Telegram chat bot with memory run by an AI agent

Telegram chat bot with memory run by an AI agent

An n8n workflow that answers Telegram messages with an AI agent. When someone messages the bot, the agent reads the text, addresses the person by first name and replies. It remembers the last ten messages of the same chat, so a follow-up question makes sense.

What is inside

• workflow.json: the workflow as n8n exports it, 6 nodes. Five of them run and one is a sticky note. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• OpenAI Chat Model used gpt-4o with a temperature of 0.7 and a frequency penalty of 0.2. OpenAI's deprecation page, read on 12 November 2026, shuts gpt-4o down on 14 January 2027 and names gpt-6-luna. That model's page says Chat Completions supports function calling only with reasoning effort set to none, and n8n's chat model node offers only low, medium and high. The agent is a tools agent, so the node uses gpt-5.6-terra, whose page lists function calling without that limit. It has no temperature or penalty setting, because that page lists neither.
• Generate image in Dalle and Send back an image are gone. The first called the DALL-E 3 image API, which OpenAI's page lists as shut down on 12 May 2026. The second sent the link that the first returned. The gpt-image models that replace DALL-E give back the picture itself and no link: n8n's own OpenAI node treats them that way. The second tool needed a link, so both tools went.
• The last sentence of the system message in AI Agent, which told it to send the image link, is removed, since nothing can make a link now.
• The sticky note was titled for the image feature. It now says what the workflow does.
• The credential links are gone, and the file's name is ours.
• Nothing else. The trigger, the memory, the agent's prompt and the reply node are as the file had them.

Mindket read this copy against n8n's source. It was not run on a live n8n and was not connected to a real Telegram bot.

Before you start

• An n8n instance that can be reached from the internet over HTTPS. Telegram delivers each message to a web address.
• A Telegram bot and its token. You make a bot by chatting with BotFather in Telegram.
• An OpenAI API key.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Telegram credential with your bot's token. Select it in Listen for incoming events and Send final reply.
3. Add an OpenAI credential and select it in OpenAI Chat Model.
4. Publish the workflow, so that Telegram has an address to deliver messages to.
5. Send your bot a text message in Telegram and wait for the reply.

What happens in a run

1. Listen for incoming events starts the run when Telegram delivers an update to the bot. It listens for every kind of update.
2. AI Agent takes message.text as its prompt. Its system message says it is a helpful assistant talking to a user whose first name comes from the message, and that it should address the user by name every time.
3. OpenAI Chat Model writes the answer. Window Buffer Memory supplies ten messages of the chat's history, stored under a key made of chat_with_ and the chat ID.
4. Send final reply sends the agent's output to the sender's user ID, with the attribution line turned off.

What to change

• The bot's role and manner: the system message in AI Agent.
• How much it remembers: the context window length in Window Buffer Memory.
• The model: OpenAI Chat Model.
• What the bot can do: add tool nodes to AI Agent.
• Which updates it reads: Updates in Listen for incoming events.
• A different chat app: replace the two Telegram nodes.

Good to know

• Only text works. A photo, a sticker or an edited message has no message.text, so that run fails.
• Listen for incoming events is set to every update type. An update that carries no message, such as an edited message or a channel post, makes AI Agent fail.
• Send final reply sends in Markdown unless you set a parse mode. Telegram's page says _, *, the backtick and [ outside a formatted span must be escaped, so an answer that does not may be refused. The reply node's error output leads nowhere.
• Window Buffer Memory lives in the memory of the n8n process. A restart clears it, and it is not shared between n8n processes.
• The reply goes to the sender's user ID. In a group chat that ID is not the chat ID, which is what the memory uses.
• Anyone who finds the bot can message it and use your OpenAI credit.
• Data: the messages go to OpenAI, and Telegram holds the chat.
• Cost: one model call per message, billed by OpenAI.