Agent Memory Turned Out to Be Text Files. Your Company Playbook Is Next

Something funny happened on the way to superintelligent AI: the frontier of agent memory turned out to be text files.

Claude Code reads a CLAUDE.md from your project folder. Cursor has rules files. AGENTS.md is an open standard for exactly this, one markdown file that tells any coding agent how your project works. No vector database. No exotic memory architecture. A file, in a folder, written in plain language.

I think this is one of the most underrated shifts in how work is changing. And it has almost nothing to do with code.

Documentation just became configuration

For most of business history, writing down how your company works was a chore with no payoff. The wiki went stale a week after you wrote it because nothing enforced it. The real process lived in people's heads and Slack scrollback.

That changes when an agent reads the doc before acting. The doc stops being documentation and becomes configuration. Write "refund requests get a reply within 24 hours, offer store credit first" in a file an agent actually reads, and that sentence is now operational. It runs.

So the incentive flips. Writing things down used to be overhead you skipped when busy. Now the written version of your process is the version that executes. Skip it and your best thinking never ships.

The skill is the sentence

If files are the memory format, the valuable skill is filling them well: describing what you want precisely enough that a machine can act on it. Not prompt tricks. Clear decisions, written down.

We built DontCode around this idea before it had a name. When you connect an external API on DontCode, you don't write the function. You write an intent: "When called with { email, firstName }, send a welcome email." The platform generates the code, and when you change the sentence, the code gets rewritten to match. The sentence is the source of truth. The code is derived from it.

The same pattern runs through the whole product. Your app's feature list is a living plan that updates itself as builds finish. Your visual direction is a blueprint you refine by chatting. What our users maintain over time is descriptions, not code.

That's also why we call our users DontCoders and mean it as a real skill. It's the exact skill the industry just converged on: describing a system clearly enough that software can build it and run it.

What to do with this

Start writing your business down as if an agent will read it. One will, sooner than you think.

• One paragraph per process. Who it's for, what triggers it, what "done" looks like.
• If you can't write the paragraph, you haven't actually decided how that process works. Better to find that out now.
• Keep it somewhere your team can see it and correct it. A file people argue over beats a spec nobody reads.

The teams that adapt fastest to agents won't be the most technical ones. They'll be the ones who already know what they want and can say it plainly.

If you want to feel what it's like when a description actually runs, that's DontCode. Describe an app, get a deployed app. There's more on how people use it over on the blog.