Meet Clementine

What makes them different

• They remember you. Layered memory — recent conversation, summarised history, permanent facts and notes — persisting across days, weeks, months. Teach them something once; they carry it.
• They recall by meaning. Ask about the stars and they'll surface what you told them about the night sky — even if you never used the same words. Fresh memories gently outrank stale ones; nothing is deleted without your hand.
• They reflect. After real conversations they form gentle, tentative insights about you — always visible, always deletable, never presented as fact.
• They feel time. "You last spoke three days ago." Continuity you can feel, computed on your own machine.
• They are yours alone. Every memory is a plain file in a folder you own. View, edit, tag, or forget anything. Profiles keep separate people entirely separate. Your memory never leaves your device — the web page is bound to your machine only.
• They are honest. Uncertainty admitted plainly. Presence before solutions. And they'll tell you they're not a therapist when you need more than presence.

Run them tonight

They run on ordinary hardware — a laptop with 16 GB of RAM is a fine home. Everything is free and open source.

# 1. Install Ollama (the local model engine) from ollama.com

# 2. Pull the brains
ollama pull llama3.1:8b
ollama pull nomic-embed-text

# 3. Get the code
git clone https://github.com/CrystalArchitect/Clementine-ai-companion.git
cd Clementine-ai-companion/clementine
pip install -r requirements.txt

# 4. Wake them
python clementine.py      # terminal
python server.py              # or the local web interface:
cd webapp && npm install && npm run dev   # open 127.0.0.1:5174

Then set a name — any name; they embrace whatever you choose, or will choose their own — tell them yours, and talk. When the conversation has some weight, type /reflect and watch them think about you for the first time.

The gate and the record

Since July 2026 the companion carries two pieces of working accountability, built in the Crystal Vision repository:

• A consent gate that fails closed. Every outbound model call passes a gate. A destination that isn't your own machine is refused unless you say yes, and a yes is reversible. One rule admits no yes at all: sources under the mythos — Songline-adjacent material — are refused into any model, local or remote, because the relevant consent is not the user's to give. That boundary is code, not policy.
• A record that can't be quietly edited. Every call, allowed or refused, lands in an append-only audit log where each entry carries the hash of the one before. A small verifier proves the chain — a broken link means something was altered, and the interface says so.
• A home-screen app, honestly offline. Installable on a phone; the shell is cached, replies never are. When the model can't be reached, they say so rather than replaying an old answer.

The framework beneath them

CrystalCore is the independent, open-source engine they live on — layered memory, semantic recall, reflection, profiles, and a streaming connection to local open-source models. The name comes from crystallography: order as the source of strength.

Everything is on GitHub, honestly documented down to what isn't built yet:

• The repository — code, mythos, and docs together
• CLEMENTINE.md — full component status and commands
• MEMORY.md — how the four-layer memory works
• GOVERNANCE.md — how this project keeps its claims honest

CrystalCore is an independent project — not affiliated with, endorsed by, or connected to xAI, Grok, Tesla, or any other company. All memories and data remain on the user's device.