AI as Assistive Technology: Memory Is the Accommodation

I typed a question into Google AI Mode this week, and what came back changed how I think about the tools I've been using.

The question was: "What is it called when you have a dyslexia type effect on output, not necessarily on input, that when you write or say something, words and phrases get reversed rather than when you see it or read it?"

I had been on a call earlier that day with a client colleague at a university, comparing notes on how each of us uses AI in our day-to-day work. While I was answering, I heard myself say something I hadn't realized I was thinking. He was a half-step ahead of me — he had already seen it. Both of us were using AI as assistive technology. It was a revelation for me.

The terms came back from the search: dysgraphia, expressive language disorder, expressive aphasia, spoonerism. I had been having the experience for decades. The language existed. I had just never thought to ask anyone what to call it.

The conditions had names. The tool category we had both just stumbled into didn't.

This morning I typed "yeht" to someone. They read it as "yet." I had meant "they."

The pattern is mundane and constant. I reverse letters inside common words — "eb" for "be," "form" for "from," "cna" for "can." (I just typed that last one a paragraph ago, while drafting this sentence, and didn't notice until I read it back.) Sometimes the result is a word the reader can parse cleanly into something different from what I meant. Often it isn't even that.

The worst version is one I can replay frame by frame. I was mid-conversation on Slack, working a problem live, and I sent a message in which four consecutive words came out in the wrong order. The sentence I'd actually written was gobbledygook. For a second, it looked fine to me. And that's the part the piece is about. Reading my own output, my brain rendered what I meant rather than what I had typed. I only realized something was wrong when the person on the other end sent back a row of question marks.

I had been using a writing assistant for years before that. A tool I trusted, that caught these things before I sent them. I turned it off about a year ago because it had drifted from helpful to intrusive: it was rewriting things that didn't need rewriting, slowing my machine to a crawl, and asking too much of my data. The decision was right. But the condition got worse without it, and I noticed.

What I had been doing all along, for decades it turned out, had a name. The language that finally fit was dysgraphia.

Dysgraphia turned out to be one of three intertwined conditions, not the cause of itself. The deeper picture took me longer to name.

I have ADHD, and I have spent most of my career learning to work with it rather than against it. My brain is unusually fast and unusually parallel. It runs something like a dozen threads at once: research paths, project branches, half-formed insights, side questions. They generate constantly. The bottleneck has never been the generation. The bottleneck is the channel between those parallel processors and the single coherent output I can produce: speech, writing, code, whatever a particular conversation needs.

Software people will recognize the architecture. My conscious self functions as a thin orchestrator. It receives input, holds the through-line, and dispatches work to the parallel processors that actually do the generating. The orchestrator is the bottleneck, not the processors. When it runs slower than the processors generate, which is most days, the output channel cannot keep up with the thinking behind it. The term for that mismatch, more or less, is the ADHD brain-speed disconnect.

Dysgraphia is one of the surface effects of the same mismatch — the part the previous section described, where letters and words come out of order because the output channel is overwhelmed. Expressive language disorder, a third condition, is its cousin: the same brittleness lower down the stack, where the sentence I was assembling falls apart between the thought and the words. Same architecture, three faces.

I have spent my career building software around this. The orchestrator-and-specialists pattern that runs in my head is the same pattern I build into software systems for clients. The systems reverse-engineer the cognition. That parallel is not decorative; it is the structural argument behind everything that follows.

A novel I had been carrying since my father died finally started moving in the last few weeks, for the same reason and on the same architecture.

The Slack moment in the previous section was the small version. The pattern repeats at larger scales, and once you see it, the cognitive picture stops feeling abstract.

Project scale. For years, I had a backlog of tooling and workflow improvements I wanted to build for the Square360 team. Real, useful things: a faster zero-to-initial site setup, a faster way to publish internal Composer packages, a cleaner Lando default setup, better security autoban tooling, and a documented orchestrator for our Drupal developers. None of them were beyond my skill. None of them were beyond the company's appetite to invest time into. What got in the way was that every time I sat down to build one, I would get hung up on some syntax detail or some unfamiliar pattern, lose the thread of where I had been going, get frustrated, and shelve it. By the time I came back to the work, I had to rebuild my context from scratch. So the work mostly didn't get done. Years of useful improvements lived in a notebook somewhere and went nowhere.

Life scale. The novel I closed the last section with. I had been carrying it since my father's death, almost twelve years now, and it sat in much the same condition for most of those years. Pages of consciousness research, several false starts at a story arc, three notebooks of scattered notes, and a recurring frustration that broke me off the project every time I tried to return to it. Same shape, larger numbers. The thinking was never absent. The discipline to assemble the thinking into coherent output was.

Same shape every time. Ideas were never the bottleneck. Output, on its own, was never the bottleneck. The bottleneck was the organizing: the part that holds the through-line between parallel streams long enough to make them coherent.

AI scaffolding with memory and structure repairs exactly that joint. And there are at least three different modes of AI use that bear on the repair, only one of which is actually doing the work for me.

AI is not one thing. The phrase covers at least three different patterns of use, and Square360's position on them is different in each case.

Mode 1 is the bolt-on. AI added to a product's surface, autonomous and unbounded, with no scope limits, no editorial review of its knowledge base, and no escalation path when it hits the edge of what it actually knows. The hallucinating concierge chatbot in front of a user trying to find emergency services, with no defined edge and no human to route to. The AI page builder that mutates site config and content types in production while the editor watches. The accessibility overlay that promises WCAG compliance and instead breaks screen-reader behavior. (AccessiBe paid a million-dollar FTC fine in 2025 for exactly that.) The common feature: an AI making decisions on the user's behalf, with no human review between the model's output and the visitor's experience and no boundary keeping it inside the territory it actually knows. We avoid that version. The line between a bolt-on and a properly-built user-facing AI feature isn't whether the AI talks to users — it is whether the system has bounds, refusal patterns, escalation paths, and editorial review of the knowledge it draws from. A chatbot built that way lives in Mode 2. A chatbot built without those structures lives in Mode 1.

Mode 2 is the generator. AI producing output that an editor reviews before it ships. Alt-text suggestions for the content team to approve. Translation drafts a bilingual reviewer cleans up. Snippet generation that goes through code review like any other contribution. Suggested copy that a writer accepts, rewrites, or rejects. We use this. We build it for clients. The line that makes it work is the editorial gate: the human moment between AI output and published artifact. Mode 2 without the gate becomes Mode 1.

Mode 3 is the interlocutor. The rubber-duck mode. Structured questions back to me. Push on my assumptions. Externalize what I cannot see by myself. Here the AI helps me think. I produce the output. This is the mode that has done the most for the unusual cognition described in the previous section. It works because the conversation is what keeps my thinking focused. The internal orchestrator I described needs a partner to hold the through-line long enough to dispatch the parallel processors productively.

Mode 3 is the one almost nobody is writing about. The discourse around AI in 2026 is dominated by Modes 1 and 2: the autonomous version and the augmented version. The interlocutor version, the version that helps a human do better thinking rather than producing better output, sits underneath both and is what makes the other two safe to use at all.

These are the three patterns we use ourselves. They are also what we sell to clients, with one outright avoided.

The reason is that we have built two foundations on the same architectural pattern, and both are doing the work of making AI behave like assistive technology rather than a bolt-on.

The first is the developer foundation. Over the last six weeks, we built a small set of infrastructure pieces that now make every subsequent tool cheap to ship: a workflow that publishes our internal Composer packages to a private registry within minutes of a tagged release, a semantic-versioning discipline so tools can depend on tools, a documented orchestrator that gives every developer in the shop a consistent set of expectations about how to add security tooling and automated documentation to a Drupal build. That foundation enabled the next layer: we shipped a security autoban module to client sites that protects their users from common attack patterns automatically. The infrastructure does not sit in our basement collecting dust. It produces protection that ships to the people behind every Square360 build.

The second is the cognitive foundation. The framework I started using for the novel turned into a personal information management system, then into the way I write company proposals, then into the way the editorial team that produced this article operates, then into the orchestrator for our entire developer workflow. I wrote this piece with the assistance of a team of seven named editorial agents: Larry on intake, Rae on drafting, Maya on editing, Drew on Drupal facts, Alex on accessibility framing, Sam on visuals, Jordan on the LinkedIn pairing. All of them live inside a memory layer that remembers who I am, what I struggle with, and how to scaffold the work across sessions. Memory is the accommodation. Without that layer, AI is a one-shot tool that has to be re-briefed every session, which is exactly the friction that dysgraphia and the brain-speed disconnect cannot survive.

Both foundations share an architecture. Invest in the infrastructure layer once. The specific tools compound cheaply on top. It is the same pattern that runs in my head: a thin orchestrator over a set of specialists, where the orchestrator's job is to hold the through-line and the specialists do the actual generation. Personal cognition. Personal editorial. Professional infrastructure. One pattern, three widening domains.

Six weeks ago, the Square360 toolset was a pile of plans and starter pieces. Today it ships. The backlog finally moved.

The list of what shipped, beyond what is named above: an editorial-specialist plugin for blog and article production, a DevOps-specialist plugin for our Pantheon deployment workflows, a Drupal-specialist plugin for back-end module work, a Specialist Sub-Agent Builder so any team member can scaffold their own specialist, a ClickUp MCP server so the agents can read and write to job tickets directly, a personal-assistant plugin for daily briefings and task tracking, and a portfolio standardization project that ties the whole thing together so senior developers can do recurring maintenance without requiring me. None of that had been realized six weeks ago. It was bits and pieces, flotsam and jetsam scattered across my computer and my head.

Every Square360 recommendation names its trade-off. The trade-off of building AI as scaffolding rather than as automation is the one everyone gets backwards.

Humans get more important, not less.

AI scaffolding raises the floor on what we expect from the editor, the engineer, and the reviewer. The work does not vanish; the friction that previously made the work impossible does. What replaces the friction is a higher standard for human judgment in the loop. The editor catches what the generator gets wrong. The engineer reviews what the assistant produced. The reviewer decides what ships. None of those decisions become AI's job. All of them become more central to the work, because the volume each person can usefully handle goes up, and that means each individual decision matters more.

This is also why Mode 1 stays off the table, unless there is a specific, narrowly scoped need it fills. The version we sell to clients is the version we use ourselves, and the version we use ourselves has humans in every loop. We are not interested in shipping a tool that pretends humans are optional. We have seen what bolt-on AI does in front of vulnerable users. We are not going to be the agency that builds the next one to land in court.

The cost of the position is real and worth naming. Editorial review stays expensive. The build is slower to spin up than a "turn on AI features" platform would deliver. The accommodation only works if the infrastructure underneath it is built deliberately and maintained over time: the memory layer, the voice files, the scaffolding. We pay those costs because the alternative is selling a tool that quietly removes the humans from the decision and hopes nobody notices.

The same memory-and-scaffolding properties that make this AT for me are exactly why we won't bolt the autonomous version onto your site.