---
title: Pensieve
emoji: 🎙️
colorFrom: gray
colorTo: red
sdk: gradio
sdk_version: 6.17.3
app_file: app.py
pinned: false
license: mit
short_description: Speak a thought, get a markdown note, then chat with them
tags:
  - gradio
  - build-small-hackathon
  - track:backyard
  - sponsor:modal
  - sponsor:cohere
  - achievement:offbrand
  - best-demo
models:
  - Qwen/Qwen3-8B
  - Qwen/Qwen3-Embedding-0.6B
  - CohereLabs/cohere-transcribe-03-2026
---

# Pensieve

Speak a thought and Pensieve turns it into a clean markdown note in the background. Browse your
growing collection of notes and ask questions across everything you have captured.

Capture is asynchronous: stop recording and the transcribe, summarise and index pipeline runs as a
background job, so you can record the next thought right away.

I built this for my dad, who is always going on walks and recording voice notes of his thoughts.
Pensieve allows him to build a catalogue of his thoughts and recall them easier.

## Demo

- Video: https://huggingface.co/spaces/build-small-hackathon/pensieve/resolve/main/pensieve-demo.mp4
- Social post: https://www.linkedin.com/feed/update/urn:li:activity:7472430389475610624/

## How it works

The front end is a dark, minimalist Gradio app with a bottom tab bar that installs to the home
screen on an iphone as a progessive web app (PWA). All AI inference runs on Modal, and every model is under 32B parameters.

- Record: capture audio, then a background job runs transcribe, summarise and index.
- Jobs: a live view of each pipeline job and its stage.
- Knowledge: a Chat and Notes view. Chat answers with RAG over your notes and cites sources. Notes
  lets you search and read your captured notes.

## Models (all < 32B)

| Role       | Model                                | Runs on       |
|------------|--------------------------------------|---------------|
| ASR        | CohereLabs/cohere-transcribe-03-2026 | Modal, L4 GPU |
| Embeddings | Qwen/Qwen3-Embedding-0.6B            | Modal, CPU    |
| LLM        | Qwen/Qwen3-8B                        | Modal, L4 GPU |


## Future work

First order of business would be to move the data into some user owned data storage (like google drive).
right now its on a dataset repo, but its not private as I have access, although each user can't see eachothers data.

I could would then speed up inference by using per-token cost API's instead of having to cold-start GPUs.
I am currently using memory snapshots and that does seem to speed things up a lot.  

Improve RAG and prompting, currently using hybrid RAG with reciprocal rank fusion (RRF), but a re-ranker couldn't hurt.