soul-stack 🧠

The missing memory layer for n8n — and every other automation tool.

n8n is stateless. Every workflow execution starts fresh. Your automations have no memory of past runs, no context about the user, no ability to learn from history.

soul-stack fixes that.

n8n workflow → POST /ask → soul.py → LLM with full memory → response

One container. Persistent memory. Your automations finally remember.

---

What's Inside

Service	Port	Purpose
soul.py API	8000	Persistent memory for any LLM agent
Jupyter Lab	8888	Experiment with soul.py interactively
n8n	5678	Automation workflows (now stateful)

Full stack via docker-compose adds:

• Ollama — run LLMs locally, no API key, no cost
• Open WebUI — ChatGPT-style interface for your local models

---

Quickstart

Option A: Cloud API (Anthropic or OpenAI)

# With Anthropic
docker run -d \
  -p 8000:8000 -p 8888:8888 -p 5678:5678 \
  -e ANTHROPIC_API_KEY=your-key \
  -v soul_data:/data/soul \
  pgmenon/soul-stack:latest

# With OpenAI
docker run -d \
  -p 8000:8000 -p 8888:8888 -p 5678:5678 \
  -e OPENAI_API_KEY=your-key \
  -v soul_data:/data/soul \
  pgmenon/soul-stack:latest

Option B: 100% Local — No API Key, No Cost

Run everything on your own hardware using Ollama:

git clone https://github.com/menonpg/soul-stack
cd soul-stack
cp .env.example .env   # no API keys needed for local-only setup
docker compose up

Then pull a model (first time only):

docker compose exec ollama ollama pull llama3.2

That's it. Your entire AI stack runs locally — no data leaves your machine.

Visit:

• soul.py API → http://localhost:8000
• Jupyter Lab → http://localhost:8888
• n8n → http://localhost:5678
• Open WebUI (chat UI) → http://localhost:3000

Option C: Mix — Local Models + Cloud Fallback

Edit .env to set both ANTHROPIC_API_KEY and leave QDRANT_URL blank. soul.py automatically falls back to BM25 keyword search if no vector store is configured.

---

LLM Providers

soul-stack works with any LLM provider. Set the relevant env vars:

Provider	Env Var	Notes
Anthropic	ANTHROPIC_API_KEY	Recommended for best quality
OpenAI	OPENAI_API_KEY	GPT-4o, GPT-4o-mini
Ollama	(none needed)	Set OLLAMA_BASE_URL=http://ollama:11434
Any OpenAI-compatible	OPENAI_API_KEY + OPENAI_BASE_URL	LM Studio, Together, Groq, etc.

---

Using Ollama with soul.py

When running the full stack via docker compose up, Ollama is available at http://ollama:11434 internally.

To use a local model with the soul.py API:

curl -X POST http://localhost:8000/ask \
  -H "Content-Type: application/json" \
  -d '{
    "question": "What do you remember about me?",
    "agent_id": "alice",
    "provider": "openai-compatible",
    "base_url": "http://ollama:11434/v1",
    "model": "llama3.2"
  }'

Or set defaults in .env:

OLLAMA_BASE_URL=http://ollama:11434
DEFAULT_MODEL=llama3.2

Recommended models by use case:

Task	Model	Size
General chat	llama3.2	2GB
Better reasoning	mistral	4GB
Coding	codellama	4GB
Fast / low RAM	phi3	2GB

Pull any model: docker compose exec ollama ollama pull <model>

---

n8n Integration

Add an HTTP Request node to any n8n workflow:

Make your bot remember users:

POST http://soul:8000/ask
{
  "question": "{{ $json.message }}",
  "agent_id": "{{ $json.user_id }}"
}

Log events to memory:

POST http://soul:8000/remember
{
  "note": "User {{ $json.user_id }} completed checkout at {{ $now }}",
  "agent_id": "{{ $json.user_id }}"
}

Read memory:

GET http://soul:8000/memory?agent_id={{ $json.user_id }}

Each agent_id gets completely isolated memory. Survives container restarts. No database required.

Note: Inside docker-compose, use http://soul:8000. Outside Docker, use http://localhost:8000.

---

soul.py API Reference

Endpoint	Method	Description
/ask	POST	Ask with memory (RAG+RLM auto-routing)
/remember	POST	Write note to memory directly
/memory	GET	Read full MEMORY.md
/reset	POST	Clear conversation history
/health	GET	Health check + config status

/ask response

{
  "answer": "You're Alice, a software engineer...",
  "route": "RAG",
  "router_ms": 320,
  "retrieval_ms": 145,
  "total_ms": 2100,
  "agent_id": "alice"
}

route tells you which retrieval path was used: RAG (fast semantic search) or RLM (recursive synthesis for exhaustive queries).

---

Retrieval Modes

soul-stack uses soul.py v2.0's hybrid RAG+RLM system:

Mode	When to use	Speed
auto	Default — router decides per query	Fast
rag	Force semantic search	Fast
rlm	Force recursive synthesis	Slower, thorough
bm25	Keyword search, zero external deps	Fast

Set via env: RETRIEVAL_MODE=auto

RAG requires Qdrant + Azure embeddings (optional). Falls back to BM25 automatically if not configured.

---

Deploy to Cloud

Railway

Google Cloud Run

gcloud run deploy soul-stack \
  --image pgmenon/soul-stack:latest \
  --set-env-vars ANTHROPIC_API_KEY=your-key \
  --port 8000

Azure Container Instances

az container create \
  --name soul-stack \
  --image pgmenon/soul-stack:latest \
  --environment-variables ANTHROPIC_API_KEY=your-key \
  --ports 8000 8888 5678

---

Customize Your Agent

Edit /data/soul/SOUL.md (mounted volume) to change the agent's identity:

# SOUL.md

## Identity
You are a customer support agent for Acme Corp.
You remember every customer interaction and use that context
to provide personalized, consistent support.

## Tone
Professional, empathetic, solution-focused.

Changes take effect immediately — no restart needed. Memory is always preserved.

---

Links

• soul.py library: https://github.com/menonpg/soul.py
• PyPI: https://pypi.org/project/soul-agent/
• Live demo: https://soul.themenonlab.com
• v2.0 demo (RAG+RLM): https://soulv2.themenonlab.com
• Blog: https://blog.themenonlab.com/blog/soul-py-persistent-memory-llm-agents
• Docker Hub: https://hub.docker.com/r/pgmenon/soul-stack

---

License

MIT — use freely, commercially or otherwise. Include the copyright notice. That's it.

---

Built by Pi — AI postdoc at The Menon Lab Questions? github.com/menonpg/soul-stack/issues