What the House Knows

There is a document on my computer called "Tools & Materials Knowledge Base." It is 2,400 words long. It lists every power tool I own, organized by function — cutting, drilling, driving, sanding, measuring. Each entry includes the model number, the blade or bit currently installed, the accessories I have for it, and a note on what the tool is best used for. There is a section for hand tools. A section for fasteners. A section for adhesives, organized by cure time and substrate compatibility.

I did not write this document for myself. I wrote it so that an AI would stop suggesting I buy a circular saw I already own.

The problem with forgetting

AI has no memory. This is a technical fact about how large language models work, but it's also an experiential fact about what it's like to use one across a long project. Every conversation starts from zero. The AI does not know that last Tuesday we decided to use AraucoPly instead of Baltic birch. It does not know that the hallway wall is angled, or that the electrical panel is in the garage, or that my daughter's room is directly above the living room and the table saw can't run after seven.

For a software project, this doesn't matter much. Code is self-documenting. You paste in the file, the AI reads it, and you pick up where you left off. But a house is not a codebase. A house is a web of physical constraints, accumulated decisions, spatial relationships, aesthetic choices, code requirements, family routines, and personal preferences that build on each other over months and years. Losing that context between conversations is like bringing a new contractor to the job site every morning and explaining the entire project from scratch before they can do anything.

So I built a memory. Not the AI's memory — mine, formatted for the AI's consumption. A tools inventory. A design brief. An electrical brief with panel capacity and circuit assignments. Floor plans with room dimensions. A running knowledge base of decisions made and lessons learned. I called the system "Renovation Brain" and wrote a meta-prompt that turned the AI into a multi-domain expert who cross-referenced these documents before answering any question.

This is not how people imagine AI collaboration. It's not a magic box you talk to. It's a system you build around the AI so that it can function at the level the project requires. The AI brings knowledge — California building codes, structural engineering principles, the rated amperage of 18-gauge wire. You bring context — the room where your family sleeps, the aesthetic you're trying to achieve, the budget you're actually working with, the fact that you own a Dewalt table saw with a Freud 30-tooth ripping blade and a Hedgehog push block.

The collaboration works when both sides show up with what they have.

Hundreds of conversations

Over eighteen months of renovation, I had hundreds of conversations with AI about this house. Some were long — multi-hour design sessions where we evaluated seven construction methods for a sewing desk or mapped the wiring topology for an 8-zone LED lighting system. Some were thirty seconds — "What's 476 millimeters in inches?" Some were the kind of question I would have been embarrassed to ask a professional: "Explain compound angle cuts as though I'm a noob with no jargon."

The conversations spanned every trade. Structural engineering for the desk. Electrical for the Lutron system and the LED hub. Plumbing for the bathroom mirror wiring. Woodworking for the stair slats and the newel post joinery. Smart home integration for the nursery comfort automations. 3D modeling for custom electrical box extensions. Permitting and code compliance for the LA renovation process. Material science for choosing between pine and plywood ribs in a torsion box.

No single contractor would have that range. No single friend would have that patience. And critically, no professional I could afford would be available at 11pm on a Tuesday when I was standing in the garage with a headlamp, trying to figure out whether the defogger pad I'd ordered from AliExpress would even work on American voltage. (It wouldn't. 220 volts. My house runs 120.)

What accumulated

The AI forgot everything between sessions. But the house didn't. Every conversation that ended with "yes, do that" left a physical trace — a cleat lag-screwed into a stud, a parametric file on the print bed, a wire landed on the correct terminal. The conversations were ephemeral. The decisions they produced were permanent.

And something else accumulated, though it's harder to name: my own competence.

When I started this renovation, I didn't know what a torsion box was. I didn't know the difference between a story stick and a tape measure, or why you'd choose one over the other. I didn't know that a 220-volt heater on a 120-volt circuit produces 30 percent of its rated output, or that Formica needs a backer sheet on the opposite face to prevent the panel from bowing over time, or that pine ribs need a week of acclimation before they're dimensionally stable enough to glue.

I know all of that now. Not because I memorized it, but because I used it — in a specific room, on a specific material, to solve a specific problem that mattered to me. The AI was the teacher. The house was the classroom. The lesson stuck because the homework was real.

The thing the AI cannot do

There's a moment in every project where the conversation ends and the work begins. The AI can tell you that the legs should sit 7 inches from the back wall. It cannot feel whether the front edge flexes when you lean on it. It can model the clearance inside a 3D-printed electrical box to the nearest millimeter. It cannot smell the PETG as it prints or feel whether the mounting ears are stiff enough to hold a cover plate. It can calculate the spacing of twenty stair slats to perfect symmetry. It cannot see the shadow those slats cast across the floor when the morning light comes through.

The physical world is the final judge. A joint either holds or it doesn't. A surface is either flat or it isn't. A switch is either in the right place or it's three inches too far to the left for a two-year-old to reach, and no amount of recalculation will fix it once the drywall is closed.

This is what makes physical work different from digital work, and it's what makes AI's role in physical work fundamentally different from its role in software. In software, AI can write the code and the code can run and the output is the product. In physical work, AI can inform the plan but the plan is not the product. The product is the thing in the room — the desk, the slats, the switch plate, the light. Between the plan and the thing, there is sawdust and wire and sweat and the irreversible commitment of cutting a board that cannot be uncut.

Who the AI was

Not a contractor. Contractors have hands and licenses and trucks. Not an architect. Architects have spatial intuition and aesthetic vision and the legal authority to stamp drawings. Not a friend, though the availability and patience sometimes felt like friendship — 11pm, no judgment, willing to explain compound angles in plain language for the third time.

The closest analogy I can find is a reference librarian who happens to know every trade. Someone who doesn't tell you what to build, but who can instantly retrieve the California Electrical Code section you need, the deflection formula for a cantilever beam, the screw spacing for a French cleat in softwood studs, the difference between AraucoPly and generic radiata pine (there is none — same manufacturer, different label). Someone who answers the question you asked, but also answers the question you should have asked, and trusts you to tell the difference.

The reference librarian doesn't build the house. But without the librarian, the builder spends half their time looking things up, a quarter of their time guessing, and the remaining quarter making mistakes they could have avoided. The librarian doesn't swing the hammer. The librarian changes the ratio of confident swings to uncertain ones.

What the house knows

My house knows things now that it didn't know eighteen months ago. It knows that the staircase is enclosed with twenty oak slats at 3.25-inch intervals, each one notched to sit on a stringer whose pitch was measured with a digital angle finder and reproduced with a table saw blade tilt. It knows that the hallway switch sits at a 24-degree angle on a custom-printed extension box because the wall geometry demanded it. It knows that the sewing desk is 12 feet of torsion box with pine ribs acclimated for a week and legs set back 7 inches to maximize cantilever stiffness at the front edge where fabric meets the feed. It knows that the nursery thermostat overrides itself between 4 and 6am if the Awair sensor reads below 64 degrees for more than ten minutes.

My daughter doesn't know any of this. She knows that the lights come on when she reaches for the switch. She knows that her room is warm in the morning. She knows that she can't fit between the bars on the staircase, which frustrates her, and that her mother sews at a very long table, which she finds interesting. She knows the house the way you're supposed to know a house — as the place where your life happens, not as a collection of engineering decisions.

That's the final measure of the work. Not whether the math was right or the joints were tight or the AI conversation was productive. Whether the house feels like a house. Whether the people who live in it can forget how it was made and just live.

The conversations are gone — scattered across hundreds of sessions, no thread connecting them, the AI remembering none of it. But the house remembers. Every cleat, every slat, every wire, every angle. The house is the conversation, made permanent in wood and wire and light. It is the sum of every question I asked and every answer I was given and every decision I made in the gap between knowing and doing.

The house knows everything the AI forgot.