Mémoire des Anciens: The Grimoire of Souls

A computer science and AI student in Togo built Mémoire des Anciens, a local-first web app that records relatives telling oral memories, transcribes them with faster-whisper, and uses Gemma via Ollama to generate a written "page-souvenir" for each recording. The app stores memories as JSON and audio files on disk, exports everything as Markdown, and falls back to displaying the raw transcription when the model returns invalid JSON or Ollama is unavailable.

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine https://dev.to/challenges/hacktoberfest-week4-2026-10-26 Some stories disappear without any noise at all. Nobody decides to lose them. They simply stop being told, and one day they are gone. When those voices go quiet, the memories go with them. Mémoire des Anciens "Elders' Memory" , with the subtitle Le Grimoire des Âmes The Grimoire of Souls , is a small web app for keeping those voices. You record a relative telling a memory, either straight from your browser's microphone or by uploading an audio file mp3, wav, m4a or webm . The app transcribes what was said, then turns the transcription into a page-souvenir , a memory page with: All the memories are gathered in the Ofrenda , a gallery of every page-souvenir. The name comes from the altars families build for Día de los Muertos, where people place photos, food and favourite things to honour those who came before. This ofrenda is made of stories and voices instead of marigolds. You can also Exporter export everything as Markdown, so your family's memories are plain text files that you own. I am a computer science and AI student in Togo. On our African continent, and in countries like Togo, many of our stories disappear in very insignificant ways: quietly, without anyone noticing. My grandfather championed our culture and the old traditions. So I built Mémoire des Anciens, a grimoire where a voice can be written down and kept. The theme asks for a ghost in the machine. Mine isn't spooky. It is the gentle kind: a voice that stays in the machine, so it isn't lost. DEMO VIDEO: https://youtu.be/1TPFicEQ2Ls https://youtu.be/1TPFicEQ2Ls Recording a memory , from the microphone or an audio file: The page-souvenir that Gemma writes from the transcription: The gallery , where every memory is kept: Préservez les histoires orales de famille : enregistrez ou téléversez un proche qui raconte un souvenir l'app le transcrit puis génère une page-souvenir . 100 % local : faster-whisper + Gemma via Ollama aucun appel à un service cloud. Prérequis : Python 3.11+, Ollama https://ollama.com , ffmpeg conseillé. python -m venv .venv && source .venv/bin/activate Windows : .venv\Scripts\activate pip install -r requirements.txt ollama pull gemma3 Le modèle Whisper « small » ~480 Mo se télécharge une seule fois au premier usage ; ensuite tout fonctionne hors ligne. ollama serve si Ollama n'est pas déjà lancé uvicorn main:app --port 8000 puis ouvrir http://localhost:8000 Variables optionnelles : OLLAMA MODEL défaut gemma3 , OLLAMA URL , WHISPER MODEL défaut small . To run it yourself: pip install -r requirements.txt ollama pull gemma3 uvicorn main:app --port 8000 Then open http://localhost:8000 . Everything runs on the machine in front of you. small model. It is a download of about 480 MB, done once. After that it works fully offline. uvicorn main:app --port 8000 . static/index.html . ./data/<id .json , next to its audio file. You can back it up by copying a folder. The code is organised into small pieces: main.py FastAPI routes services/transcription.py faster-whisper services/analysis.py Gemma via Ollama services/storage.py JSON + audio on disk static/index.html the frontend tests/test app.py pytest, with the models mocked Models and endpoints are set with environment variables: OLLAMA MODEL , OLLAMA URL and WHISPER MODEL . Changing the model needs no code change. When the model fails. Language models don't always return clean, valid JSON. If Ollama is off, or Gemma answers with invalid JSON after one retry , the app doesn't throw the memory away. It keeps the raw transcription and shows it . A recording of someone's elder should never be lost because a model had a bad moment. Tests. The tests use pytest with the models mocked. They check the app logic without downloading Whisper or running Gemma. Limits. I want to be honest about two of them: localhost or over HTTPS. Browsers require that. Uploading a file works either way. For this project, the open pieces are what make it work. Family voices are private. An elder telling a story is not training data. With cloud transcription, the recording of their voice, their name and their family's names would all go to a server I don't control. Here the audio, the transcription and the memory pages never leave the machine. Open-weight models make that possible. It works offline. After the one-time Whisper download, nothing needs the internet. That matters wherever the connection is poor or absent. It also means the app works when a family gathers, and those are the moments when stories get told. It costs nothing to run. There is no API key, no subscription and no per-minute fee. A project about preserving memory shouldn't be something a family can't afford to keep using. No service can disappear on it. Memories are meant to last decades. A hosted product can shut down, change its pricing or change its terms. Here the memories are JSON, audio files and Markdown exports in a folder that belongs to the family. Whisper and Gemma weights can be kept on disk. You can swap the models. Because the model is an environment variable, it can be replaced with another Ollama model or a different Whisper size. And the Ewe and Kabiyè limitation is exactly where this openness matters. I can't fix it with a closed API, but with open models it can be improved. Whisper doesn't handle Ewe or Kabiyè, and these are languages spoken in Togo. The next step is to look at open models that handle them better, or to adapt an open model to them. That would be the true goal of this project: that the elders can speak in the language they remember in. Where open beat closed, concretely. As a student in Togo, a paid speech-to-text API billed per minute, plus an LLM API billed per token, is not something I could keep paying for, and neither could the families I want to record. The internet connection is not always there either. With faster-whisper and Gemma on my own laptop, I recorded, transcribed and generated memory pages with no account, no bill and no connection. The first memory I kept was told by Mama, my great-aunt: "Halloween au Toggo Toggo". Hearing her voice play back next to a page with her words, the people and places she named, and the quotes worth keeping, is exactly what I built this for. Her story now lives in a folder on my machine, not on someone else's server. Best Use of Gemma. Gemma is the heart of the project. It reads each transcription and turns it into the page-souvenir: the summary, the portrait, the people, places and dates, and the quotes worth keeping. It runs locally through Ollama, so the stories stay on the family's own computer. The model is configurable through OLLAMA MODEL , and the app degrades gracefully by keeping the raw transcription if the model is unavailable or returns invalid output. Thanks for reading. If you have an elder whose voice you want to keep, record them this week.