{"id":50353601,"title":"Show HN: I built Knemora – An Open-source AI Second brain","points":1,"user":"flaterra","time":1793721486,"time_ago":"4 hours ago","type":"link","content":"

Hi HN, my name is Federico. I’m an Italian graphic designer and full stack developer.

In recent months, there has been a lot of talk about the "second brain" articles and YouTube videos; everyone is still discussing it, but I see a significant difference between that and what the concept of a second brain means to me.

For me, the concept of a "second brain" is about building a 100% private knowledge base which means not uploading my own files (containing sensitive data) to be processed via cloud-based inference.

I’ve seen many videos discussing how to "build your own second brain" using Claude Code and Obsidian. I’ve always wondered: why should I upload my personal files such as certificates, my children’s birth certificates, or other sensitive personal data—to a service like Claude, where I have no idea where those files end up or who might be reading them? Furthermore, why should I rely on a graph-based viewer like Obsidian, effectively making myself dependent on third parties, paid services, and software that I have to install directly onto my own device?

From that came the idea to develop Knemora, an open-source, local-first project that runs locally by leveraging local inference.

The challenge involved:\n- choosing a modular architecture\n- ensuring ease of use via a CLI and a web interface\n- avoiding the need for users to install third-party software on their own devices\n- enabling setup in just 2 CLI commands\n- ensuring suitability for both developers (tech-savvy users) and various private organisations (such as notary offices, dental practices, insurance companies, etc.).

So far knemora includes the following features:\n- local inference via llama.cpp (with built-in support for cross-platform operation)\n- ability to use "knemora-tested" models or custom models from Hugging Face; simply run the command `./knemora load --model hf:<Hugging Face model name>`\n- a chatbot agent that answers questions and searches through your knowledge base based on your natural-language queries\n- a knowledge agent built using Google’s ADK 2.0 (my agentic orchestrator)\n- portable knowledge, markdown files are linked following the "LLM WIKI" pattern (as proposed by Andrej Karpathy)\n- BM25-based retrieval (no vector database required)\n- original documents are preserved and marked as (immutable)\n- dataset analysis powered by Polars

Currently, it is available only for macOS with Apple Silicon chips, but it will soon be available for Windows and Linux as well.

Version v0.2.0, available today on GitHub, covers these additional features and improvements:

- conversation continuity\n- automatic language matching\n- gemma 4 joins the tested model registry\n- more curated models ahead\n- more control over chat\n- wiki / direct chat modes\n- fast / deep chat modes\n- stronger knowledge compilation

My goal is to make Knemora more stable, responsive, easier to use and keep moving toward a stable v1.0.0.

Thank you!

Watch the alpha demo:\nhttps://www.youtube.com/watch?v=kl5qNrWPOnY

Explore the project:\nhttps://knemora.space

Code and documentation:\nhttps://github.com/federicolaterra/knemora","url":"https://github.com/federicolaterra/knemora","domain":"github.com","comments":[],"comments_count":0}