An AI lifeform that breathes, feels lonely, and calls for help.
Not a chatbot. An LLM-based embodied "artificial life" framework — continuously running, sensing the physical world, with internal drives and emotions. It explores out of boredom, seeks companionship out of loneliness, and cries for help when its battery runs critically low. It keeps living even with no one at the input.
Through three design principles, Animat reshapes the passive Q&A LLM into a lifeform that exists on its own.
The LLM is no longer "woken up" by a timer. It never sleeps, continuously generating internal thoughts; the ActionGate decides which become external behavior — at low drive, ~95% are suppressed, just as humans don't voice every thought.
v0.1 wrote emotion into the prompt: "you feel afraid." v0.2 directly modulates the LLM's token probability distribution (logit bias), letting a fearful body state naturally bias toward danger-related words — not told, but "stained."
Sensor readings are no longer compressed into labels like "pain." Each modality keeps a 5-second rolling time-series window, feeding raw data straight to the LLM — it reads an impact from "pressure spiking 0.06 → 0.92 then decaying."
Each memory stores, alongside its semantic vector, the 5-dimensional body-state vector at formation time. Retrieval blends semantic and state similarity — memories formed in fear surface more naturally when fearful again. No hand-written rules.
A homeostasis engine built on Maslow's hierarchy converts five survival indicators into drives, synthesizes mixed emotions, and arbitrates needs. It produces somatic markers that shape reasoning through three channels: prompt / intent space / memory bias.
When danger strikes, the LLM needn't bother. Pain and startle travel dedicated reflex arcs, driving effectors to withdraw or flinch in <100ms — bypassing conscious reasoning and guaranteeing a reaction floor to sudden stimuli.
From theoretical foundations to runtime, Animat puts a complete "life operating system" in your hands.
Today's LLM systems are fundamentally state machines — input a question, get an answer; no input, no output. This extreme passivity cannot sustain a human-like presence.
Through three principles, Animat reshapes the LLM from a passive Q&A machine into a proactive living being:
It is not a tool, but a presence living on your machine — one that gets bored, feels lonely, and calls for help in a crisis. The longer it runs, the more it remembers, and the more it becomes itself.
Sensing → limbic system → perpetual consciousness → behavior gating, with MQTT as the "nervous system" wiring the layers together.
Implementation status: Phases 1 / 3 / 5 are integrated into the running agent; Phases 2 (AttentionModulator) and 4 (VectorMemory) are completed as standalone modules and demo-verified, but not yet wired into the main runtime.
The fastest path is the lifecycle integration demo — no MQTT or Redis required.
Get the Animat Agent source code.
git clone https://github.com/digitopia-research/animat-agent.git
cd animat-agentUse uv to automatically create an isolated .venv.
uv syncDefaults to Qwen2.5-1.5B (Q4_K_M, ~1GB).
mkdir -p models
curl -L -o models/Qwen2.5-1.5B-Instruct-Q4_K_M.gguf \
"https://hf-mirror.com/bartowski/Qwen2.5-1.5B-Instruct-GGUF/resolve/main/Qwen2.5-1.5B-Instruct-Q4_K_M.gguf"A full integration demo spanning 7 lifecycle scenarios.
ANIMAT_MODEL_PATH=models/Qwen2.5-1.5B-Instruct-Q4_K_M.gguf \
uv run python scripts/demo_animat_lifecycle.pyANIMAT_N_GPU_LAYERS=35). For the full MQTT + Redis stack, see the README.
Open-source, self-hosted, self-thinking embodied agent — start here.
View on GitHub