Your inbox is a ledger. Nobody is keeping it.
Every message you send or receive changes what you owe and what you're owed. Most people track that in memory — which is exactly where commitments go to die.
Three columns, and that's all
Strip the context out of a mailbox and almost everything left is one of three things. Not topics. Not projects. Obligations, each pointing in a direction.
Asksthey want · you owe
Someone needs something from you. You haven’t answered yet, and every hour it sits there it gets more expensive to answer well.
Commitmentsyou promised · you owe
You said you’d do it — sometimes in reply to an ask, more often volunteered halfway down a paragraph about something else. It isn’t done.
Waiting onyou asked · they owe
You asked someone for something and they haven’t delivered. The column people forget entirely — and the one that quietly blocks the most work.
An ask is not a commitment. Someone asking you for a proposal, and you agreeing to send one, are different states with different consequences — and the moment between them is where most dropped work happens.
Why a summary doesn't fix this
An AI summary tells you what a thread says. Useful, and increasingly free — your mail provider probably does it already.
What it cannot tell you is that on Tuesday you promised a client the revised scope, that it's Friday, and that nobody has mentioned it since. That fact doesn't live in any single thread. It lives across your mailbox, and it only exists if something has been keeping count.
Summaries are recomputed from scratch every time you open something. They're worth the same on day ninety as on day one. A ledger accumulates. It's the one thing in an inbox that gets more valuable the longer it runs, because what it tracks is state rather than content.
Two events, one model
The ledger stays current because every message is an event that can change it. Not a summarisation pass — a set of questions with a small, closed set of answers.
Because the answers are constrained, the ledger changes only when something in the message actually changes it. A reply that says thanks, got it closes nothing. A reply that says here's the file closes the thing you were waiting for.
Every line points at a sentence
The way a tracker like this fails is not by missing things. It fails by inventing them — turning happy to help with that into a commitment you never made. Once that happens twice, you start checking every entry against the thread, and at that point reading your inbox was faster.
So nothing enters or leaves the ledger without the words that caused it.
Book flight tickets from Hanoi to Saigon on 1st September
“Pl. find the flight tickets attached”
That quote comes verbatim out of the message that closed the entry. You can check it in one click. When Compose is wrong, it's visibly wrong — which is the only kind of wrong you can live with in software that keeps score on your behalf.
The moment that matters
Nothing here needs you to file, tag, snooze or triage. You asked someone for something nine days ago. This morning they delivered it, buried in the third paragraph of a reply about something else.
asked 20 Jul · re: Q3 engagement · 9 days open
Closed — they provided it. “revised SOW is attached, ignore the earlier one”
You didn't open that email. You didn't remember you were waiting. The column just got one line shorter, and the person you were about to chase is no longer someone you need to chase.
What a kept ledger is actually for
Once the state exists, the useful things stop being about email at all.
Walking into a meeting. Not a summary of recent threads — what each person in the room owes you, what you owe them, and what's still open between you. Including the thing you promised six weeks ago and haven't thought about since.
Seeing what's blocked. Your waiting-on column is a dependency list. It has always been one. It has just never been written down anywhere you could look at it.
Knowing what to chase. Most people over-chase the recent and forget the old. A ledger has no recency bias.
It won't be right every time
Email is written by humans and humans are ambiguous. I'll look into it might be a commitment. Happy to help usually is not. Compose will get some of those wrong, and anyone telling you their tool won't is selling you something.
What it will be is checkable. Every entry carries the sentence that created it and the sentence that closed it. Nothing is asserted that you can't verify in one click.
Use it for a week on your real mailbox. If it's catching things you would have dropped, we should talk. If it's getting something wrong, tell us where it broke — that's a conversation, not a support ticket.