The gap
Cloud assistants guess, forget, and send your life somewhere else. A local chat window is not the same as a brain that already knows your disk.
AI
Your AI. Your machine.
A personal, self-learning assistant in one binary. No database server, no container runtime, no cloud account. It remembers what it learns about you, and what it learns stays here.
- Go
- SQLite
- Ollama
- WebKitGTK
- whisper.cpp
This page is the overview. You get the files on GitHub or the store. Hosting plans are the only thing we sell here.
Where does that project live?
A real path on this disk, recalled from memory. Not a guess.
Reads run on their own. Anything that changes something waits for you.
Why it exists
What you take
One Go binary and a SQLite file. It recalls real paths, learns with your approval, and never forwards the memory it assembled from your machine.
What changes
Download one binary, point it at Ollama, and talk. Memory stays on this machine in every privacy mode.
PN Brain is not a model runner and not a ChatGPT clone. It is a personal assistant that lives on your disk, remembers with permission, and keeps what it learned here.
Who it is for
Use it when
- People who want a local assistant that already knows their projects, instead of a blank chat.
- Operators who will not send assembled personal memory to a hosted model.
- Homes that want spoken answers and optional Home Assistant control without a cloud account.
Built for
- Developers who already run Ollama.
- Anyone who wants a private assistant on a Linux desktop, with optional speech.
- People who will move the brain to another disk or machine and expect the memory to come with it.
Privacy is enforced in code
What the brain has learned about you never leaves, in any mode. Not in open, not with any setting. An unrecognised privacy value is treated as private, because the safe reading of a typo is the strict one.
private Default
Nothing. Local model, local embeddings, no web.
research
Search queries only. Conversations and memory stay here.
open
What you type may go to a third-party model. Memory still stays.
What it actually does
Answers with memory
It recalls what it already knows about your projects and documents. "Where does that project live" gets a real path rather than a guess.
Learns on its own
It scans your code and documents, and proposes things worth remembering after conversations. Anything it inferred waits for you to approve. Anything it can verify against the disk promotes itself.
Acts, with permission
It can read files, list directories, search and read the web, and control a smart home. Anything that changes something stops and asks first, showing exactly what will happen.
Draws what it knows
The interface renders a live map of memory, built from real embedding similarity. It is a picture of what is stored, not decoration.
Listens and speaks
Conversation mode hears you until you stop talking, answers, reads the answer aloud, and listens again. Both halves are local. Voice can ask but cannot act. Spoken turns are sent without tools, so a misheard sentence has nothing to reach.
Speed, honestly
On a CPU with no GPU this is slow, and the shape of the slowness is worth knowing. Prompt size is very nearly the whole reply time.
Prompt
~10 tok/s
Four cores, 7B model
Reply
~2.5 tok/s
Generation on CPU
Spoken turn
~40 s
Listen, answer, speak
Resident reply
6.8 s
15.8 s after an embed eviction
A GPU, or open mode with a hosted model, are the only things that change the order of magnitude. In open mode, what the brain has learned about you still never leaves. Keep OLLAMA_MAX_LOADED_MODELS at 2, or every reply reloads the chat model after recall embeds your question.
Specs
- Pricing
- Open Source
- License
- MIT
- Deployment
- Self-hosted
- Maturity
- General Availability
- Support
- Community
Stack
- Go
- SQLite
- Ollama
- WebKitGTK
- whisper.cpp
Run it on this machine
Download the AppImage, or build the Go binary. Point it at Ollama with a chat model and nomic-embed-text. pn-brain-doctor reports what is missing instead of refusing to start.
This page is the overview. You get the files on GitHub or the store. Hosting plans are the only thing we sell here.
Questions
In private mode, nothing. Research sends search queries only. Open may send what you type to a third-party model. What the brain has learned about you never leaves, in any mode. There is no setting that changes this.
Ollama with a chat model and nomic-embed-text. The native window needs libwebkit2gtk-4.1 and libgtk-3 on Linux. Speech is optional: speech-dispatcher, espeak-ng, or piper to speak, whisper.cpp to listen. Run pn-brain-doctor to see what is missing.
On a CPU with no GPU, yes. Measured on four cores with a 7B model: about ten tokens a second for the prompt and two and a half for the answer. A spoken turn takes around 40 seconds. A GPU, or open mode, changes the order of magnitude. Memory still stays here.
No. Voice can ask but cannot act. Spoken turns are sent without tools. Anything that changes something is typed, and still stops for approval.
The brain looks for a .brain-root.json marker across mounted drives, so the data can live on an external disk. With no marker, it creates one under ~/.local/share/pn-brain. Settings live beside the data in brain.conf.
Open the GitHub repository from this page. This site is the overview. It does not take payment for products. The licence is MIT.