ErnOS AI: Local Agent
ErnOS is a high-performance AI agent engine that runs entirely on your hardware. No cloud. No telemetry. No API keys required. Point it at any GGUF model via llama-server, and you get a full agentic system: a dual-layer inference engine with ReAct reasoning, a 31-tool executor, a 7-tier persistent memory system, an observer audit pipeline, autonomous learning, and a 12-tab WebUI dashboard — all compiled into a single Rust binary.
https://github.com/MettaMazza/ErnOSAgent
(Still a work in progress)
Smart Reasoning & Autonomy
Two-Speed Thinking: Uses fast responses for simple queries and autonomous, multi-step planning for complex tasks.
Self-Correction: Detects failure loops and automatically pivots to new approaches instead of crashing.
🛡️ Built-in Quality Control
Observer System: A background auditor automatically intercepts and forces retries for hallucinations, laziness, or ignored instructions.
Ironclad Safety: Hardcoded, core-level boundaries prevent unauthorized system access or destructive actions.
🛠️ The Toolbelt (22 Local Tools)
System Access: Executes terminal commands, reads/writes files, and edits codebases directly.
Web & Media: Includes a headless browser, multi-provider web search, and local image generation.
Sub-Agents: Spawns child agents for background task delegation.
🧬 Deep, Persistent Memory
7-Tier System: Mimics human memory with active scratchpads, comprehensive timelines, and saved user preferences.
Skill Building: Converts complex problem-solving experiences into reusable procedures for instant future execution.
📈 Continuous Self-Improvement
Background Learning: Continuously analyzes interactions to adapt to preferences and correct behavior.
Sleep Cycles: Periodically compresses memories, prunes useless data, and solidifies new skills.
Self-Training: Uses past successes and failures to automatically retrain and upgrade its core model.
🔬 "Under the Hood" Control
Brain Inspection: Allows developers to view internal neural activations to understand the AI's decision-making.
Steering: Enables real-time instruction injection to alter personality or behavior mid-process.
🌐 User Interface & Flexibility
12-Tab Dashboard: A comprehensive web UI for chatting, managing memory, monitoring tools live, and adjusting settings.
Voice & Video: Supports live, multimodal audio and video interactions.
Model Freedom: Seamlessly swap between local models (e.g., Llama, Gemma) and external APIs (e.g., OpenAI) without code changes.