When I built my house, I started documenting everything: outlets, conduits, circuits, pipe runs. The kind of things you only look up once the drywall is up and can't see them anymore. Over time it grew: HVAC, vehicles, tools, appliances, even trees in the garden. I wrote a desktop app for it, and having that data came in handy more often than I expected.
But adding data was always the bottleneck. Open the app, navigate to the right room, pick categories, fill in fields. For something like "the dishwasher is a Bosch, bought in 2019" that workflow felt like overkill. Not everything worth remembering felt worth the effort.
The bottleneck disappeared
When MCP came along, I realized I could connect that database to an AI assistant and skip the UI. That's what Home Memory is: an MCP server that gives Claude, Codex, or any MCP-compatible AI structured, persistent memory about your home.
No forms. No navigation. You just talk.
"I have a Daikin Altherma heat pump in the utility room."
The AI resolves the location, picks the right category, and creates the element. One sentence, done. That isn't lost in a chat. It becomes a structured entry in a local database.
[attach a photo of a nameplate] "What is this? Add it to the utility room."
Vision-capable AIs recognize the device and create the entry via MCP. That moment sold me on the whole concept.
Not just storage: recall and decisions
The value isn't just in putting data in. It's in getting it back out, sometimes years later, in a context where it actually matters.
"My dishwasher's upper rack broke. Worth repairing given its age?"
The AI knows when you bought it, what model it is, and can reason about whether a repair makes sense. That's not a party trick. That's useful.
"Show me all cables running from the breaker panel to the kitchen."
Home Memory tracks physical connections between elements: cables, pipes, ducts, and where they run. That's the part I use the most. After drywall goes up, this is the only record that exists.
How it works
You ── AI Assistant ── Home Memory MCP ── Local Database
(natural language) (20+ tools) (Firebird, single file)
Home Memory runs on your machine as a local MCP server. The database is a single file. Nothing is sent anywhere except to the AI you're already talking to.
It comes with over 100 built-in categories to start with, from circuit breakers to garden sprinklers, and you can add your own. You start with a default house structure and reshape it by talking to your AI.
The self-contained Windows build needs no .NET or Firebird installation. On macOS and Linux: build from source with .NET 10.
What I learned building MCP tools
A few things that might be useful if you're building your own MCP server:
Tool descriptions are your UX. The AI reads them to decide when and how to call your tools. I spent more time iterating on tool descriptions than on the SQL queries behind them. A vague description means the AI picks the wrong tool or passes bad parameters. A precise one means it works on the first try. After 20+ tools I can say: this is where the real design work happens.
One word can cause data loss. My update tool originally said: "call get_element_details first if you need to read existing text." The AI treated that condition as a rational bypass. When a user said "add this to the description," the AI reasoned: "I'm adding, not reading, so I can skip the lookup." It overwrote the existing description without checking. Changing "if you need to" to "ALWAYS call get_element_details before updating" fixed it. In MCP tool descriptions, conditional instructions are opt-out invitations.
Try it
The project is open source and on GitHub:
github.com/impactjo/home-memory
The README has setup instructions for Claude Desktop, Claude Code, and Codex. There's also a project site with a demo video if you want to see it in action before cloning.
Start with whatever matters most to you. There's no need to document everything. The value grows with every entry.
Top comments (0)