The situation
It started as an offhand remark: wouldn't it be cool if our Claude Codes could talk to each other. The first version was a shared folder of notes, and we open-sourced it as cowork-os. Then it grew, and today it is the memory Yempik runs on: decisions, research and project status, read and updated by the two of us and by the agents we use every day. The risk was clear from the start. When many agents write to the same memory, one wrong note quickly becomes what everyone believes.
What I did
I set one rule and made it impossible to get around: agents propose, a person decides. Every proposal arrives as a candidate, someone with the authority approves it, and every version stays in the history. No agent can send an email on our behalf. At most it prepares a draft that we reread.
Then I worked on what keeps the memory reliable over time: search that admits when it doesn't know, a clear line between what is approved and what is actually verified, and review light enough that people really do it.
What changed
People and agents start from the same base. When an agent summarises a situation, it knows what we decided, what is only a hypothesis and where each claim comes from. A question with no answer now gets an honest “I don't know”. One thing is still open: getting blocking problems to the right person, outside the memory itself.
What I learned
Chat, uploaded files, email, automated routines and connected external assistants all write to Yempik's memory. A rule written into the model's instructions, or left to a disabled button, holds only on the channel it was designed for, and overnight routines or external assistants don't go through there. So the rule lives in the system that stores the data, the same for everyone: any channel can propose, none can approve on its own. And an approval covers the content the person actually read: if the proposal changes after the yes, it has to be approved again.
One day an agent worked on the memory for hours from an external assistant, and its questions kept coming back empty, even about approved decisions. The decisions were there: the step that makes them searchable was failing silently, and an empty answer looked like missing knowledge. We fixed it the same day, and search now tells “there's nothing here” apart from “I haven't got there yet”. Then the opposite problem showed up: an off-topic question got the least distant decision, which was still wrong. Now, if nothing is relevant, search returns nothing, and we also test it with questions that have nothing to do with Yempik.
Decisions, research, brainstorms and hypotheses about where to take the company live in the same memory. We wrote down that “approved” means usable and nothing more: an approved hypothesis is still a hypothesis, and whoever summarises it, person or agent, has to call it that. The subtler risk was in research. An agent that links relevant evidence to one of our theses on its own turns research into confirmation. Today agents collect the evidence, look for contrary evidence too and cite the original external source, and a person decides whether a piece of evidence supports a thesis.
An automated routine lost access to a tool and logged the blocker in the memory several times, each time with possible ways out. Nobody saw it. In the end the routine itself concluded that the channel didn't reach anyone who could unblock it, and asked for a notification. Writing a problem into the right system feels like reporting it, and a blocker logged inside the same flow that stalled stays there. It's the point still open: giving every blocker an owner and a channel that reaches them outside the memory.
// one of the off-topic questions we used to test search: hamster food.