Telegram bot with a Supabase memory for each person

Telegram bot with a Supabase memory for each person

An n8n workflow that gives a Telegram bot a memory for each person. When someone writes to the bot for the first time, the
workflow opens an OpenAI conversation for them and saves its id in a Supabase table, next to their chat number. Every
later message continues the same conversation, so the bot remembers the chat after a restart or a new day.

What is inside

• workflow.json: the workflow as n8n exports it, ten nodes and five sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The original copy used OpenAI's Assistants API, with a thread, a message, a run and a list of messages for each reply.
  OpenAI's deprecations page, read on 12 November 2026, gives 26 August 2026 as the shutdown date of the Assistants API and
  names the Responses API and the Conversations API as its replacements. This copy uses them.
• Create conversation replaces the call that made a thread. Ask OpenAI replaces the three calls that sent the message,
  started the run and read the messages back. One call to /v1/responses returns the answer, and the new node
  Pick the answer reads it.
• The original copy named an assistant that belonged to its author, and the assistant held the bot's rules. The rules now sit in a
  new node, Bot settings, together with the model name, which is gpt-6-luna.
• The Supabase column openai_thread_id is now openai_conversation_id, and the table query is rewritten. The id column
  is a uuid on purpose, because If User exists checks that id is text.
• Get New Message is now on node version 1.2, which can limit who may use the bot. See the first item under Good to know.
• The eight sticky notes linked to their author's pages. Five new notes replace them, with no outside links.
• Every credential reference is removed, so you pick your own.

Before you start

• An n8n instance with a public HTTPS address. Telegram delivers updates by an HTTPS POST to the webhook address.
• A Telegram bot and its token. You make one by chatting with BotFather inside Telegram.
• A Supabase project. Its address and its secret key go into one n8n credential.
• An OpenAI API key with some credit.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. In the Supabase SQL editor, run this query to make the table:

create table public.telegram_users (
  id uuid primary key default gen_random_uuid(),
  telegram_id bigint not null unique,
  openai_conversation_id text not null,
  created_at timestamptz not null default now()
)

1. Run a second query, alter table public.telegram_users enable row level security. The table is then closed to the public
   keys of your project, and the secret key that n8n uses still works.
2. In Get New Message, add a Telegram credential with your bot token. Select the same credential in Send Message to User.
3. In Find User and Create User, add a Supabase credential with the address of your project and its secret key.
4. In Create conversation and Ask OpenAI, add an OpenAI credential with your API key.
5. Open Bot settings and rewrite the rules in instructions. You can also change the model there.
6. Press Test workflow and send a message to your bot. A row appears in telegram_users. Send a second message and the bot
   continues the chat, with no new row.
7. Turn the workflow on, so that the bot keeps answering when you are away.

How a run goes

1. Get New Message starts the run when someone sends a message to your bot.
2. Find User looks for the person's chat number in telegram_users. It outputs an empty item when there is no row.
3. If User exists sends a known person straight to Merge.
4. For a new person, Create conversation opens a conversation at OpenAI, and Create User saves its id with their chat number.
   The saved row goes to the second input of Merge.
5. Merge passes on whichever row arrived, with the conversation id in openai_conversation_id.
6. Bot settings adds the model name and the rules.
7. Ask OpenAI sends the message text with the conversation id. OpenAI adds the earlier items of that conversation to the input.
8. Pick the answer takes the text of the model's reply. If there is none, it writes a short apology.
9. Send Message to User sends it to the chat the message came from.

Change the look

• The rules and the model: Bot settings.
• The apology: the text at the end of the expression in Pick the answer.
• A fresh start for one person: delete their row in telegram_users. Their next message opens a new conversation.
• Who may use the bot: see the first item under Good to know.

Good to know

• Anyone who finds the bot can write to it and spend your OpenAI credit. To limit it, open Get New Message, add the field
  Restrict to User IDs under Additional Fields and list the Telegram user numbers that may use it. Messages from other
  people are then dropped with no error. The field is enforced on node version 1.2 and later.
• The memory lives at OpenAI. OpenAI's guide on conversation state, read on 12 November 2026, says that conversation objects and their items are not subject to the 30 day TTL.
• Each answer sends all earlier items of that conversation as input, so a long chat costs more for each new message. On
  OpenAI's model page, gpt-6-luna costs 0.10 US dollars per million input tokens and 0.50 per million output tokens.
• OpenAI's reference does not say whether instructions carry over inside a conversation, so this copy sends them with
  every message.
• The bot reads text only. A photo or a voice message has no text, so OpenAI refuses the call and the sender gets no reply.
• Replies go out in Telegram's Markdown mode, which is the n8n default. A lone _ or * in an answer can make Telegram
  refuse the message.
• The secret key of Supabase gives full access to your data. Keep it in the n8n credential and nowhere else.
• Your data: each message goes to OpenAI, and the chat number and conversation id are stored in your Supabase table.
• Not run: we did not call Telegram, Supabase or OpenAI to test this file. We ran it in a test runner with stand-in answers
  for a new person, a known person and a reply with no text.
• Credentials: the JSON holds none.