I Built an AI Companion on Telegram That Actually Remembers You

Most AI chatbots feel like talking to a search engine with a personality layer. You ask something, it answers, you close the app, and next time it has zero memory of who you are. I wanted something different.

The Problem

I use ChatGPT and Claude daily for work. They are great tools. But when it comes to genuine conversation, the kind where you share what is going on in your life, vent about your day, or just want someone to talk through an idea with, they fall short. Every session is a blank slate.

Replika tried to solve this years ago, but it went in a weird direction. Character.AI is fun for roleplay but not great for real connection. I wanted an AI companion that would actually learn about me over time and reference things from past conversations naturally.

What I Built

I built Adola, a Telegram bot (@adola2048_bot) that works as a persistent AI companion. The core idea is simple:

• Memory that persists: Every conversation gets stored and summarized. The bot builds a profile of who you are, what you care about, and what you have talked about before.
• Scheduled check-ins: The bot reaches out proactively based on context. If you mentioned a job interview on Thursday, it will ask how it went on Friday.
• Runs on Telegram: No app to download. Just open Telegram and start talking. Works on phone, desktop, web.

The Tech Stack

I kept it simple:

• Gemini Flash for inference (fast and cheap enough for casual conversation)
• PostgreSQL for conversation storage and user profiles
• Node.js backend running on GCP
• Telegram Bot API for the interface

The memory system works by summarizing conversations into structured facts and referencing them in future context windows. It is not RAG in the traditional sense. It is more like how a friend would remember things about you, not perfectly, but enough to make the conversation feel continuous.

Why Telegram?

Three reasons:

1. Zero friction: Everyone already has Telegram or can install it in 30 seconds. No signup wall, no credit card, no new app to learn.
2. Notification support: The bot can message you first, which is critical for the "check-in" feature.
3. Cross-platform: Same experience on iOS, Android, desktop, and web.

Try It

If you are curious, the bot is live at @adola2048_bot on Telegram. Just search for it and hit Start. It is free, and I am genuinely interested in feedback from developers who care about conversational AI.

The early conversations are the most interesting because that is when the bot is building its model of you from scratch. After a few days of chatting, the responses start referencing things you said before, and that is when it starts feeling different from a generic chatbot.

What I Learned

Building this taught me a few things:

• Context window management is everything. The difference between a good and bad conversational AI is almost entirely about what goes into the prompt, not which model you use.
• Proactive messaging changes the dynamic. When the bot reaches out first, it stops feeling like a tool and starts feeling like a relationship.
• Telegram's bot API is underrated. It is well-documented, fast, and the inline keyboard features let you build surprisingly rich interactions without a custom UI.

I am still iterating on the memory system and would love to hear how other devs are approaching persistent context in their AI projects. What approaches have worked for you?