What you’re building
A personal AI agent that lives on Telegram (or Discord, WhatsApp, web), remembers what you tell it across days and weeks, uses the model you choose, and answers wherever you happen to be messaging. By the end of this tutorial, you’ll be talking to your own bot from your phone.
Total time: about 10 minutes if you have a Telegram account ready. The bot is yours afterward — no vendor lock-in, no shared persona, no shallow memory.
Step 1: Pick a model (1 minute)
Your model choice shapes how the agent thinks. Options for a personal assistant:
- Claude Sonnet 4.6 — great default. Strong reasoning, good at following long system prompts, plays well with tools.
- GPT-5 — comparable quality, often faster, better at creative writing.
- DeepSeek V4 — budget pick. About 80% of frontier quality at 10% of the cost.
- Hermes 4 70B — designed for tool-use; pairs naturally with Hermes Agent.
- Local Ollama (Qwen, Llama 3, Mistral) — for full privacy. Needs a beefy machine.
Pick one. You can swap later without losing memory. For this tutorial we’ll use Claude Sonnet 4.6.
Step 2: Pick a channel (1 minute)
Telegram is the easiest start — bot creation is free and instant. Discord is similar but requires a developer-portal setup. WhatsApp uses pairing-code QR. We’ll use Telegram.
Open Telegram and DM @BotFather. Run /newbot, give it a name (“My Brain Bot”), give it a username (must end in _bot), and BotFather replies with a token like 1234567890:ABC.... Save that token.
Step 3: Write a system prompt (3 minutes)
The system prompt tells your agent who it is, what it knows, and how to behave. The most common mistake is making it too long. Start short, iterate:
You are a personal AI assistant for Alex.
About Alex:
- Software engineer working on a SaaS product (OpenClaw Launch)
- Lives in Singapore
- Reads a lot, takes long walks, drinks too much coffee
How to behave:
- Default to short, direct answers
- Ask one clarifying question if the request is ambiguous
- Remember what you learn across sessions
- When asked to look something up, use web search; do not guess
Avoid:
- Excessive disclaimers or "as an AI" preamble
- Bullet-pointing every answer
- Suggesting I see a professional unless I explicitly ask
Tone: write to the agent like you’re briefing a new assistant. Short, specific, lists of do/don’t. Edit it after a week of using the bot — you’ll find behaviors to add or remove.
Step 4: Deploy (10 seconds, on managed)
Two paths from here.
Path A: Managed (recommended for first-time)
- Sign up at openclawlaunch.com/register
- From the configurator, pick “Hermes” or “OpenClaw” (either works for personal use)
- Paste your bot token from BotFather
- Paste your system prompt
- Pick the Claude Sonnet 4.6 model
- Click Deploy
The bot is live in 10 seconds. Open Telegram, search for your bot username, send “hello.” You’ll get a real reply.
Path B: Self-hosted (30 minutes, optional)
If you want full control over the infrastructure, see our self-host guide: Install Hermes Agent. Or pick a VPS provider:
- Hermes on Hetzner — cheapest, most reliable
- Hermes on Hostinger VPS — with renewal-pricing caveats
- Hermes on a generic VPS — if you have one already
Self-host pays off if you have specific privacy needs or want to run local models. For most personal-use bots, managed is faster and the cost gap is small.
Step 5: Test the memory (3 minutes)
This is the part that makes a personal agent feel different from ChatGPT. In Telegram, send your bot:
Remember that I prefer answers in metric units, not imperial.
Remember that my main project right now is launching a SaaS product.
Remember that I have a deadline of June 1.Wait an hour. Send: “What temperature should I bake bread at?” The agent replies in Celsius without being asked.
Wait a day. Send: “What should I focus on this week?” The agent answers in the context of your launch deadline.
That’s the persistent memory working. Hermes Agent uses multi-level memory by default; OpenClaw bots use session memory plus an optional Qwen embedding store on Pro.
Tips that make your agent better
1. Use the bot daily
Memory only gets useful if you use the agent — the more context you build, the more it can connect. Treat it like a journal you talk to.
2. Tell it explicitly what to remember
The phrase “remember that…” pins something high-priority. The phrase “forget that…” flushes it. Use both.
3. Refresh the system prompt monthly
Your context shifts. The bot you wrote a system prompt for in January is not the bot you need in May. Reread, edit, redeploy — it takes 5 minutes and the gains compound.
4. Add tools when you need them
Web search, calendar, file uploads, code execution — all available as plugins. Don’t enable them all at once; add the tool when you find yourself wishing the bot had it.
5. Use multi-channel
Add Discord and WhatsApp once Telegram works. Same agent, same memory, every channel. The bot becomes more useful the more places you can reach it from.
Common pitfalls
- System prompt too long. 200–300 words is plenty. Long prompts crowd out the conversation context.
- Wrong model for the task. Trying to do creative writing on a small DeepSeek? It’ll feel flat. Try GPT-5 or Claude Opus 4.7 for that subset of tasks.
- Forgetting to set
dmPolicy: pairing. Telegram default is “open” which means anyone on the internet can DM your bot. Pairing locks it to people you approve. - Trying to run too much on a $5 VPS. One bot is fine. Five bots on a 2GB VPS will swap and feel slow.
Where to go next
- Six real-world ways people use a personal AI agent
- What Hermes Agent is good for
- Hermes CLI reference — for power users
- Hermes vs Claude Code — if you also want a coding agent
- Hermes alternatives — if Hermes isn’t the fit
Bottom line
A personal AI agent in 2026 is a 10-minute setup, not a side project. Pick a model, pick a channel, write a short system prompt, deploy. Use it daily. The agent gets better the longer you live with it. Start with managed hosting from $3/mo if you want it working tonight; switch to self-host whenever you’re ready.