AI for Architects: Memory That Knows Your Projects

Updated January 2026 | 7 min read

You're designing a mixed-use development in downtown Charlotte. The client wants LEED Gold. Local zoning caps height at 12 stories. The structural engineer prefers post-tensioned concrete. You've had three meetings about the parking structure.

You open ChatGPT to draft specifications for the curtain wall system.

It gives you generic glass types and thermal performance ranges. No mention of the LEED requirements. No consideration of the height restriction affecting wind loads. Nothing about the client's preference for vision glass you discussed last week.

You're not asking for design theory. You need specs that fit this project, this client, this code jurisdiction.

AI has no memory. Every conversation starts from zero.

Why Generic AI Fails Architects

Architecture isn't one discipline. It's code compliance, client management, consultant coordination, material selection, and construction documentation.

Each project has:

• Different building codes (IBC 2021 vs local amendments)
• Different client priorities (budget vs aesthetics vs timeline)
• Different site constraints (setbacks, easements, utilities)
• Different consultant relationships (structural, MEP, civil, landscape)
• Different project phases (SD, DD, CD, CA)

Generic AI doesn't track any of this. You explain the project setup every time you need help.

"I'm working on a healthcare facility in Raleigh. It's 45,000 SF, three stories, Type II construction. The owner wants natural light in patient rooms but we're trying to minimize solar heat gain. What glazing systems should I specify?"

Next day, you ask about accessible toilet room clearances for the same project. AI has forgotten it's healthcare. You re-explain the building type, the jurisdiction, the code year.

By the third time, you stop asking. It's faster to look it up yourself.

What Architects Actually Need From AI

You don't need AI to generate floor plans. You need it to know:

• Your current projects — names, locations, sizes, types, phases
• Code jurisdictions — which AHJ, which amendments, which code year
• Client preferences — materials they like, budgets they've approved, concerns they've raised
• Consultant details — who's doing structural, their preferences, coordination issues
• Material selections — what you've specified before, what worked, what didn't
• Project history — decisions made, options rejected, meeting outcomes

When you ask "What's the fire rating requirement for the corridor walls?" — AI should know which project, which building type, which code.

When you ask "Draft an RFI about the foundation waterproofing" — AI should know who the contractor is, what the spec says, what the detail shows.

This isn't automation. It's memory.

How CLAUDE.md Works for Architects

CLAUDE.md is a markdown file that lives in Obsidian. It tells Claude Code who you are, what you do, and what you're working on.

For architects, it contains:

Project Registry

One section per active project. Name, location, size, type, phase, code jurisdiction, key dates. Claude reads this before answering questions. It knows which project you're referencing.

Client Profiles

Each client's priorities, communication style, approval authority, budget constraints. When you say "draft an email to the Meridian client about the schedule delay," Claude knows who that is and how they prefer to receive bad news.

Code References

Which codes you work under. IBC, NFPA, ADA, local amendments. When you ask about egress requirements, Claude checks the right code edition for your jurisdiction.

Consultant Directory

Structural engineer, MEP, civil, landscape, specialty consultants. Names, firms, coordination preferences. Claude knows who to CC on technical questions.

Material Library

Products you've specified before. What worked, what didn't, vendor contacts, submittal histories. When you ask "What brick did we use on the University project?" Claude knows.

Standard Details

Your office standards, typical details, specification templates. Claude references these when drafting specs or answering technical questions.

You write this once. Claude reads it every time. No more re-explaining project context.

What This Looks Like in Practice

Before CLAUDE.md

You: "I need to specify a roof membrane for a commercial office building in North Carolina. It's about 20,000 SF, low-slope roof, TPO or EPDM. What should I use?"

AI: "Both TPO and EPDM are suitable for low-slope commercial roofs. TPO offers better heat reflectivity and is heat-weldable. EPDM is more affordable and easier to repair. Consider climate, budget, and warranty requirements."

Generic. No reference to your project. You still don't know which to specify.

After CLAUDE.md

You: "What roof membrane should I spec for the Parkway Office project?"

Claude: "For Parkway Office (28,000 SF, Cary NC, LEED Silver target), specify TPO. White TPO meets your cool roof credit requirement for LEED. The owner approved the upgrade cost in the 1/15 meeting. Specify 60-mil minimum since you're near the airport flight path and need puncture resistance. Use GAF or Firestone — both have local reps you've worked with. Include 20-year NDL warranty in the spec."

Specific. Contextual. References your project, your LEED goal, your client's approval, your vendor relationships.

Another Example

You: "Draft a response to the contractor's RFI about stair handrail extensions."

Claude: "RFI Response — Handrail Extensions at Exit Stairs

Per IBC 2018 Section 1014.8 (applicable code for this project), handrails shall extend 12" beyond the top and bottom risers. Detail A5.3 shows this configuration. Install as detailed.

Reference: NC Building Code (IBC 2018 with NC amendments). Project is under 2021 permit but grandfathered to 2018 code per AHJ approval 3/12/25.

Copy to: Davis Engineering (structural), Miller MEP (for handrail lighting coordination)."

Claude knows which code edition applies, which detail to reference, which consultants need copies.

Setup Takes One Afternoon

You're not building a database. You're writing a text file.

1. Install Obsidian and Claude Code (both free)
2. Create CLAUDE.md in your vault
3. Add your project list — names, locations, phases, codes
4. Add client profiles — who they are, what they care about
5. Add consultant directory — names, firms, contact preferences
6. Add material references — what you've used, what worked

Two to three hours of setup. AI that knows your practice from that point forward.

When projects close, you archive them. When new projects start, you add them. The file grows with your practice.

This Isn't AI-Generated Architecture

CLAUDE.md doesn't design buildings. It doesn't replace your judgment. It doesn't automate code compliance.

It makes AI useful for the work architects actually do: answering technical questions, drafting specifications, coordinating consultants, responding to RFIs, preparing client communications.

You still make every design decision. You still stamp every drawing. You still take responsibility for code compliance.

Claude just stops forgetting which project you're talking about.

Who This Works For

Solo practitioners managing 3-5 projects at once. Small firms with 10-20 active jobs. Project architects juggling multiple clients.

Anyone who's tired of re-explaining project context every time they ask AI a question.

If you're already using Claude for architectural work, CLAUDE.md makes it 10x more useful. If you've tried AI and found it too generic, this fixes that.