How to Send Follow-Up Questions to Gemini with Correct Thread History in n8n

Create a new workflow with a Webhook node. Set HTTP Method to POST and Path to /gemini-chat. Set Response Mode to 'Using Respond to Webhook Node'. The incoming payload should include a 'message' field and a 'sessionId' field. If the caller does not provide a sessionId, you will generate one from the user's identifier (IP, userId, or a UUID) in the next step.

1// Expected payload:
2// POST /webhook/gemini-chat
3// {
4//   "message": "What are the best restaurants in Paris?",
5//   "sessionId": "user_42_session_abc123",
6//   "userId": "user_42"
7// }

Add a Code node after the Webhook to normalize the session ID. If the caller provides a sessionId, use it. If not, generate one from the userId or create a new UUID. The session ID must be consistent across multiple requests from the same user so the memory node can load the correct conversation history.

1const input = $input.first().json;
2
3// Use provided sessionId, or derive from userId, or generate new
4let sessionId = input.sessionId;
5if (!sessionId) {
6  if (input.userId) {
7    sessionId = `session_${input.userId}_${new Date().toISOString().split('T')[0]}`;
8  } else {
9    sessionId = `session_${crypto.randomUUID()}`;
10  }
11}
12
13return [{
14  json: {
15    message: input.message,
16    sessionId,
17    userId: input.userId || 'anonymous'
18  }
19}];

Add an AI Agent node. Set the model to Google Gemini (gemini-2.0-flash or gemini-1.5-pro). In the Memory section, add a Postgres Chat Memory node (or Redis Chat Memory). Set the Session ID field to the expression {{ $json.sessionId }}. The memory node automatically loads all previous messages for that session and appends them to the Gemini request as conversation history. Set Context Window Length to a reasonable number (e.g., 20 messages) to prevent token limit issues on long conversations.

1// AI Agent node configuration:
2// Model: Google Gemini (gemini-2.0-flash)
3// System Message: "You are a helpful travel assistant."
4// Memory: Postgres Chat Memory
5//   - Session ID: {{ $json.sessionId }}
6//   - Context Window Length: 20
7//   - Connection: Your PostgreSQL credential

After the AI Agent node, add a Respond to Webhook node. Set the Response Body to the AI Agent's output text. Also include the sessionId in the response so the caller can use it for subsequent requests. This completes the request-response cycle for each turn in the conversation.

1// Respond to Webhook body expression:
2// {
3//   "response": "{{ $json.output }}",
4//   "sessionId": "{{ $('Ensure Session ID').first().json.sessionId }}"
5// }

Test the complete flow by sending a series of messages with the same sessionId. First send 'What are the best restaurants in Paris?' with sessionId 'test_session_1'. Then send 'Which one has the best desserts?' with the same sessionId. The second response should reference the restaurants from the first answer, proving that conversation history is being loaded correctly. Then send a message with a different sessionId to verify conversation isolation.

Create a scheduled workflow that runs daily (using a Schedule Trigger node) to clean up old conversation sessions. Add a Code node or PostgreSQL node that deletes memory records older than your retention period (e.g., 7 days). This prevents the memory table from growing indefinitely and helps with privacy compliance.

1// PostgreSQL node query for cleanup:
2// DELETE FROM n8n_chat_histories 
3// WHERE created_at < NOW() - INTERVAL '7 days';