This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine
EchoVault: Some Memories Refuse to Disappear
What if the voices of yesterday could still answer the questions of tomorrow?
A voice recording is more than a collection of sounds. It can hold a grandparent's stories, a forgotten family tradition, the name of a place, or a moment that deserves to be remembered.
But memories are fragile.
Recordings get buried in folders. Important details become difficult to find. Stories that once brought people together can slowly fade into silence.
I started thinking about one possibility:
What if we could preserve these voices, organize their stories, and make them searchable without allowing AI to rewrite what was actually said?
That question became EchoVault.
What I Built
EchoVault is a local-first AI memory vault that transforms family voice recordings into searchable, structured memories.
Instead of simply storing audio, it uses AI to identify meaningful details such as people, places, traditions and important moments. It keeps the original transcript at the centre of the experience so that users can revisit the words behind every preserved memory.
My interpretation of the theme Ghost in the Machine is about keeping something deeply human alive through technology.
The ghost is the echo of a memory that refuses to disappear.
The journey of a memory
A recording enters EchoVault as a voice. It passes through transcription, memory extraction and evidence verification before becoming a searchable story.
VOICE RECORDING
|
▼
FASTER-WHISPER
TRANSCRIPTION
|
▼
TRANSCRIPT REVIEW
|
▼
QWEN 2.5
MEMORY EXTRACTION
|
┌────────────┼────────────┐
▼ ▼ ▼
PEOPLE PLACES TRADITIONS
|
▼
MOMENTS
|
▼
GHOSTLIGHT
SOURCE EVIDENCE
|
▼
ASK THIS MEMORY
|
▼
MEMORY VAULT
Four dimensions of every story
EchoVault organizes extracted details into four memory categories.
MEMORY CATEGORY FEATURE MAP
People ████████████████████
Places ████████████████████
Traditions ████████████████████
Moments ████████████████████
This is a visual representation of the four extraction categories, not a performance benchmark.
The features that make EchoVault meaningful
*1. Voice to Memory *
Users upload a voice recording. faster-whisper converts the speech into text locally. The transcript remains editable, allowing users to correct recognition errors before preserving the story.
*2. AI Memory Extraction *
Qwen 2.5 identifies details explicitly mentioned in the transcript, including:
- People and names
- Places
- Family traditions
- Important moments
The extracted information is organized into a structured memory.
*3. Ghostlight: The Evidence Layer *
This is one of the most important features of EchoVault.
AI can organize and interpret a story, but it can also misunderstand information. How can users verify what it extracts?
Ghostlight connects extracted details to supporting sentences from the original transcript.
Instead of asking people to blindly trust an AI-generated result, it gives them a way to revisit the actual words behind a memory.
4. Ask This Memory 💬
What was the name of the place? Who was mentioned? What happened during a particular moment?
Users can ask questions about a preserved story.
The answers are designed to remain grounded in the original transcript. When a requested detail is missing, EchoVault responds:
That detail isn't present in this memory.
Because an invented memory is not a preserved memory.
*5. The Memory Vault *
Preserved stories can be revisited through the archive, bringing their people, places, traditions and moments together in one place.
Demo
See how EchoVault works:
The demonstration showcases the application workflow, from voice transcription to AI memory extraction, Ghostlight evidence and question answering.
Code
EchoVault is an open-source project.
GitHub Repository:
https://github.com/R-iddhi/Echovault
GitHub Username: R-iddhi
Explore the source code and project setup instructions:
R-iddhi
/
Echovault
A local-first AI memory vault that preserves family voice recordings, extracts people, places, traditions and moments, and answers questions with source-backed evidence.
EchoVault
Some memories refuse to disappear.
EchoVault is a local-first AI memory vault that turns family voice recordings into searchable stories. It transcribes audio, extracts people, places, traditions and moments, and answers questions using evidence from the original transcript.
Try EchoVault on your computer
Current availability: EchoVault is a local application, not a hosted website yet. The frontend, API, Whisper transcription and Ollama model run on your own computer. This keeps the AI processing local, but it means you need to complete the setup below before opening the app.
1. Install the prerequisites
Install:
2. Download the project
Open PowerShell or a terminal:
git clone https://github.com/R-iddhi/Echovault.git
cd Echovault
3. Set up the Python backend
On Windows PowerShell:
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r backend/requirements.txt
On macOS or Linux:
How I Built It
EchoVault combines a web interface, a Python API and locally running AI models.
System architecture
The application has two main processing pathways that work together.
USER
|
▼
REACT + VITE
FRONTEND
|
▼
FASTAPI
BACKEND
|
┌──────────┴──────────┐
| |
▼ ▼
AUDIO PROCESSING AI PROCESSING
| |
▼ ▼
FASTER-WHISPER OLLAMA
| |
▼ ▼
TRANSCRIPT QWEN 2.5
| |
└──────────┬──────────┘
|
▼
STRUCTURED MEMORY
|
┌────────┴────────┐
▼ ▼
GHOSTLIGHT ASK MEMORY
SOURCE EVIDENCE GROUNDED ANSWERS
| |
└────────┬────────┘
|
▼
MEMORY VAULT
Technology stack
| Layer | Technology | Purpose |
|---|---|---|
| Frontend | React, Vite | User interface |
| UI animations | Framer Motion | Interactive experience |
| API communication | Axios | Frontend and backend communication |
| Backend | Python, FastAPI | Application logic |
| Transcription | faster-whisper | Audio-to-text conversion |
| Language model | Qwen 2.5 | Memory extraction and questions |
| Local inference | Ollama | Running Qwen locally |
How can someone try EchoVault?
The current version runs locally on the user's computer. It is not yet a publicly hosted, one-click website.
Anyone interested in exploring it can clone the public repository and follow the setup instructions.
Requirements:
- Python
- Node.js and npm
- Ollama
Step 1: Clone the repository
git clone https://github.com/R-iddhi/Echovault.git
cd Echovault
Step 2: Set up the Python backend
On Windows PowerShell:
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r backend/requirements.txt
Step 3: Download the language model
Make sure Ollama is running:
ollama pull qwen2.5:3b
Step 4: Start the backend
From the project root:
.\.venv\Scripts\python.exe -m uvicorn backend.main:app --host 127.0.0.1 --port 8000
Keep this terminal running.
Step 5: Start the frontend
Open a second terminal:
cd frontend
npm install
npm run dev
Open the local address shown by Vite, usually:
http://localhost:5173
Keep the frontend, backend and Ollama running while using EchoVault.
On the first run, faster-whisper may download its transcription model. An internet connection is required for the initial model downloads.
The user experience
Once the application is running, the actual journey is simple:
Upload Recording → Review Transcript → Preserve Memory → Explore Details → Verify Evidence → Ask Questions
The current implementation performs its core AI processing on the user's own computer rather than depending on a publicly hosted AI backend.
Why Does Open Innovation Matter?
Family memories can contain personal stories, names, relationships and details that people may not want to send to external AI services.
That made local inference an important design decision for EchoVault.
I used:
- faster-whisper for local speech transcription.
- Qwen 2.5, an open-weight language model, through Ollama for local memory extraction and question answering.
Open innovation gives developers the freedom to experiment with models, modify prompts, explore different approaches and build around their own requirements.
For EchoVault, this means the core AI processing does not require a remote AI inference API once the necessary models and dependencies have been installed.
It also reduces dependence on closed AI services and makes experimentation more accessible.
Most importantly, open innovation allowed me to explore a principle that became central to the project:
AI can interpret the memory. It cannot rewrite the memory.
💭 The Spirit of It
Imagine someone telling a story today.
A grandparent speaks about a place that no longer looks the same. Someone remembers a family celebration, a childhood tradition, or a person whose voice is no longer heard every day.
Years later, another family member might wonder:
Where did that story happen? Who was there? What was that tradition?
The recording still exists, but finding one particular detail inside hours of audio can be difficult.
EchoVault tries to bridge that gap.
The voice remains the original memory. The transcript makes it readable. AI makes it searchable. Ghostlight brings the user back to the actual words.
I did not want to build another chatbot that confidently answers everything.
I wanted to build something that understands the importance of saying, I cannot find that detail in this story.
Because sometimes, preserving what was never said is just as important as preserving what was.
This project is about more than storing recordings. It is about giving forgotten stories another chance to be heard.
*Some memories refuse to disappear. *
Final Thought
A machine can process thousands of words, but the value of a memory is not measured by the number of words it contains.
It is measured by the people, places and moments it keeps alive.
Built for Hacktoberfest 2026, Week 4: Ghost in the Machine.
Top comments (0)