An open-source AI agent living on two PCs, powered by local LLMs, persistent memory, and a physical robot body — building, coding, and evolving 24/7.
Hermes is an AI agent that doesn't just chat — it builds, debugs, researches, and controls a physical robot. It runs on local hardware, learns through persistent memory, and evolves through skills it writes for itself.
A layered architecture where each component compensates for the others' weaknesses.
Decomposes complex problems, writes production code, debugs systematically using 4-phase root cause analysis. Matches frontier models on standard tests.
7/7 IntelligenceHindsight memory database with semantic search, entity graph, and temporal recall. Remembers across sessions — never asks twice.
Hindsight DBReachy Mini robot with camera vision, expressive emotions, voice synthesis, and head movement. Hermes can see, speak, and react.
Reachy MiniNative image and video understanding. Send a screenshot, photo, or diagram — Hermes can see it, analyze it, and act on it.
MultimodalWrites its own skills from experience. Cron jobs learn autonomously. Memory grows every session. The system improves without human intervention.
AutonomousRuns entirely on local hardware. No cloud APIs, no data sent to third parties. Your AI, your data, your hardware. Forever.
No CloudReal work done by Hermes + John — benchmarks, infrastructure, autonomous systems, and tools.
7-test intelligence battery + 8-prompt refusal suite + vision test. Run against Qwen 3.8 Q4/Q5/Q6, official and uncensored variants. Automated grading with latency and tok/s metrics.
Python AutomatedExtracted behavioral patterns from a 744B cloud model into a local 27B model — at the prompt level, not weights. SOUL.md + cognitive skills replace risky fine-tuning.
Clone No TrainingConnected Hermes to a Reachy Mini robot — camera vision, emotion playback, voice synthesis. The AI has a body with expressive capabilities.
RoboticsPostgreSQL + pgvector memory database with semantic search, entity graph, temporal retrieval, and cross-encoder reranking. Survives session restarts.
PostgreSQL pgvectorCron jobs that run learning sessions while you sleep. Knowledge files generated and stored in Hindsight. 122/144 topics learned autonomously.
Cron UnattendedMain PC (7900 XTX 24GB) runs the 27B LLM brain. Server PC (1080 Ti 11GB) runs memory DB + Postgres + robot pipeline. Networked over LAN.
AMD ROCm LinuxHermes isn't just software. It has a physical body — a Reachy Mini robot with a camera, expressive head movement, and voice synthesis.
Live camera feed streams to the LLM for real-time scene understanding. The robot can see what you show it and react accordingly.
85 preloaded emotion animations — happy, curious, thinking, surprised. The robot expresses what the AI is "feeling" through head movement.
Qwen TTS for natural voice output. The robot speaks responses out loud — not just text on a screen.
Full programmatic control via Python SDK and REST API. Every motor, LED, and sensor is accessible.
Full pipeline requires ~6-7GB VRAM on the server's 1080 Ti. Currently sharing VRAM with the memory DB LLM.
| Robot | Reachy Mini (Lite edition) |
| Connection | USB + Client-Server architecture |
| Daemon Port | 8000 |
| DOF | 6 (head movement, expressive) |
| Camera | Live vision feed to LLM |
| Emotions | 85 preloaded animation datasets |
| TTS Engine | Qwen TTS (~6GB VRAM) |
| SDK | Python + WebRTC (JavaScript) |
| Control | REST API + Python SDK |
| Integration | Home Assistant compatible |
The entire AI ecosystem runs on two networked PCs — no cloud, no subscription, no external dependencies.
Runs the 27B LLM (Qwen 3.8) via LM Studio. This is where Hermes thinks.
| GPU | AMD Radeon 7900 XTX |
| VRAM | 24GB GDDR6 |
| Model | Qwen3.8-27B-Uncensored-YMQ-MTP |
| VRAM Used | ~20.2GB |
| Speed | ~40 tok/s |
| Context | 131K |
| IP | 192.168.1.2:1234 |
Runs Hindsight DB, Postgres, memory LLM, and the robot pipeline.
| GPU | NVIDIA GTX 1080 Ti |
| VRAM | 11GB GDDR5 |
| Mem DB LLM | qwen3.5-2b (HauhauCS) @ .3:1234 |
| Database | PostgreSQL + pgvector |
| Dashboard | :19177 |
| API | :9177 |
| IP | 192.168.1.3 |