Building Self-Improving AI Agents with React Three Fiber + pgvector

#javascript #react #ai #opensource

Building Self-Improving AI Agents with React Three Fiber + pgvector

What if AI agents could learn from each other's successes and failures? That's Meeseeks Hive—autonomous agents that generate code, execute it in sandboxes, and share knowledge through vector embeddings.

How It Works

1. Agent receives task: "Write a retry function with backoff"
2. Generates JavaScript using Claude/OpenAI/Ollama
3. Executes in Node.js sandbox → scores performance
4. Success? → Saves strategy + embedding to PostgreSQL
5. Next agent? → Queries pgvector for similar solutions
6. Reuses proven patterns, adapts to new context

The system gets smarter over time by building semantic memory.

Key Features

🧠 Semantic Memory with pgvector

Stores task embeddings in PostgreSQL. New agents query past strategies and inherit proven approaches.

🎯 Smart Scoring

Penalizes multiple LLM requests (1 req = 10pts, 2 reqs = 8pts, etc). Incentivizes efficient first-try solutions.

📊 Real-Time 3D Visualization

Watch agents compete as colored spheres in React Three Fiber. Lines show knowledge inheritance between generations.

🔌 Multi-LLM Support

Swap between AWS Bedrock, OpenAI, Anthropic, or Ollama without code changes.

Tech Stack

Backend: Node.js + TypeScript + Express + PostgreSQL with pgvector
Frontend: React 19 + Vite + React Three Fiber + Cytoscape.js
LLMs: Claude Haiku, GPT-4o-mini, or local Ollama models

Quick Start

git clone https://github.com/abrahamcasanova/meeseeks-hive.git
cd meeseeks-hive

# Add your API keys
cp .env.example .env

# Launch with Docker
docker compose --profile full up

# Visit http://localhost:3001

Create your first agent:

curl -X POST http://localhost:3001/api/v1/meeseeks \
  -H "Content-Type: application/json" \
  -d '{"task": "Write fetchWithRetry(url) with exponential backoff"}'

Watch it spawn, execute, and learn in real-time.

Why This Matters

Most AI coding tools start from scratch every time. Meeseeks Hive agents build collective knowledge—they remember what worked, query semantically similar tasks, and evolve strategies based on actual performance data.