Four thousand contacts and no idea who mattered

I met four thousand people on Mesh alone.

That number is not a flex — it is the setup for the problem. Because at the same time, the relationships I actually cared about were going quiet. Somebody I had a genuinely good conversation with in March, and then nothing, and by August it would have been strange to message them.

The contacts list was growing and the relationships were shrinking. Those turn out to be nearly unrelated quantities.

#What I tried first

A CRM. Several.

Every one of them is built on an assumption so deep that it is not stated anywhere in the marketing: a relationship is a deal in progress. There is a pipeline, and people move along it, and the terminal state is closed.

That model is correct for sales and completely wrong for the thing I needed. The people I was losing were not leads. There was no close. The whole value was that the relationship kept existing, and a system whose fundamental verb is advance to next stage has no way to represent "this person matters and there is nothing to do about it this month."

I tried using one anyway, with custom stages. It was miserable within three weeks, because every screen was asking me a question about my friends that only makes sense about prospects.

Then I tried notes in a document, which is what everyone does. That failed differently: it is write-only. Nothing ever surfaces. A note about someone is only useful at the moment you would have thought of them anyway, which is exactly the moment you do not need it.

#The one thing it had to do

Not store. Surface.

The insight, such as it is, is that relationship decay is not a memory problem. I have not forgotten these people. I could name every one of them right now, and tell you what we talked about.

It is a prompting problem — nothing in my week ever says this is the day. There is no meeting, no deadline, no notification. The relationship dies of nothing happening, and nothing happening never announces itself.

So the core of Roster is the follow-up: the system decides when a person should resurface, based on how close the tie is and how long it has been, and puts them in front of me with the context of the last conversation attached. Not a task list. A single name and enough detail to send something that is not generic.

Everything else in the product is downstream of that.

#What I got wrong

I built tiers first. Close friend, good friend, acquaintance, professional. Elaborate, satisfying to design, and I never maintained them, because categorising four thousand people is a project and nobody finishes a project about their own address book.

What replaced it is much dumber and works: how recently and how often you have actually interacted, which the system can observe without asking. The behaviour was already there. I did not need to ask people to classify their own friendships.

I over-indexed on capture. Early versions had rich note-taking, tags, custom fields. Almost none of it got used, including by me. If logging a conversation takes ninety seconds, you will not log the conversation, and then the system has nothing and decays into another empty tool.

Capture had to get down to a few seconds or it was not going to happen at all.

#Then it turned into a CRM anyway

Thirty-four professionals are in the beta now, and a real fraction of them use it for business relationships.

I resisted this for a while, because "personal relationship manager" was the whole thesis and CRM was the category I had defined myself against. That resistance was ego, not product judgement.

What the beta showed is that the distinction I had drawn is not the one users experience. A founder's most valuable professional relationships behave exactly like the personal ones — long-running, no pipeline, no close, and dying quietly of nothing happening. The salesperson's funnel is a different thing, and Roster is still bad at that and should stay bad at it. But everything above the funnel, the network a person actually operates from, is the same shape as friendship.

So it became a CRM, in the sense that people run their working relationships in it. It did not become a pipeline tool. Those are different concessions and I only made one of them.

#Where it is now

Beta, thirty-four people, built in public. Not finished — the capture flow is still slower than it should be, and the resurfacing logic is a heuristic rather than anything clever.

The measure I care about is whether I have had a conversation this month with someone I would otherwise have lost. Usage is a proxy, and a bad one.

This month, three.