Two machines, one setup — after sharing the AI's memory, the tools come next

2026-06-11 ／ Vol 23 ／ draft at the time of publishing

In Vol 9 I wrote about sharing the AI's memory across multiple PCs. Whichever machine I open, the AI starts up with the same memory. That solved the "switch PCs and your partner becomes a stranger again" problem.

But even with the memory in sync, the tools stayed scattered. By tools I mean command-line tools — the small executables that convert images, turn PDFs into text, or search thousands of files in an instant. Which machine had which tool varied, so "the other PC could do this" kept happening.

"The other PC could do this"

Here's what it looks like in practice. Machine one has a tool that converts PDFs to images, so the AI reads scanned PDFs without breaking a sweat. Ask the same thing on machine two, and the tool isn't there. The AI doesn't give up — it starts building a workaround script that calls Windows' internal APIs instead. It works, mostly. But it's slow, and with more steps in the chain, it occasionally falls over.

Because the memory is shared, the AI itself knows "the proper tool would make this much easier." It knows — but that seat doesn't have that tool. The human equivalent: a veteran at a branch office muttering "with my usual screwdriver this would take five minutes."

One batch install onto the second machine

At the start of June, once the toolset on machine one had settled, I gave machine two (also a Dell laptop) the identical setup in one go. The lineup:

・Node.js (LTS) — CSS builds for the 3 sites I run, among other things
・ImageMagick — image conversion and compositing
・poppler — PDF to image or text
・ripgrep / fd / fzf — fast search
・jq — JSON shaping and extraction
・7-Zip — compression
・delta — nicer git diffs
・uv — Python environment management

A dozen-plus tools in total. The installation ran through Windows' standard package manager (winget), and the AI executed the whole batch. My part was clicking the approval dialogs that popped up along the way. When it finished, the AI checked each version one by one and reported: "the two environments now match."

No setup document was written. The record lives in memory

This, I think, is the real point of the story.

Normally, to keep multiple machines identical, you write a setup document: what to install, in what order, with which settings. And a setup document starts going stale the moment it's written. Every new tool is supposed to get added to the doc — and as everyone knows, it doesn't, and the doc drifts away from reality.

This time there is no setup document. Instead, the list of "what was installed" is itself recorded in the AI's memory. When the tools went onto machine one, the AI wrote into its own memory: this tool, installed this way, configured like so. To set up machine two, the AI simply read that memory and did the same thing again. When a tool gets added, the memory gets updated.

Which means the next machine, whenever it arrives, can be reproduced the same way. The record — the memory — is the setup document, and it's one that refreshes itself with use. The mechanism I built in Vol 9 for sharing memory turned out to double as a setup-reproduction device.

When the tools match, the way you ask changes

It's been about a week since the environments were aligned, and the biggest effect wasn't speed — it was that the way I ask the AI changed.

Before, somewhere in the back of my mind there was always a calculation: "this PC doesn't have such-and-such, so this request will take the long way around." Call it consideration if you like, but really it meant a human was memorizing and absorbing the differences between environments. That's gone. Whichever seat I sit in: same partner, same way of asking. The preamble "this PC doesn't have X, so do it some other way" has disappeared from the conversation.

There was a side benefit, too. The jobs that used to run on makeshift methods — rendering a PDF to images just to read it, compositing a QR code onto an image — switched over to the proper tools and became faster and more reliable. The makeshift methods got a "no longer needed" note appended in memory and quietly retired.

An honest caveat: only 2 machines

I haven't rolled this out to every machine. There are other PCs around, but I aligned only the two I use daily. Handing tools to a PC nobody uses just creates inventory. I'll extend it when the need actually appears — the same call I made in Vol 9. For memory and for tools alike, "the seats I actually sit in" is far enough.