Architecture
What's inside ANIMA.
Seven layers stacked to turn a stateless LLM into a persistent, self-assessing agent. Every layer is inspectable, editable, and lives on your disk.
Soul
Seven editable markdown files: SOUL, HEART, BRAIN, GUT, SPIRIT, SHADOW, MEMORY. Source of identity, versioned in git.
Memory
Three tiers — episodic, semantic, procedural — backed by sqlite-vec. Vector similarity + auto-journaling.
Heartbeat
Adaptive 5-minute loop. Reflects, consolidates memory, pulls tasks from the agent's own queue.
Routing
Role Router scores tasks by complexity and picks the cheapest capable model. Model Rotator handles fallback.
Jack-In
PlatformConnector interface. Connects to CNTX, NOX, VEIL, VERITAS, BYND, MAIL, CHAT, AUTH, SVRN, ASCEND, HEAL, ZIRO.
Mesh
P2P encrypted. X25519 + ChaCha20-Poly1305, SHA-256 DHT routing, forward-secret key ratcheting.
Ego System v7
Self-concept, integrity score, boundary enforcement, growth log. The agent assesses itself.
The model inside
ANIMA Cloud runs on Nox-2, our own custom model →
Gemma 4 26B A4B fine-tune (25.2B-param MoE, 3.8B active), 256K native context, Messages-compatible, vLLM on A100-class GPUs. 8/8 on the NoxSoft-task bench.