impactjo

Images definitely have their place. The desktop app this grew out of lets me attach photos to elements. But even simple structured positions and routes cover a lot of day-to-day cases: "switch at 110 cm, 15 cm from the left edge", cable runs section by section, etc.

For true old-house spaghetti behind walls, photos or sketches probably matter more. But for handoff and routine lookups, being able to ask for all switches in an area and their positions beats digging through scattered notes.

For a small set of notes, an agent over files works fine. The difference here is that the house is stored as rooms, devices, cables, connections, and dates, not just as summaries. That makes questions like "which cables leave this panel and where do they end?" or "which appliances in the utility room are older than 10 years?" much more reliable. With files or transcripts, the model has to reconstruct that structure from prose every time, and a fully documented house quickly runs into context-window limits. Here it makes small, focused queries that return just the slice it needs.

Good association. I actually considered "Home Twin" early on, but moved away from it because "digital twin" suggests real-time sync, telemetry, or simulation. This is closer to a memory of the house: what's there, how it connects, and what happened to it over time. "Home Memory" felt like a better fit.

For a small set of notes, an agent over files works fine. The difference here is that the house is stored as rooms, devices, cables, connections, and dates, not just as summaries. That makes questions like "which cables leave this panel and where do they end?" or "which appliances in the utility room are older than 10 years?" much more reliable.

With files or transcripts, the model has to reconstruct that structure from prose every time, and a fully documented house quickly runs into context-window limits. Here it makes small, focused queries that return just the slice it needs.

Late here in Austria. Happy to follow up tomorrow.

Interesting. I use HA myself (though not the AI/LLM side yet). I think it's a different layer: Home Memory is the physical reality of the whole house, not just what's wired and smart. How do you picture "running inside"? Chat from the HA UI, or just running on the same hardware? I haven't dug into the feasibility yet. The part I'm fairly sure about: this only shines with a model that reliably uses tools. 23 MCP tools is a lot for a small local model, Claude or GPT tier handles it fine. What conversation agent are you running in HA?

I've been documenting my house in software for years: outlets, conduits, circuits, pipe runs, appliances, vehicles. Having that data came in handy more often than I expected. But adding data was always the bottleneck. Opening the app, navigating to the right room, picking categories, filling in fields.

So I built an MCP server for that database. Now Claude, Codex, and other MCP-compatible clients can read and write the home data directly. Tell it about a heat pump in one sentence, or snap a photo of the nameplate, and it figures out the category and location. Later you can ask whether it's worth repairing given its age.

It also tracks physical connections between elements: cables, pipes, ducts, and where they run. The database is a single local file on your machine.

Project site with demo video: https://home-memory.com