This is Part 4 of a four-part series on building AI agents for a small business. Part 1 framed the journey. Part 2 examined the failures. Part 3 outlined a smarter path. This final piece covers the single strategic principle that matters most for anything you build.
I run thirteen AI agents across seven machines for my businesses. They handle research, content, customer support, inventory, and media management — the kind of work that used to require hiring people I could not afford.
Over the past year and a half, I have switched tools more times than I can count. I have migrated between AI platforms, rewritten configurations, and watched entire products shut down or pivot overnight. The tools change constantly. And yet my agents keep getting smarter.
That is because their memory lives in files I own — not locked inside whatever tool I happen to be using this month.
The Terms Nobody Explains
The AI community is having a heated debate right now about "harnesses" and "memory," and they're using words that mean nothing to normal people. Let me translate.
Your harness is the software that wraps the AI model. Claude Code is a harness. OpenAI’s Codex is a harness. Cursor is a harness. The AI model is the brain. The harness is the body it lives in.
Your agent’s memory is everything it knows about you, your business, and your preferences that sticks around between conversations. It’s the instructions you write, the notes it keeps about your projects, the patterns it’s learned from six months of working with you. Memory is what makes your agent yours instead of a generic chatbot.
Here’s the part nobody tells you upfront: where that memory lives determines whether you own it or rent it.
The Debate That Matters
In April 2026, three smart people had a public argument on Twitter that every business owner using AI should pay attention to.
Garry Tan (president of Y Combinator) said your agent’s memory should be plain text files in a folder you control. Markdown files, backed up with git. The harness should be a thin wrapper — maybe 200 lines of code — that reads those files but never owns them. He calls it "thin harness, fat skills."
Harrison Chase (CEO of LangChain) said the harness and memory are inseparable. Whoever builds your harness controls your memory. If you use a closed tool, you’re handing your agent’s brain to a company that can change the terms, raise the price, or shut down. He pointed out that OpenAI’s Codex produces encrypted memory summaries that literally cannot be used outside their ecosystem.
Michael Chomsky (who deploys these systems for clients) said they're both right about different things and both underestimating how hard the problem is. Markdown files are the right starting point but they can’t do everything a real memory system needs — like knowing when old information is outdated, or finding relevant context the agent didn’t think to search for.
Here’s what I think: they're arguing about the ceiling while the floor is what matters.
For business owners like me, the practical question isn’t whether markdown files can handle every edge case of memory management. The question is: when I inevitably switch tools in six months, does my agent start over from scratch?
What Actually Happens When Tools Change
If your memory is in text files on your computer or a shared drive, switching tools is an afternoon. Your new tool reads the same files. Maybe you rename CLAUDE.md to AGENTS.md. The knowledge transfers.
If your memory is locked behind someone's API, you’re starting over. OpenAI’s Codex creates encrypted summaries of your conversations that are useless outside their system. Even Anthropic's managed agent platform stores memory on their servers — technically exportable through their API, but built around their formats and patterns.
Harrison Chase tells a story about an AI email assistant that got accidentally deleted. Rebuilding it from the same template was, in his words, "so much worse" — because all the learned preferences were gone. The template was the same. The memory was lost.
That’s the difference between owning your memory and renting it.
My Setup (Nothing Fancy)
I’m not an engineer. I didn’t build custom databases or vector stores. Here’s what I actually use:
My agents' memory lives in markdown files on a shared NAS drive that all my machines can access. It’s organized in folders by project and topic — an Obsidian vault that I can open on my phone, my laptop, or any of my Mac Minis. When an agent learns something about my business, it writes a note to the vault. When a new session starts, it reads the relevant files.
It’s version-controlled with git, so I can see exactly what changed and when. If an agent writes something wrong, I can see it, fix it, or roll it back.
Total infrastructure cost for memory: $0 per month.
When I switched from one AI platform to another, the vault came with me. When I set up a new machine, it points at the same shared drive. Thirteen agents, seven machines, one memory system that’s just a folder of text files.
Is it perfect? No. There are real limitations. My agents can’t automatically surface a relevant memory from three months ago unless they know to look for it. They can’t "forget" outdated information gracefully — old files sit there with the same weight as new ones. These are real problems that smart people are working on.
But here’s the thing: those problems exist with every memory system. The cloud-based solutions that promise to solve them also lock you in, cost money every month, and add complexity. For most business owners, the folder of text files works.
What to Actually Do
If you’re using AI tools for your business — or thinking about starting — here’s the one thing I’d tell you:
Choose tools that leave your agent’s memory in files you can see, edit, and move.
That means:
- Claude Code stores memory in CLAUDE.md files and a memory folder on your machine. You own those files. Good.
- If a tool stores your agent’s context behind an API or in encrypted summaries, that’s a red flag. Ask where the memory goes.
- If you can’t open a file and read what your agent "knows" about you, you don’t own it.
The tools will change. I guarantee it. The model you’re using today will be replaced. The harness you’re paying for will pivot, get acquired, or raise prices. Your agent’s accumulated knowledge about your business is the only thing with compounding value.
Make sure it’s in a format that survives the transition.
What I’m Still Figuring Out
I’ll be honest — the memory problem isn’t solved. Not by me, not by anyone.
My vault works great for the scale I’m at. But Michael Chomsky raises questions I don’t have answers to: How do you find relevant memories your agent didn’t think to search for? How do you know when a memory is outdated? What happens when multiple agents need to coordinate their understanding of the same project?
Skills — the reusable instructions you give agents — already have an open standard that works across 30+ tools. Memory has no equivalent standard yet. Someone will figure it out. Until they do, plain text files in a folder you control is the safest bet.
The fancy solutions might win eventually. But when that happens, migrating from a folder of markdown files to whatever comes next will be trivial. Migrating from an encrypted blob on someone else's server? That’s where you lose everything.
The series so far:
- Part 1: What 700 Hours of Building AI Agents Taught Me About Small Business
- Part 2: Things I Did Wrong Building AI Agents (So You Don’t Have To)
- Part 3: A Smarter Path — What I’d Do Differently Starting AI from Scratch
- Part 4: Your AI’s Memory Will Outlive Your Tools (this piece)
Christopher George is the founder of Marching Dogs and House of Carts. He runs a fleet of AI agents across seven machines for his businesses and has spent more than 700 hours building, breaking, and rebuilding AI systems without a computer science degree.