Private, self-hosted audio memory — recorded anywhere, processed by providers you choose, and recalled naturally.
NeoRecall runs as a service on your own machine. It receives audio from web, desktop, and wearable clients, sends it to a separately deployed or hosted transcription service, builds searchable local memory, and uses the external language-model provider you configure to write memories.
Important
npm registry distribution has been shut down. Install NeoRecall directly from GitHub using the repository installer below. The npm command is still used locally to install Node.js dependencies and link the CLI to the checkout.
The repository installer clones NeoRecall directly from GitHub (requires Git, Node.js 20+, and npm for dependencies):
bash <(curl -fsSL https://raw.githubusercontent.com/NeoLabs-Systems/NeoRecall/main/install.sh)Open http://localhost:4500 once neorecall status reports a
running service. The macOS and Windows apps can do all of this without a
terminal: Set up NeoRecall on this computer installs the server, then asks
for the transcription and language-model services and tests both.
- Audio stays under your control. Clients retain every chunk until the server proves its transcript is durable and its temporary audio copy is gone.
- Inference stays replaceable. NeoRecall bundles no speech-recognition or language model. Use a hosted provider or a compatible service you deploy elsewhere; provider and model choices are configurable in the desktop app,
.env, and the admin dashboard. - Voices are recognized on your machine. A transcription service returns words, never who said them, so NeoRecall detects speech and tells speakers apart itself — 31 MB of models, not gigabytes. Silence is never sent anywhere, and the same voice is still recognized weeks later.
- Long recordings remain bounded. Memories are written as soon as a conversation ends, and a transcript longer than the configured external model context is read in windows and folded back into one memory.
- Built for recorders that never stop. Audio uploads, transcription, and conversation detection all run during capture, a conversation that is still recording gets a provisional title and summary you can read before it ends, and one real-world occasion still becomes exactly one memory.
- Offline first. Browser and desktop clients buffer independently decodable chunks and resume idempotent uploads when the server returns.
- Native NeoAgent and MCP recall. Connect from NeoAgent or paste the MCP URL into Claude, ChatGPT, or Cursor to search memories and transcript evidence through the same read-only retrieval tools.
Read the documentation here.
Use GitHub Discussions for questions and GitHub Issues for reproducible bugs. Security reports belong in the process described by SECURITY.md.
NeoRecall is licensed under the GNU Affero General Public License v3.0 only.
Made with ❤️ by Neo · NeoLabs Systems