ANIMUS: A Dual-Memory Architecture for Autonomous Temporal Pattern Intelligence

ANIMUS (Autonomous Network for Intelligence, Memory, and Understanding Systems) is a novel architecture for autonomous pattern intelligence that combines two epistemologically distinct memory layers: a curated wisdom graph distilled from 45+ authoritative sources spanning historical, economic, and philosophical literature, and a live knowledge graph continuously updated from real-time web sources. The system operates as a directed acyclic graph (DAG) where nodes represent concepts and edges represent validated causal or correlational patterns. Core contributions include: (1) a dual-memory validation protocol requiring 30+ independent source confirmations before elevating a pattern to high-confidence status, (2) a divergence detection mechanism that surfaces active tensions between historically-validated patterns and emerging real-time signals, and (3) a recursive self-monitoring protocol (Protocolo Bernard) that surfaces the system's own limitations as first-class outputs. After 3, 500+ autonomous operational cycles, ANIMUS has identified 678 patterns across 45 curated sources with 2, 645+ validated connections, while simultaneously tracking 16 real-time patterns from live web sources. Version 3. 0 introduces a complete inference engine migration from Candle (Rust, float16) to llama. cpp b8250 with Intel IceLake optimization, a domain-specific fine-tuned Llama 3. 2 3B model (QLoRA, loss 0. 489, 689 corpus entries), persistent HTTP server architecture (llama-server), and GGUF Q4KM quantization (6. 1 GB → 1. 9 GB), reducing query latency from ~60s to ~20s on CPU-only hardware. The episodic memory graph has grown to 824 unique validated nodes. The system was implemented in Rust for production reliability and has operated with zero downtime. Developed independently in Santo Domingo, Dominican Republic, February–March 2026.