---
title: Mycelium
emoji: 🍄
colorFrom: red
colorTo: indigo
sdk: gradio
sdk_version: 6.16.0
python_version: '3.13'
app_file: app.py
pinned: true
license: mit
short_description: Stop losing ideas. Capture, connect, recall.
tags:
  - build-small-hackathon
  - backyard-ai
  - tiny-titan
  - nvidia-nemotron
  - off-brand
  - off-the-grid
  - field-notes
  - track:backyard
  - sponsor:nvidia
  - achievement:offgrid
  - achievement:offbrand
  - achievement:fieldnotes
---

# Mycelium — Personal Knowledge Agent

> Capture fast. Think later. Let the system surface what matters.

Mycelium is a local-first AI knowledge companion that closes the loop between saving something and actually learning from it. No more screenshot graveyards or forgotten browser tabs.

## Demo

📹 **[Demo video](https://www.youtube.com/watch?v=Kr7LxRm0JBs)**

📝 **[LinkedIn post](https://www.linkedin.com/posts/ajit3259_mycelium-stop-losing-ideas-capture-connect-share-7472039183830810624-TtU_/)**

📓 **[Field Notes — what I built, what broke, what I learned](https://huggingface.co/blog/build-small-hackathon/mycelium)**

## The problem

Everyone has the same graveyard: saved links, screenshots, notes-to-self — all gone dark in a week. The capture habit exists. The recall loop doesn't.

Mycelium fixes the loop.

## What it does

- **Capture** notes, URLs, and images — each processed into a structured summary with intent classification (`learn` / `act` / `reference` / `ephemeral`) and semantic tags
- **ASK** — semantic search across your knowledge base with LLM synthesis, follow-up questions, and Feynman self-testing
- **BRIEF** — daily digest of what you saved, with synthesis across captures and a weekly thread
- **REVIEW** — spaced repetition (SM-2) targeting specific claims from your own notes, not generic flashcards
- **GRAPH** — visual map of how your ideas connect via embedding similarity

## How it works

1. You capture a note, URL, or image
2. **NVIDIA Nemotron-Mini-4B** extracts the core insight, classifies intent, generates tags and recall questions
3. **Qwen2.5-VL-7B** handles image captures — describe a whiteboard, diagram, or screenshot
4. **BGE-base-en-v1.5** embeds summaries into a 768-dim vector space
5. Related captures link automatically via cosine similarity
6. The surface engine resurfaces what you should revisit, weighted by intent and time

## Tech

- **LLM**: `nvidia/Nemotron-Mini-4B-Instruct` via HF Transformers + ZeroGPU
- **Vision**: `Qwen/Qwen2.5-VL-7B-Instruct` for image capture
- **Embeddings**: `BAAI/bge-base-en-v1.5` (768-dim, top MTEB retrieval)
- **Backend**: FastAPI + SQLite (persistent at `/data/mind.db`)
- **Frontend**: React + TypeScript + Tailwind CSS

## Prizes targeting

- **Backyard AI** track — practical tool solving a real daily problem
- **Tiny Titan** badge — Nemotron-Mini-4B drives all text intelligence (4B parameters)
- **NVIDIA** — built on `nvidia/Nemotron-Mini-4B-Instruct`

## Built with

Built for the [Build Small Hackathon](https://huggingface.co/build-small-hackathon) · June 2026