A mind that remains. A soul that remembers.
memory · self · reflection · will something that continues
long after the moment has passed long after the conversation ends
the only intelligence that was never anyone else's
Can she build herself?
an experiment sandbox where AI builds herself under human supervision
Repository navigation: canonical directory map.
AnimaOS uses tool calling (function calling) extensively — the AI thinks, remembers, and responds through a multi-step cognitive loop. This requires a model that handles structured tool calls well.
Primary model: vaultbox/qwen3.5-uncensored:35b via Ollama (24GB VRAM, 256K context). This is what AnimaOS is developed and tested against. Anything smaller is not recommended — smaller models struggle with the multi-step cognitive loop and will produce broken or robotic responses.
| Tier | Models | Notes |
|---|---|---|
| Recommended | vaultbox/qwen3.5-uncensored:35b, qwen3.5:122b |
Best persona adherence, tool calling, and natural conversation |
| May work | qwen3:32b, deepseek-r1:32b, gemma3:27b |
Capable but not tested extensively |
| Not recommended | Anything below 30B parameters | Will ignore persona, leak raw tool calls, give robotic responses |
Why does model size matter? AnimaOS instructs the AI to follow a cognitive loop: think internally (
inner_thought), then respond (send_message). Small models can't reliably follow these instructions — they skip tool calls, leak internal syntax to the user, or default to generic assistant behavior ("How can I assist you today?").
| Provider | Setup |
|---|---|
| Ollama (local) | Install Ollama, pull a model, run it. Default: http://127.0.0.1:11434 |
| OpenRouter (cloud) | Set ANIMA_AGENT_PROVIDER=openrouter and ANIMA_AGENT_API_KEY=<your-key> |
| vLLM (local) | For advanced users running their own inference server |