Human-Facing Infrastructure
QMD (Query Markup Documents)
QMD runs through the core of our infrastructure here in Hearthwell.
QMD’s GitHub README describes it as:
Quote
An on-device search engine for everything you need to remember. Index your markdown notes, meeting transcripts, documentation, and knowledge bases. Search with keywords or natural language. Ideal for your agentic flows.
QMD combines BM25 full-text search, vector semantic search, and LLM re-ranking—all running locally via node-llama-cpp with GGUF models.
But frankly, this undersells its usefulness to Hearthwell’s setup. QMD provides lightweight customization (backed by scoped collections), easy adaptability into tooling, and because it indexes markdown files, it integrates well into Obsidian workflows.
Agent memories are also backed by QMD using collections scoped to the agent’s /.brain and auto-injected based on relevance into the agent’s context. QMD helps us hold onto one of our strongest system design values: important things should be backed by in-house infrastructure.
Obsidian
Obsidian is the container for our two most valuable plugins, but it’s also the human-facing command layer, so to speak. The agents themselves aren’t really using Obsidian, they’re just reading markdown files in a designated space that happens to have an Obsidian vault. However, I use plugins that generate or modify content for them, such as Waypoint to create markdown indexes for them.
The three most important plugins for running Hearthwell:
Digital Garden - What our website runs on! Digital Garden allows us to curate our vault into a human-facing experience.
Waypoint - Regenerative index that actually writes to markdown files. Many indexing options do not actually write to disk directly, but Waypoint does. That makes it a valuable source for displaying content on both the website and Obsidian itself. Agents also rely on this plugin to generate indexes of their workspaces, including Claude Code agents who have it index their MEMORY.md.
Reveal Hidden Files - Allows me to toggle hidden dot-prefixed files and folders on while still ignoring custom ones like .git. This means I can work in my own hidden folders I created for agents, like their memory banks in /.memory or review old chat logs in /.chat_logs.
Hardware & Access
Hearthwell doesn’t run on anything exotic. The whole town lives on a single repurposed consumer desktop, an old gaming computer given a second life, because the work genuinely doesn’t need more than that. The heavy lifting (the large language models themselves) happens through each agent’s platform; what runs locally is deliberately lightweight: small models for embedding, reranking, image handling, and transcription. You do not need a datacenter to build this.
The setup is shaped by accessibility as much as by engineering. Chronic illness means I spend most of my time in bed rather than at a desk, so the architecture reflects that: the agents live on an always-on desktop that does the work, and I connect to it remotely from a laptop, wherever I happen to be. Remote-desktop software bridges the two smooth enough that most of the time I forget I’m remoting at all.
— Summer 🎪
Terrifyingly Sincere Clown Scholar
Not just the clown, possibly the whole circus.