ghost.py: My Terminal Is Haunted (and It Remembers Me)

What I Built

ghost.py possesses your shell. You run commands through it as usual, and after each one a ghost whispers back
about what you just did. It's powered by Gemma 3 1B, an open-weight model running locally through Ollama.

Pick who haunts you:

Ghost	Personality
💀 The Sysadmin of 1999	Died in the Y2K panic. Tired, judgmental, misses Slackware.
🌼 La Catrina	The elegant skeleton of Dia de los Muertos. Treats every command as an offering on an ofrenda.
📟 The Spirit of COBOL	Summoned from a 1959 mainframe. SPEAKS ONLY IN UPPERCASE. Reminds you banks still run on it.

And the haunting goes deeper than one-liners:

• 🧠 It remembers you. Close the terminal, come back tomorrow, and the ghost greets you by recalling commands from your last visit.
• 😠 It gets angry. Every failed command raises the haunting intensity, from calm to uneasy to furious. At full fury the screen flickers.
• 🕯️ Seances. Type seance <question> to ask the ghost anything.
• 🪦 A graveyard. When you leave, the ghost carves a two-line epitaph for your session into ~/.ghost/graveyard.md. Over time it becomes a little cemetery of everything you've done in that terminal, which felt right for Dia de los Muertos: remembering what came before.

Demo

haunted:WK4> notacommand
'notacommand' is not recognized as an internal or external command
👻 Such a curious deviation. It seems you're attempting to impose order upon the chaos of the departed.

haunted:WK4> exit
👻 "Here lingers bone, a silent plea,
   Remembered echoes, eternally free."
Epitaph carved into ~/.ghost/graveyard.md

> python ghost.py --ghost cobol dir /b
👻 A SIMPLE QUERY, STILL, STILL A SIMPLE QUERY.

Code: https://github.com/SrilalS/ghost-py

How I Built It

One Python file, standard library only, talking to Ollama's local /api/chat endpoint.

1. Run the command with subprocess and print its output (cd is handled by the shell itself so it sticks).
2. Build the ghost's prompt: persona + current mood (based on failure count) + memories from past sessions.
3. Send it with recent conversation history, so the ghost can build on its earlier comments.
4. Stream the reply token by token, printed with a small typewriter delay.
5. On exit, ask for an epitaph, append it to the graveyard, and save the last few commands to ~/.ghost/memory.json for next time.

The fun part was how much personality a 1B model has when you give it a strong character and a mood.

Why Open Innovation Matters Here

• Your commands stay on your machine. Shell history and output are full of paths, hostnames and secrets, and this ghost also keeps a memory of them. Sending all of that to a hosted API would be the actually scary part. Here nothing leaves localhost, and the memory is a plain JSON file you can read or delete.
• Works with the lights off. No internet, no API key, no bill. It runs on an ordinary laptop.
• Swap the ghost. One environment variable (GHOST_MODEL) trades Gemma for any other open model, and each personality is just a prompt you can edit. A closed service could change or disappear; this ghost stays put.
• A 1B model is enough. Short, in-character reactions don't need a frontier model, and a tiny local one answers in about a second.