Building Crabs 🦀 — My Personal AI Assistant

Why

I wanted more than a ChatGPT window that forgets who I am every time I close the tab. I wanted an AI assistant that lives on my server, is accessible from anywhere via Discord, and actually knows what I'm working on. It should be able to read my book notes, understand my coursework, and help me brainstorm on active projects without me copy-pasting context every single time.

So I built "crabs 🦀" — a no-nonsense, server-native assistant powered by OpenClaw. Named after Mr. Krabs, obviously.

Nothing in this architecture is groundbreaking computer science. The real breakthrough is accessibility. Between working full-time, taking two courses in the UT Austin OMSAI program, and trying to have a life, I don't have entire weekends to burn reading API docs. But by leveraging AI pair programming, a complex personal infrastructure build that would normally take weeks was dialed in over a few short evenings.

The Infrastructure

The entire setup runs on a remarkably cheap Hetzner VPS (2 vCPUs) running Ubuntu 24.04.

Untethering Obsidian with CouchDB

Instead of relying on expensive sync subscriptions, I deployed a self-hosted CouchDB instance on the VPS. Using the Obsidian LiveSync plugin, my vault syncs across laptop, phone, and any other device. A background service (livesync-bridge.service) pulls the latest markdown files into a server directory (/srv/vault/), so the AI always has real-time access to the exact same notes I'm typing on my phone.

The $20 Mistake

The initial setup had a learning curve. I enabled OpenClaw's "heartbeat" feature thinking it'd be cool for the agent to proactively check in on things. What I didn't realize was how heavy those automated background context-windows were. The bot silently burned through $20 in API costs just talking to itself. Heartbeats: disabled. Lesson: learned.

The Architecture

The Brain: Gemma-4-31b-it

Google's instruction-tuned model that natively reasons (outputting <think> blocks before answering). I configured OpenClaw to intercept and hide these thinking blocks from the Discord UI — I get the reasoning without the wall of text.

Two-Tiered Memory (the part I'm most proud of)

If you just give an LLM access to thousands of markdown files, it chokes. I needed to separate "what we were just talking about" from "my permanent knowledge base."

Short-Term Memory (memory_search): OpenClaw handles conversational memory. As Discord chats get long, its compaction engine summarizes older messages into daily markdown summaries. Follow-up questions the next day pull from these auto-generated summaries.

Long-Term Memory (LightRAG + Obsidian): My actual "second brain" lives in /srv/vault/. LightRAG ingests my markdown files and builds a knowledge graph — not just keyword matching. When the agent runs vault-search, it uses "mix" mode: semantic vector search and graph traversal simultaneously.

The graph traversal is where it shines. If I ask "How do the themes from the books I read this year connect to my 2026 career goals?", LightRAG traverses entity relationships connecting reading logs to personal goals — cross-folder insights that keyword search would completely miss.

The Zero-Guessing Policy

Smart models get lazy and rely on pre-trained weights when they think they already know something. The SOUL.md instruction set is draconian:

1. NEVER answer using general pre-trained knowledge
2. Always use vault-search first for my specific notes and context
3. Only use memory_search for recent conversational context
4. Obsidian vault is strictly read-only unless I explicitly say otherwise

When It Clicked

I sent this prompt:

"Use vault-search to find my 'Personal Reflection' on DDIA, and tell me how my takeaway connects to my technical career goals for 2026."

Checking the server logs:

1. The agent recognized it couldn't answer from general knowledge (zero-guessing policy fired)
2. It ran memory_search first — found nothing useful
3. It independently read its own config file (SKILL.md) to relearn the curl syntax for vault search
4. It hit LightRAG with mode: mix, triggering graph traversal
5. LightRAG returned my note about p99 latency and premature optimization, cross-referenced it with my "ProtoBird" project, and realized ProtoBird is the exact empirical optimization that DDIA warned me to wait for

It didn't summarize. It traversed a graph database, audited its own code, connected a textbook to my active GitHub repo, and replied in Discord within seconds.

Less chatbot, more extension of my own brain.