A Telegram bot built as a parallel layer to a productivity system, designed to intercept moments of negative impulse and convert them into structured reflection conversations.

How it works:

Every user message triggers an AI intervention conversation grounded in the user's Chief Definite AIM, recent insights, current goal pacing, and reflection history

Voice messages auto-transcribed via OpenAI Whisper, treated as text turns in the conversation

/done command ends the session, prompts trigger word + outcome (avoided / indulged / unclear), persists everything to SQLite for downstream daily summary integration

/restart wipes session state for fresh context

System prompt is brother-tone, AIM-grounded, 80–200 words per response, deliberately not therapy-speak

Runs in the same Node process as a separate productivity bot — two bots, two tokens, one codebase.

Tech: Node.js, Telegram Bot API, Claude API, OpenAI Whisper, SQLite.

The idea of intercepting negative impulse moments in real time is genuinely clever. most productivity tools wait for you to come to them — this flips it. voice transcription via Whisper is a smart touch too, removes friction completely. how's the response latency holding up in practice?

Telegram AI Agent with Memory, Tools, and Android Self-Hosting

Bzhela started as an experimental Telegram AI agent but evolved into a full research project focused on long-term memory, controlled context, and autonomous infrastructure. The goal was to build a context-aware assistant that tracks conversation flow and interacts with real-world tools without losing coherence in an active group chat.

The core of the system was a custom memory architecture. To prevent context collapse, I built a rolling strategy where older messages were compressed into a summary while fresh messages remained raw. I used Redis for short-term history and Pinecone for long-term fact retrieval. The agent actively extracted and managed facts, applying time-to-live limits to temporary memories so the database stayed clean.

This setup allowed the agent to be proactive rather than purely reactive. It operated with practical integrations tied to daily life. The agent checked local power outage schedules, managed Google Calendar events, and interacted bidirectionally with an iPhone via Shortcuts. It could trigger phone-side scenarios, like setting alarms or sending warnings based on battery levels.

The infrastructure was deliberately unconventional. Instead of a standard cloud server, the agent ran on a self-hosted Android phone without root. I set up a runtime using Termux, proot Ubuntu, and Node.js, managed by PM2. Cloudflare Tunnel ensured secure remote access. This gave the agent physical survivability via a power bank and access to real-world network signals.

Over several months, the system maintained a stable context, managing around 150 compressed facts and proactively messaging based on real-world triggers. It proved that an AI agent can act as a reliable, system-level tool rather than just a simple chatbot wrapper.

Key technical aspects: - Custom memory stack using Pinecone and Redis - Rolling context and automated chat summarization - Event-driven proactive messaging - Integrations with Google Calendar, webhooks, and iOS Shortcuts - Self-hosted Android runtime with Termux and Cloudflare Tunnel

Building rolling memory with Redis + Pinecone on a self-hosted Android stack is serious agent engineering → this goes far beyond a typical “Telegram chatbot.”