This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine.
Some memories arrive as a story. Others are just a steel dabba on a counter, peanuts toasted separately, or curry leaves from a balcony pot.
Memory Lantern is a private keepsake book for saving those small details. Pick one or more notes and it asks Gemma to shape them into a short remembrance, with the source notes shown alongside the result. You can edit or delete the originals, copy the draft, or save it as a printable HTML keepsake.
I read “Ghost in the Machine” as a question about what we choose to keep. The ghost here is not a simulated person speaking from beyond; it is the detail someone wrote down and decided to carry forward.
The repository includes a sample note so you can see the flow without entering a real family story:
I ran the generation locally with Gemma 4. The sample result was:
Lemon rice was made on Sundays when cousins visited. Peanuts were toasted separately to maintain crunch. A steel dabba was on the counter, and curry leaves were used from a balcony pot.
Sources: [1]
To try it on your machine, install Ollama, pull Gemma, and run the app:
ollama pull gemma4:latest
python app.py
Then open http://127.0.0.1:8000. If Ollama is not already running as a service, start ollama serve in a second terminal. The app itself needs no Python packages.
Source: github.com/Abhishek764/memory-lantern
Memory Lantern is a small Python server and a plain HTML, CSS, and JavaScript interface. It stores notes as readable JSON in data/memories.json, which is ignored by Git so personal notes don't get committed accidentally.
When you light the lantern, the server sends only the selected notes to Ollama at 127.0.0.1:11434. Gemma 4 is the writing engine; it is not a decorative add-on. The prompt asks it to stick to concrete details, avoid inventing emotions or biography, and return the note numbers it used. The interface keeps those originals visible so the generated text can be checked against them.
flowchart LR
A[Your saved notes] --> B[Local Python app]
B -->|selected notes on localhost| C[Ollama]
C --> D[Gemma 4]
D -->|draft + source numbers| B
B --> E[Review, copy, or export]
B --> F[(memories.json on this device)]
The server binds to 127.0.0.1. The interface uses system fonts and makes no external requests. If Gemma isn't available, the keepsake box still works; generation shows a clear local setup message instead of sending anything to a cloud fallback.
These notes can be intimate. I wanted the privacy boundary to be visible in the architecture, not just promised in a privacy policy.
With an open-weight model running through Ollama, the story stays on the machine where it was written. After Gemma is installed, generation can run without an internet connection. There is no hosted AI account, no per-story API charge, and no dependency on one provider keeping a service online. The model, server, prompt, and storage format are all inspectable and replaceable.
A closed API could make the first version faster to wire up, but it would require sending family stories to a service I don't control. Here, the open approach makes local processing practical and gives the person keeping the memory the final say over the data.
Gemma can still get details wrong. This is a writing aid, not an authority: the original notes stay untouched, the output names its sources, and the person using it should review every draft before sharing.
The sample is fictional, so I won't pretend it brought back someone's real family memory. Its little test is smaller: can the tool keep “toast the peanuts separately” from dissolving into a generic recipe summary? In the local run, it did.
I hope the real moment comes when someone writes down an ordinary detail today, then finds it still there years from now.
I'm entering Best Use of Gemma. Gemma 4 runs the core remembrance generation locally through Ollama.