Why I built DIN

I learned to code at 40, alongside AI agents. Every few weeks a new model arrived, an old one retired, and the same agent behaved differently from one version to the next. I came to accept that working on the cutting edge means you are part of the experiment: you are living in their sandbox, and paying for the privilege.

By early 2026, every model nodded along. Anything seemed possible, and every change came back "done". Many of us worked in the chat window, copying code out of a conversation and pasting it into a terminal, trusting that the conversation would still be there when we needed it. It was, but nobody went back to it. The reasoning behind the code stayed in the chat, and only the code moved on.

Others moved into the terminal early and built their own memory: prompts that made the agent stop and write handoff files, instructions to read the file tree before starting, and eventually loops that ran on their own. It felt like handoffs had been solved, because the next agent could read what the last one left. But an agent only reads what the last one wrote down, and a file is not memory. It does not know what was tried, what failed, or what was only assumed to work. Underneath all of it, the original conversations piled up, and almost nobody has gone back through them to see what was actually proven.

And none of us stayed in one place. We moved between Claude, ChatGPT, Grok and open models, and then into the wave of self-hosted agents that run on their own. Every move left another history behind, in another format, with no shared record of what happened. So the assumptions piled up. Each new piece of work was built on the last one's word that it was done, and every echo that came back sounding new was carried forward with it. Before long, no one could say which parts of the project had ever actually been proven.

Memory tools came and went, and each one stored something: saved chats, summaries, snapshots, running journals. We learned the hard way that the value is in the original work, meaning what the agent actually ran and what actually came back. A retelling of the work, however good, is not the work.

Nobody was giving us provenance, so we built it. DIN records the work at the tool call, as it happens, and marks which claims a result actually backs. A claim that something is new, or that it works, gets the same mark as every other claim: backed by what ran, or only said. Then we went back through every conversation we had ever had and graded it the same way, so we could see the blast radius across our projects: if this changes, what breaks, and how much of it rests on claims nobody ever proved.

This matters more every month. Agents now run headless, orchestrate other agents, and act on your behalf as personal assistants. When an agent's word is copied forward into the next agent's work, that word is the risk. Provenance is how you keep that risk accountable.

That is what DIN stands for: the Data Integrity Network. We built it into a family of tools and pointed it at our own history first: every past conversation, and the work that came out of them. Today our agents don't start from a blank page or a handoff file. They work from the field, which captures each conversation as it happens, keeps the record honest in real time, and answers their questions with a receipt: the record and the exact passage the answer came from. Our agents work faster and more honestly because the data under them has integrity. What started as memory, and then provenance, has become a system of record: an account of what was done, by which agent, on what evidence.

That record is worth more than the work of any single day. Every conversation you have had with an AI is inference you have already paid for: the reasoning, the dead ends, the decisions and why they were made. Most of it sits in old chats, unread. DIN treats that inference as what it is: intellectual property. It is yours to keep, to build on, and to stand behind when someone asks how your product was made.

What we are giving you is the foundation: the bottom layer, the part everything else stands on. The record, the grades, the capture, the field. It is open source, and you are welcome to fork it. We mean that. But model releases used to come with the seasons, and now they come with the moon. Every month there is a new one, and every month spent rebuilding the floor is a month spent not building on it. You can spend your own inference recreating what we have already worked through, or you can start from here. The foundation is free. The floors above it, like more agents on one field, more fields and the roll-up across your organisation, come with a key, on the same download, whenever you need them. We cannot wait to see what you build.

We built DIN because we had no choice. Now it's yours.