The ellma Guide

Most AI assistants are stateless — they forget you the moment the conversation ends. ellma is different. It's a persistent, personal AI that lives in Telegram and grows with you over time.

Built on OpenClaw, ellma combines a powerful AI runtime with pre-configured plugins, Swedish service integrations, and a suite of unique features that no general-purpose chatbot offers. Here's what's under the hood.

ellma builds a personal knowledge graph about you — your preferences, decisions, important people, recurring tasks, and more. This isn't keyword matching; it's semantic search powered by vector embeddings.

Ask “what did I decide about the office renovation?” and ellma finds the relevant context even if you never used those exact words. The QMD (Queryable Memory Database) stores structured facts alongside embeddings, so your assistant can reason about relationships between pieces of knowledge.

Every night, while you sleep, ellma enters Dream Mode. It processes the day's conversations, consolidates memories, identifies patterns, and generates insights.

Think of it like how your brain consolidates memories during sleep. Dream Mode strengthens important memories, prunes irrelevant ones, and surfaces connections you might have missed. When you wake up, your assistant is smarter than it was yesterday.

Named after the concept of recursive self-improvement, the Ralph Loop is how ellma manages complex engineering tasks. When you need code generated, a script written, or a technical problem solved, the Ralph Loop kicks in.

Your assistant (the “PM”) breaks the task into a clear brief, spawns a specialized coding agent, monitors its work, verifies the output, and reports back. It's AI managing AI — the same model that produced the very website you're reading right now.

ellma doesn't just execute tasks — it gets better at executing tasks. Compound Engineering is the systematic process of self-improvement: every mistake is documented, every lesson learned is encoded into skills, and every interaction refines the system.

Over weeks and months, this compounds. The assistant that serves you in month three is dramatically more capable than the one in month one — not because the underlying model changed, but because the system around it evolved.

Ordering groceries from Mathem normally takes 30-45 minutes of browsing, searching, and clicking. With ellma, it takes one message: “order the usual.”

ellma remembers your typical grocery list, navigates the Mathem website using browser automation, builds your cart, and asks for confirmation before placing the order. Over 10+ hours of development went into making this navigation reliable and fast.

OpenClaw is the powerful open-source runtime that ellma is built on. You could set it up yourself — but here's what you'd need to do:

• Provision a server and configure the runtime
• Set up Telegram bot integration
• Configure plugins (web search, browser, memory)
• Build and maintain a workspace template
• Set up monitoring, alerts, and auto-recovery
• Build the Mathem integration from scratch
• Configure Dream Mode and memory systems
• Handle updates, security patches, and scaling

ellma does all of this for you. Subscribe, provide your Telegram bot token and Claude API key, and you're live in 5 minutes. We handle the infrastructure, updates, and improvements — you just use it.