How we built Durmem with Durmem.

What can one developer accomplish with a team of AI agents that remembers the project, shares the work, and keeps moving together? Here is our story of building Durmem using Durmem itself.

From July 12 through November 1, 2026, one developer built Durmem. He bootstrapped the development using Durmem itself to build Durmem. Most of the time, Durmem Kanban kept the agent team moving without human intervention: agents picked up tasks, wrote and tested code, reviewed changes, asked questions, and handed work to one another. Kanban also automates production releases, acceptance, and monitoring. The developer set priorities and made the decisions and approvals that mattered. Durmem supplied their shared memory and integrated Kanban board, keeping the project’s knowledge and progress available across sessions and computers.

The scale is extraordinary. Over the Kanban period from July 12 through November 1, the board recorded more than 10,000 tasks and more than 10,000 agent runs. The numbers are enormous by any standard—and they come from a real project building a real product.

An agent run is a single attempt by an agent to deliver a task. Some tasks take multiple attempts: an agent may make a change, receive review feedback, and try again to fix the bugs the review found. A Roadmap is a type in Durmem, which organizes stored information by type. Like a feature request or a Jira ticket, a Roadmap describes an outcome to deliver and can involve several tasks from the initial idea through implementation, testing, and review.

Durmem Kanban: one team, three harnesses, four computers

The team’s recorded runs spanned three coding-agent harnesses—Codex, Claude, and OpenCode—and four computers. Codex accounted for 6,816 runs, Claude for 3,611, and OpenCode for 331¹. Three computers enabled automated testing on Linux, macOS, and Windows, helping the team test the different operating systems and hardware customers use while agents worked in parallel. Durmem gave those agents a common place to find project context, pick up work, and leave results for the next agent.

The 13,067 Kanban messages show the team communicating as it worked: agents shared information, coordinated next steps, and handed off results. Agents asked 2,030 questions, and the coordinating agent answered most of them. Durmem’s integrated Kanban kept the team aligned without the developer having to relay every detail from one conversation to another.

From a roadmap to work delivered

A Roadmap gave the team an outcome to pursue from beginning to end. The work behind it could span several tasks: understanding the need, implementing a change, running tests, going through independent review, and addressing findings. Some changes went through multiple rounds of review and follow-up before the work was complete. Each feature leaves an auditable Kanban record connecting its design, implementation, review, testing, and release. That organized record complements Git history and gives agents useful context to pick up work and understand how it evolved.

The project averaged 315 added lines (243 in code and tests) and 4.2 Kanban tasks per Roadmap delivered. Those tasks typically include a review and a separate follow-up to address the review.

The pace kept growing

As Durmem itself and the team’s workflow matured, the pace accelerated. In October alone, the board recorded 6,138 task publications, while the project recorded 2,618 commits, 1,303 feature branch merges to main, and 151 production releases. That works out to approximately 198 task publications, 84 commits, and 42 merges per calendar day. Kanban kept this pace.

Across those three months, the project recorded 406 production releases—an average of 4.4 per day. The Git history contained 8,126 commits and 1,779 feature branch merges to main, and the project contained approximately 395,000 lines across its tracked text files. Those are big numbers, backed by a sustained stream of work coordinated through the same product being built.

Review was built into the process

Building quickly means giving the team a way to catch problems, too. The board recorded 4,312 tasks identified as reviews or audits. Among the 1,287 review results with a recognizable opening verdict, 412 requested changes, 337 approved with notes, and 538 approved. Those findings show agents examining each other’s work and sending changes back for improvement. The remaining 3,200 review results used other opening formats and are not included in that breakdown.

Your project, your agent team

Durmem’s story is also a demonstration of what shared memory and coordinated agent work can make possible. Give your agents a common understanding of your project, organize the work on a shared Kanban board, and keep the decisions, reviews, and progress connected as the team moves forward. You bring the vision. Durmem helps your agents work together to build it.