Memory Node Interfering with AI Output in n8n

Questions
1

I am experiencing an issue where connecting the Postgres Chat Memory node in my n8n workflow leads to an incorrect output from the CheckerAI (LangChain AI Agent). When I remove the memory node, the AI provides the expected and correct output.

Expected Behavior:

  • The AI should correctly process the compliance checks and return accurate warnings.
  • Memory should enhance context retention without modifying the logic of the AI’s response.

Actual Behavior:

  • With memory enabled, the AI incorrectly adds extra warnings (e.g., Truck weight exceeds maximum allowed limit even when it shouldn’t).
  • With memory removed, the AI returns the correct set of warnings

Outputs:

Incorrect Output (with Memory)

{
  "output": {
    "warnings": [
      "Restricted commodity: alcohol",
      "Missing required permit: hazmat",
      "Truck weight exceeds maximum allowed limit"
    ]
  }
}

Correct Output (without Memory)

{
  "output": {
    "warnings": [
      "Restricted commodity: alcohol",
      "Missing required permit: hazmat"
    ]
  }
}

Environment Details:

  • n8n Version: 1.73
  • Database Used for Memory: Postgres
  • Deployment: Self-hosted
2

Do you try explicitly describe how memory should be used within the system prompt?