I Kept Losing the Work Between the Work

The call ended at 4:52pm. I stood up from the desk, walked down two flights of stairs and out onto the street, and by the time I got to the bike I had made four commitments out loud that did not exist an hour earlier. Chase the staffing plan. Get the revised scope back to the client by Thursday. Ask Bo whether the migration blocks anything. Follow up on an invoice that had been quietly ageing for six weeks.

Four things. Small things. The kind of thing any competent operator handles without a second thought.

I got home an hour later and remembered two of them.

This was not a one-off. This was every day for about nine years. I run an engineering consultancy with teams in Vietnam and a commercial entity in the Netherlands. I have a small hot sauce brand on the side that is more work than it sounds. Before all of that I spent years in operations roles at two London startups that both got acquired, which is a polite way of saying I spent years being the person who was supposed to remember everything.

And the thing I kept noticing was this: I did not have a productivity problem. My system was fine. I had a capture problem. The work was not falling apart in the doing. It was falling apart in the four-minute gap between deciding to do something and being anywhere near a keyboard.

The capture tax

Every task manager ever built assumes you are sitting down.

That is the unspoken premise. Todoist, Things, Asana, Notion, all of them. They are beautifully designed for a person at a desk with two hands free and a clear head, deliberately entering a task they have already fully formed in their mind. Under those conditions they are excellent. I used Todoist for four years and I have very little bad to say about it.

But that is not when tasks arrive. Tasks arrive:

• Walking out of a meeting

• On the bike

• Ten seconds after you hang up

• In the shower, which nobody has solved and I am not claiming to

• Mid-conversation, when someone says something that quietly creates three days of work for you

• At 11pm when you are half asleep and your brain finally lets go of something

In every one of those moments, the cost of opening an app, tapping into a text field, typing out a coherent task title, picking a project, setting a due date, and saving it is higher than the cost of just trusting yourself to remember. So you trust yourself to remember. And you are wrong maybe a third of the time, which is not enough to feel like a crisis and is absolutely enough to slowly erode the trust your team has in you.

I started calling this the capture tax. It is not that the tools are slow. It is that they are slow relative to the moment, and the moment lasts about eight seconds.

Voice was everywhere and nowhere

The obvious answer is voice. I am not the first person to think this.

So I tried the obvious things. Siri into Apple Reminders. The voice button inside Todoist. Dictation into a notes app and a weekly ritual of cleaning it up. Voice memos, which is where good intentions go to die.

They all failed in the same way, and it took me an embarrassingly long time to see the pattern. In every one of those tools, voice is a transcription feature bolted onto a typing product. You speak, it turns your words into a string, and then it drops that string into a text field that was designed for someone typing carefully. So you end up with a task called "yeah um chase Hoang about the Q3 staffing thing before Friday and loop in Dale I think" sitting in your inbox with no due date, no project, and no owner, and now you have to go and fix it. Which means you have to sit down. Which is the exact thing you were trying to avoid.

The tools were not wrong about voice. They were wrong about where the hard part is.

Transcription is solved. Interpretation is not.

Speech to text is essentially finished as a problem. It has been for a couple of years. The models are fast, they are accurate, they handle my London accent and they handle it in a noisy café in Da Nang.

The hard part, the part nobody had built a product around, is what happens next.

Take that sentence again: "Chase Hoang about the Q3 staffing plan before Friday and loop in Dale."

To a transcription tool that is one string of text. To a human it is obviously structured:

• Task: chase the Q3 staffing plan

• Person: Hoang

• Deadline: Friday, which means the 15th, which means it is due in two days

• Stakeholder: Dale

• Project: whatever bucket Q3 staffing lives in

A human reads that structure instantly and without effort. And for the first time, so can a machine, reliably enough to build a product on. That was the unlock. Not voice. Voice was available to everyone. The unlock was that language models got good enough that you can speak a messy human sentence and have something on the other end actually understand what you meant, not just what you said.

That is what Meridian is. You talk. It listens, works out what you actually meant, and files it properly. No text field. No cleanup pass. No sitting down.

What that forced in the build

Deciding voice is the primary input rather than an alternate input changes almost everything downstream, and most of the changes are unglamorous.

Latency became the whole product. If capture takes longer than about two seconds from finishing your sentence to seeing the task appear, the magic dies and you go back to not bothering. That single constraint dictated the entire architecture: streaming transcription rather than record-then-upload, optimistic local writes, AI calls proxied server side so the app is not waiting on multiple round trips it does not control.

Offline had to work. You do not get to tell someone their task manager is unavailable because they are in a lift or on a plane or in a stairwell with bad signal. Capture works offline and reconciles later. This was not a v2 feature. A capture tool that fails at capture is not a tool.

The AI had to be invisible. There is no chat window in Meridian. There is no assistant persona. Nothing calls itself an agent. The intelligence sits entirely underneath the surface doing one job, which is turning speech into structure. I feel strongly about this. The current wave of products treat AI as the interface, which mostly means adding a chat box to something and asking the user to do prompt engineering to complete a task they could have done with a button. The interesting version is the opposite: AI as the thing that removes the interface entirely.

Trust had to be earnable. If the parsing is wrong even 15% of the time you stop speaking freely and start dictating carefully, which is just typing with your mouth and is somehow worse than typing. Getting from "impressive demo" to "I trust this with a commitment I made to a client" was most of the last six months of work and almost none of the fun part.

What I got wrong

The first version tried to be a second brain. Notes, links, projects, references, a whole knowledge layer. I built quite a lot of it before admitting that I was building the thing that already exists in twenty flavours rather than the thing I actually needed.

I cut it. Meridian does one job. It gets things out of your head and into a system, correctly structured, in the eight seconds you have available. Everything else was noise.

The second thing I got wrong was assuming this was a personal itch that maybe forty other people shared. It is not. Every operator, founder, consultant, contractor, parent, and generally overcommitted person I have shown it to has the same reaction, which is not "nice app" but "wait, can I keep this."

What Meridian is not

I would rather set this out plainly than have you find out after a trial.

It is not an AI scheduler. It will not rearrange your calendar or decide what you should be doing at 2pm. Motion does that and does it well, and it is a genuinely different product for a genuinely different problem. Meridian is upstream of that. It is about getting the thing captured at all.

It is not free, and it is not going to be. It is $14.99 a month or $129 a year with a seven day trial and no card required to start. Every capture costs me real money in transcription and inference. A free tier would mean either degrading the thing that makes it work or selling something else, and I am not doing either.

It is not finished. It is on iOS, macOS and the web today. What is live is the core loop, done properly, rather than a wide surface done thinly.

Where this goes

The direction is a single place where everything you have committed to actually lives, whether you spoke it, typed it, or it arrived in an email. Capture is the wedge, not the destination. But capture had to be right first, because a hub full of half-formed rubbish is worse than no hub at all.

I am building this on my own, in the evenings and the gaps, separately from the consultancy. That is partly a legal and structural thing and partly because I wanted at least one thing where I could make the call in ten seconds without a meeting.

If any of this sounds like your week, the trial is seven days and the first thing I would ask you to do is not open the app at your desk. Open it walking out of your next meeting, say the thing you would otherwise have promised yourself you would remember, and see if it is still there and still correct when you sit down.

That is the whole pitch. Everything else is detail.

Try Meridian free for 7 days

Will Macfarlane is the founder of Meridian. He lives in Vietnam, and has forgotten more tasks than he cares to admit.