[Help Request] AI Agent in n8n — How to Achieve a Fully Conversational, Multi-Turn Flow?

Help me Build my Workflow
1

Hello n8n community,

I hope you’re doing great!

I am currently building an AI Agent-based flow in n8n, aimed at handling dynamic, multi-turn conversations through WhatsApp and integrating with Airtable to fetch product information for my brand, MARAS GOURMET.

I would greatly appreciate your help, as I am facing some challenges regarding conversation continuity.

:brain: What I have built so far:

  • Trigger: WhatsApp webhook receiving user messages.
  • AI Agent: Powered by OpenAI (gpt-4-turbo) with a detailed custom system prompt.
  • Memory: Using “Simple Memory” (Memory Buffer Window) with a session tied to the user’s WhatsApp number.
  • Airtable Integration: Searching products based on user queries.
  • WhatsApp Response: Sending back AI-generated replies dynamically.

The AI Agent starts the conversation correctly (asking for product information), but does not maintain a natural multi-turn dialogue.
Instead of smoothly collecting data step-by-step (email, product interest, offer pricing, etc.), it behaves more like a “one-shot” interaction: asking once, then stopping.

Even though I have:

  • Configured a Simple Memory node with contextWindowLength = 7.
  • A system prompt that instructs it to move through the conversation progressively.
  • Airtable working correctly for product lookup.

➔ The conversation is still not flowing naturally across multiple steps.
Each user response seems isolated instead of building on the previous ones.

Key questions where I need help:

  1. How can I instruct the AI Agent to truly maintain context and progress through the conversation logically?
    (e.g., Ask for product → confirm → offer pricing → offer purchase.)
  2. Is there a best practice for “slot filling” behavior (email, product, location, etc.) inside an n8n AI Agent flow?
  3. Is Simple Memory enough for this?
    Or should I implement a different memory type (like vector memory, Redis memory, or manual slot tracking)?
  4. Are there working examples of fully conversational flows (multi-turn, slot-filling, dynamic branching) that I could reference?
  5. Any optimization tips for System Instructions to ensure the agent acts like a real human conversation, not just answering and stopping?

:hammer_and_wrench: Technical details:

  • n8n Version: (your version here, e.g., 1.29.1)
  • Hosting: (Self-hosted / n8n Cloud — specify)
  • Language Model: OpenAI GPT-4 Turbo
  • Memory Node: Memory Buffer Window (context length 7)
  • Trigger: WhatsApp API (using WhatsApp Trigger node)
  • Database: Airtable (Personal Access Token)
2 Likes