n8n-a2e

Powered by [Context-Time Training (CTT)](https://github.com/MauricioPerera/repomemory-v2) — the production validation of entity-based memory architecture applied to n8n automation.

> Key finding: A 1B-parameter model achieves 86% deploy-ready workflows with three lightweight guard rails (feedback loop + plan normalizer + inline retry). Small models fail on format, not logic — structured error feedback closes the gap without fine-tuning or larger models.

n8n-a2e transforms natural language descriptions into fully valid n8n workflow JSON, deploys them to a running n8n instance via REST API, and learns from successful deployments to improve future compositions.

"Create a workflow that watches Slack and logs messages to Google Sheets"
                        ↓
              ┌─────────────────┐
              │   1. RECALL     │  TF-IDF search → relevant nodes + patterns
              ├─────────────────┤
              │   2. COMPOSE    │  LLM generates WorkflowPlan → valid JSON
              ├─────────────────┤
              │   3. VALIDATE   │  Check params, credentials, connections
              ├─────────────────┤
              │   4. DEPLOY     │  POST to n8n REST API
              ├─────────────────┤
              │   5. LEARN      │  Save as reusable pattern
              └─────────────────┘
                        ↓
            Active workflow in n8n

• Zero runtime dependencies — pure Node.js built-ins only (crypto, fs, path, fetch)
• 436 n8n node definitions extracted from GitHub
• 8 built-in workflow patterns as seeds
• TF-IDF search with Porter stemming and domain-specific query expansion
• Content-addressable storage with SHA-256 deduplication (Git-inspired)
• 4 LLM providers — Claude, OpenAI, Ollama (local), Cloudflare Workers AI
• MCP server with 6 tools for external AI agent integration
• Interactive chat mode for conversation