DEV Community

Reed Dev
Reed Dev

Posted on

How to Build a Telegram Bot That Remembers Users (Node.js + Docker)

#tutorial #telegram #docker #node

Most Telegram bot tutorials show you how to echo messages back. This tutorial shows how to build a bot that actually remembers who it is talking to.

What We Are Building

A Telegram bot where each user gets persistent memory. The bot stores facts about you in a markdown file and reads it before every response. After a week of chatting, the bot knows your name, your job, your interests, and what you talked about last time.

You can try the finished version here: t.me/adola2048_bot

Architecture 

Telegram -> Webhook -> Gateway -> Per-User Container -> AI Model
                                       |
                                  MEMORY.md (persistent)
Enter fullscreen mode Exit fullscreen mode

Each user gets their own Docker container with a bind-mounted workspace directory. Inside that directory:

  • MEMORY.md - everything the AI knows about this user
  • SOUL.md - personality configuration
  • SCHEDULES.json - proactive check-in times

Step 1: The Gateway

The gateway receives Telegram webhooks and routes to the right container:

app.post("/webhook", async (req, reply) => {
  const chatId = req.body.message?.chat?.id;
  const text = req.body.message?.text;
  const container = await getOrCreateContainer(chatId);
  const response = await sendToAgent(container, text);
  await sendTelegramMessage(chatId, response);
});
Enter fullscreen mode Exit fullscreen mode

Step 2: Per-User Containers

Each container runs an AI agent with access to the workspace:

services:
  user-container:
    image: adola-agent
    volumes:
      - ./data/users/${USER_ID}/workspace:/workspace
    environment:
      - MODEL=google/gemini-2.5-flash
Enter fullscreen mode Exit fullscreen mode

The agent reads MEMORY.md at the start of every conversation and writes updates when it learns something new.

Step 3: Memory Management

The AI manages its own memory file. A typical MEMORY.md after a few conversations:

# About This Person
- Name: Alex
- Location: Berlin
- Job: Frontend developer at a startup
- Interests: climbing, cooking, sci-fi books

# Recent Context
- Feb 10: Mentioned job interview at Google on Wednesday
- Feb 11: Nervous about the interview, practiced questions together
Enter fullscreen mode Exit fullscreen mode

Step 4: Proactive Check-ins

A scheduler reads SCHEDULES.json every 30 seconds and fires messages when due:

[
  {
    "task": "Ask about Google interview",
    "due": "2026-02-12T18:00:00Z",
    "recurring": false
  }
]
Enter fullscreen mode Exit fullscreen mode

The AI creates these entries itself by writing to the file.

Results

This architecture supports 9 concurrent users on a $35/month GCP instance. Containers that are idle get stopped automatically and restart in under 2 seconds when a new message arrives.

The memory feature is what keeps users coming back. Try it yourself: t.me/adola2048_bot

Top comments (0)