I commissioned the same research twice

I commissioned the same piece of research twice, and I did not notice until somebody started it a third time.

Here is how it goes. You put an agent on something real. It works for an hour and comes back with something genuinely good: a churn taxonomy, the reason a parser has been failing, a straight call on which of two designs to keep. You read it, you agree with it, you use the part you needed that afternoon, and you close the window.

Three weeks later you need it again. The window is gone. So you ask for it again, you pay for it again, and you get an answer that is slightly different from the first one. That is worse than getting nothing, because now you have two answers and no way to tell which one you already agreed with.

Nexus is my fix, and it is deliberately boring. It is a git repository with a contract.

It closes four holes, and I want to be specific about them because “AI memory” is a phrase that means nothing. Output landing in conversations instead of files costs you the same research commissioned twice. Settled calls getting re-argued costs you four rounds of attention on one defect. Known broken things getting shipped past costs you features built for users who already left. And sixteen projects against one person’s memory costs you a context switch tax every single time you come back to something.

So there are findings, recallable by project and tag. Decisions, each carrying an explicit list of what it forecloses, which is the field that actually stops an argument restarting. Open loops, aged and ranked by what they cost the business rather than by when they were logged. Patterns, for something that has now happened twice and wants fixing at the cause instead of being corrected a third time. Runbooks, with the exact commands. And tasks, with an owner and a due date, so a commitment somebody made out loud does not evaporate when the call ends.

The thing that makes it usable is that agents reach it over MCP. A session does not get handed context by me remembering to paste it. It asks, before it starts, what is already known. Any Claude, any GPT, a local model, on any machine, without being told in advance what exists.

The rule I had to add later is that writing to it is not free. Everything has to clear a signal bar, or you end up with a second chat log that happens to be in git, which helps nobody.

If you have solved this a better way I would honestly like to hear about it. Send me a note.