what did we agree on in the March version", and it does not read the PDF for you.
Mnemosyne OS reads them into a memory on your computer. You ask in plain language. The answer is written from the passages it found, and it names the files, so you open the right page instead of trusting a paragraph.
How an answer is built
Found first, written second. Never the other way round.
Your question is compared to every passage in the vault, by meaning and by the exact words. The best passages go to the model with your question. The model writes from them. It does not answer from general knowledge. It answers from your own notes, your documents, your files. Developers call this RAG, retrieval-augmented generation; here it runs offline.
When nothing matches, the answer says so. A gap is a gap, not an invitation to invent.
- What does the 2024 contract say about the notice period?
- Summarise the three reports on the Lyon site, and where they disagree.
- Which invoices mention the reference AX-2291?
- What did I note after the meeting with the architect?
Offline
Unplug the network. It still answers.
Reading, chunking, indexing and search all run on your processor. The model that writes the answer can too: pick a local model, download it once, and the whole loop runs with the cable out. Nothing is sent, because there is nothing to send it to.
Prefer a stronger cloud model? Use your own API key, or prepaid credits. Then only your question and the passages found travel, to the provider you chose. You decide per conversation.
Free, no account
The complete OS, for free. No subscription.
You download, you launch. Nobody asks your name or your email. Everything on this page runs without a license and without credits, with a local model.
How it works
Five steps, once. Then it just runs.
- 01 Create a vault "+ New Vault". One per domain of your life, a file on your disk.
- 02 Give it a folder "Add a folder to watch…". Whatever lands there enters memory, and again whenever it changes.
- 03 Let it index Each document is cut into passages, each passage gets a vector. Locally.
- 04 Choose the model Local for offline, or your own key. You can change your mind per conversation.
- 05 Ask The answer cites its files. Open the page it points at.
The detailed walkthrough lives in the documentation: Your first vault ↗ ·Watched folders ↗
Questions
What people ask before installing.
- Can I chat with my PDFs offline?
- Yes. Reading, indexing and search run on your machine. Pick a local model and download it once; from then on the whole loop works with the network unplugged. A cloud model is a choice you make per conversation, never a requirement.
- Which file formats does it read?
- PDF, Word (.docx), Markdown, plain text, LaTeX, spreadsheets (.xlsx, .csv), JSON and YAML. Images (jpg, png, webp) once image memory is turned on. PowerPoint is not read: export the deck to PDF.
- Does it upload my documents anywhere?
- No. Your files stay in your folders and the index is a file on your disk. With a local model nothing leaves the machine. With a cloud model, only your question and the passages that answer it are sent, to the provider you chose.
- How is this different from uploading files to a chatbot?
- Nothing is uploaded, and nothing is forgotten at the end of the session. A vault is permanent memory on your disk: you add a folder once, every change is picked up, and every answer cites the file it drew from.
- How many documents can a vault hold?
- A vault is a database file on your disk; the practical limit is your disk and the time to index. A long document enters whole: past roughly three hundred pages the remainder is announced as unread rather than silently cut.
- Does it work with scanned PDFs?
- A scan has no text layer, so it needs an OCR engine. A free one installs from Settings → OCR. Until then the scan waits, and the app says so.
- Can I use my own model?
- Yes: a local GGUF model on your machine, your own API key with a provider, or your own server. The memory does not move when you change the model.
- Is it free?
- The download is the complete OS, free, with no account and no email. Everything on this page runs without a license and without credits, with a local model. Cloud credits are prepaid and optional.