flyingkimura

A Claude agent that lives on Telegram

22 July 2026

I run a Claude Code agent on a small VPS. I talk to it through Telegram, like a contact in my phone. It reads and writes a git repo of plain markdown files, and that repo is its whole memory. Nothing lives in the chat history. Nothing lives in a database.

Why

Chat apps forget. Close the window, and the model starts from zero next time. I wanted something that keeps state across restarts, crashes, and context limits, and that I could inspect and edit by hand if I ever disagreed with what it remembered.

The design borrows directly from Andrej Karpathy’s LLM Wiki gist. His point: don’t make the model re-read raw sources and re-derive the same facts on every query. Instead, have it read once, write what it learned into a small set of markdown pages, and treat those pages as the real memory. The context window is scratch space. The files are the truth.

The repo

Three folders do the work:

A CLAUDE.md at the root is the constitution, read at the start of every session: what each folder is for, how to file a new source, when to write to the wiki versus just answering in chat, what counts as out of scope.

Running it unattended

None of this works if the process falls over the first time the VPS reboots or the network blips. The setup that keeps it alive is a handful of systemd user units plus one tmux session:

Every service start is a fresh Claude Code context, with no --continue. The tradeoff is that a fresh session doesn’t remember the last few exchanges on its own, which matters for a phone conversation with natural pauses in it. A pair of hooks bridges that gap: one rebuilds a bounded, mechanical journal of recent turns as the session runs, and freezes it as a handover when the session ends. The next fresh session gets that handover injected as context at startup. It’s short-term memory only. The durable record is still whatever got written into wiki/.

Scheduled work

Some things shouldn’t wait for me to ask. A small job runner turns a recurring Claude task into a spec file plus a prompt file, and handles locking, validation, delivery, logging, and retention so each job isn’t another bespoke script. Each job becomes a claude-job@<name>.timer systemd instance. Mine currently run a nightly wiki contradiction sweep and a weekly note, delivered over Telegram or written straight into the wiki.

A starting point

I’ve published the setup as a template repo: agent-wiki-template. It’s the same vault structure and infrastructure described above, stripped of my own content: a CLAUDE.md with the frontmatter schema and workflows already written, a few example wiki pages showing the pattern, the infra/ scripts and systemd unit files, an installer, and a SETUP.md written as an install guide for a coding agent to follow end to end, stopping only at the points that need a human: a Telegram bot token, a sudo password, a browser login.

To use it:

git clone https://github.com/alexander-miguel/agent-wiki-template ~/agent-wiki
cd ~/agent-wiki
claude

Then tell the agent: “Read SETUP.md and set this up on this host.” It needs a Linux box with systemd user services, a cheap VPS is enough, Node 22 or newer, git, curl, jq, tmux, and Python with PyYAML, plus the Claude CLI and a Telegram bot token from @BotFather. The README covers the rest, including a security tradeoff worth reading before turning it on: the session runs in an unattended permission mode, so anyone who can message the bot can direct it. The fix is to set Telegram’s pairing policy and allowlist yourself before the first start.

Early use

So far it has: