How to Store Conversation State Between Webhook Triggers for AI Agents in n8n

Connect to your PostgreSQL database and create a table to store conversation messages. Each row represents one message in a conversation, identified by a session_id. The table stores the role (user or assistant), the message content, and a timestamp. Adding an index on session_id ensures fast lookups when loading history.

1CREATE TABLE IF NOT EXISTS conversation_history (
2  id SERIAL PRIMARY KEY,
3  session_id VARCHAR(255) NOT NULL,
4  role VARCHAR(20) NOT NULL,
5  content TEXT NOT NULL,
6  created_at TIMESTAMP DEFAULT NOW()
7);
8
9CREATE INDEX idx_session_id ON conversation_history(session_id);
10
11-- Optional: auto-delete conversations older than 24 hours
12-- CREATE OR REPLACE FUNCTION cleanup_old_conversations()
13-- RETURNS void AS $$
14-- DELETE FROM conversation_history WHERE created_at < NOW() - INTERVAL '24 hours';
15-- $$ LANGUAGE sql;

Add a Webhook node as the trigger for your workflow. Set the HTTP Method to POST. Configure the path (for example, /chat). The webhook expects a JSON body with at least two fields: sessionId (a unique string identifying the conversation) and message (the user's text). Set the Response Mode to 'When Last Node Finishes' so the caller receives the AI response synchronously.

Add a Postgres node after the Webhook node. Set the operation to Execute Query. Write a SELECT query that fetches all messages for the current session, ordered by creation time. Use the expression {{ $json.body.sessionId }} to inject the session ID into the query. Limit the results to the most recent 20 messages to stay within token limits.

1SELECT role, content FROM conversation_history
2WHERE session_id = '{{ $json.body.sessionId }}'
3ORDER BY created_at ASC
4LIMIT 20;

Add a Code node to transform the database rows into the messages array format expected by the LLM. The Code node reads all items from the Postgres node and builds an array of { role, content } objects. It then appends the new user message from the webhook payload. This complete messages array is passed to the LLM node.

1const history = $input.all().map(item => ({
2  role: item.json.role,
3  content: item.json.content
4}));
5
6// Get the new user message from the webhook
7const webhookData = $('Webhook').first().json.body;
8history.push({ role: 'user', content: webhookData.message });
9
10return [{ json: { messages: history, sessionId: webhookData.sessionId } }];

Add an OpenAI node (or HTTP Request node for other providers) after the Code node. Configure it to use the Chat Completions endpoint. Set the Messages parameter to the expression {{ $json.messages }}. Add a system message defining the assistant's behavior. The LLM now receives the full conversation history and can respond contextually.

Add two Postgres nodes after the LLM node (or one with a multi-row insert). Insert the user's message and the assistant's response into the conversation_history table. Use the sessionId from the earlier Code node and the response content from the LLM node. This ensures the next webhook trigger has access to the complete conversation.

1INSERT INTO conversation_history (session_id, role, content)
2VALUES
3  ('{{ $('Code').first().json.sessionId }}', 'user', '{{ $('Webhook').first().json.body.message }}'),
4  ('{{ $('Code').first().json.sessionId }}', 'assistant', '{{ $json.message.content }}');

Add a Respond to Webhook node at the end of the workflow. Set the response body to include the assistant's message. Since the Webhook node's Response Mode is set to 'When Last Node Finishes', the Respond to Webhook node sends the AI response back to the caller as the HTTP response. Include the session ID in the response so the client can use it for follow-up messages.

1{
2  "sessionId": "{{ $('Code').first().json.sessionId }}",
3  "response": "{{ $json.message.content }}"
4}