heoster-jarvis-ai-assistant

HarmoniCus (formerly Heoster Jarvis) is an open-source personal AI assistant designed to feel less like a chatbot and more like a persistent cognitive partner. Built by Codeex AI and available on GitHub, it orchestrates multiple AI models—GPT-class and open-source LLaMA variants included—through a proprietary Chain of Resonance algorithm. The goal is not to answer queries in isolation but to maintain a continuous, context-aware thread that remembers not just what you said, but often how you said it. It's aimed squarely at developers, researchers, and hobbyists who want a fully self-hosted assistant with no cloud dependency and full control over their stack. The core differentiator is Temporal Echo Memory, which stores both short-term and long-term recall using a hybrid FAISS + Redis vector store. This isn't simple conversation history; the system claims to capture the emotional and semantic texture of interactions, building what it calls a cognitive fingerprint over time. The Pulse-Listening Engine continuously adapts to tone, context, and intent, while the Multilingual Harmonic Engine supports 50+ languages with cultural nuance rather than mere translation. For live data, the Real-Time Web Symbiosis streams from APIs, news feeds, and public datasets without manual triggers. Under the hood, it's a Python stack: PyTorch, Hugging Face Transformers, LangChain v4+, FastAPI with WebSockets for bidirectional streaming, and a React 19 frontend with a Vue.js fallback. Setup involves cloning the repo, configuring your own API keys in a .env file (with an encrypted key vault, no hardcoded keys), and running harmonicus_launcher.py, which downloads roughly 2GB of model weights. It runs on 8GB RAM minimum, 16GB recommended. It's MIT-licensed, so free to use and modify. Compared to tools like Open Interpreter or privateGPT, HarmoniCus leans harder into memory and real-time streaming, but its sparse documentation and lack of official community support make it a niche sandbox for tinkerers, not a daily driver.