# AI Was Solo. Now It's a Team Sport.

By Wes Pomeroy. Published April 30, 2026. https://wespomeroy.com/writing/shared-ai-project-memory-for-teams/

In a team huddle today, a colleague pulled down a project I'd been working on for weeks. Her AI knew everything. She picked up right where I left off.

It's the first time anyone other than me has been able to step into the running thread of one of these projects. Everything I'd been working through for weeks — the dead ends, the half-decisions, the things I was still figuring out. All of it sitting in her AI's head the moment she started.

Most teamwork with AI today is a relay race. One person at a time. Whoever's next has to catch up by reading the chat, the screenshots, the long messages explaining what the previous person figured out. The AI itself remembers none of it.

That stops if you can hand off the project itself, memory intact. Which is what happened today.

## How we got here

Only weeks of work to get here. The nights ran late — happily late.

It started with a personal system I'd built for my own work — just enough so the AI didn't forget what the project was about every time I opened a new chat. Pieces of memory it could pick up across sessions. Some structure for how a session started and ended. Nothing fancy.

I remade a basic version inside the firewall to get us started.

The next gate was access. The first version ran on Claude's command line. That was a wall for anyone who wasn't an engineer. So we built our own browser-based interface. It emulated Claude Desktop, and in places surpassed it. Multiple chats at once with whatever model fit the task, attachments and archive search, a settings layer tuned to how people actually work, a separate panel for background tasks so the chat stayed clean, a desktop shortcut so it felt like any other app — and the work didn't disappear when you closed the browser. The memory system held it all together underneath.

Once the interface existed, people used it. Dashboards. Analysis. The kind of operational automation that adds up once a few colleagues see what's possible.

Then came shared context. We put group reference files on a central server — some we wrote ourselves, some we boiled down from documents and conversations the company already had piled up. Any session could pull them in. Working on something that needed your team's vocabulary and history? Pull it down and start a thousand miles ahead.

Then we built a way to run a Claude on the server itself. We called it a 'group chat' — yours on the laptop, plus a second one on the server, both running on your account. Each looking at its own files. They took turns, sent files back and forth, picked up where they left off if the line dropped.

Today we added the part that lets people work the same project at once.

## What changes

A team can now work the same project at the same time, on different machines, and have it all merge back together cleanly when they're done. That's a normal way of working in software development — has been for twenty years. They've had branching, merging, version control. A hundred developers can build the same code at once without tripping over each other.

That's never been the way teams work on the project itself. The thinking. The decisions. The drafts and debates and partial conclusions. Those have always been one-person-at-a-time until somebody pastes the tenth screenshot or gathers everyone in a room.

That's what's now different.

## Why I'm writing this down

Most of the AI conversation right now is about which model. Which one. How big. How fast. Those are real questions.

The bigger questions are beyond the model. How the AI keeps its memory. How its work piles up over time. How a team works on the same thing together.

We're working together on things that weren't possible before — by any one person, or any team.

It's Awesome in the original sense of the word. I feel legit awe on a weekly basis now.