What is worth keeping.

A conversation is a process. A conclusion is a result. On screen they look alike — both are text, both are true — but only one of them is worth anything in a month.

One question does most of the work: does a sentence survive being lifted out of the conversation? Read a line from the middle of a chat with everything around it removed. “Yes, that’s what I thought” survives as nothing. “That would push it into next quarter” survives as almost nothing — push what, into which quarter, from where. “The countertop quote came back at $9,400, about a third over the estimate” survives intact. It carries its own context. It can be handed to someone.

The same question explains why a summary is not a conclusion. A summary is a smaller version of the conversation, and it stays proportional to it: a long talk gets a long summary. A conclusion is proportional to what you learned, which is a different quantity altogether. Two hours can end in one line. Five minutes can end in three. If what you kept has the same shape as what you said, you kept the wrong thing.

Chats are filed by when they happened. Nobody thinks in that order — nobody thinks the Tuesday afternoon chat. They think the kitchen, the renewal, my mother’s care, the move. Filing by folder helps, and it is the right instinct. But a folder files the conversations, not what came out of them, and four transcripts in one place are still four transcripts.

Anything genuinely worth keeping takes many sessions. The kitchen took one conversation about layout, one about quartz against laminate, one late one about whether the contractor’s timeline was realistic, and a short one three weeks later when the quote changed. Filed by session, that is four records of nothing. Filed by the thing, it is one record that got better four times.

So the unit is the thing, and the word for it is a project — meaning the thing itself, not the chats about it, and not a project in the work sense either. A project is anything with moving parts that you will come back to. A renewal is a project. A diagnosis is a project. Choosing a school is a project. Name the folder you would make if you were going to make a folder. That name is the project, and everything you keep goes underneath it.

And because the record belongs to the thing and not to the session, it does not care which tool the session happened in, or whether there was a session at all. A phone call can update it. So can a letter, a price you saw in a window, a thing your neighbor mentioned. The record outlives the software it was made in, which is the only property that makes a record worth starting.

Almost everything worth keeping from a conversation is one of four things. The reason to separate them is not tidiness. It is that they end in four different ways.

A current fact is something true right now that could stop being true. Kitchen budget, $38,000. It has a name and a value; the name is stable and the value moves. Current facts end by being replaced.

A decision is a choice you made. Move the sink to the island. It does not expire and it does not need updating — a decision made in March is still a decision that was made in March, however much has changed around it. A decision ends by being reversed, which is a separate event and should be recorded as one.

An open question is something you do not know and need to. Does moving the sink need a permit? It ends by being answered.

A note is everything else that is true and useful and none of the above — a caution, a piece of context, something you understood and would rather not work out twice. Ramirez does the demolition himself; the tiling is subcontracted. A note ends by quietly ceasing to matter.

Notes are the widest of the four, which makes them the ones to watch — not because a note is second-class, but because it is the right home for a thing with no name and no date, and the wrong home for a thing you were in too much of a hurry to name. If it would sit under a name with a value, it is a fact. If it settles something, it is a decision. Keep the rest, and keep it short: a page of notes is a folder of prose again, and you have already seen how that ends.

The four are not subject categories. They are categories of how a thing dies — replaced, reversed, answered, outgrown. That turns out to be exactly the information you need in order to maintain a record without rereading all of it.

The instinct at the end of a good conversation is to keep all of it, because all of it felt valuable while it was happening. Almost none of it was. What you keep is worth less the more of it there is to read, and you will always be reading it in a hurry.

That gives you the only filter you need. Not is this true — most of it is true. Ask instead: would I want this handed to me ten minutes before the meeting? Everything else can go.

In practice, these are the things people keep and should not.

The reasoning, in full. Keep the reason; drop the derivation. One “because” clause, not the path you took to reach it.

Options dismissed early. Keep the one you nearly chose; the other five were never live.

Explanations written for you. If an assistant explained how something works and you understood it, the explanation was scaffolding and it can come down. If you did not fully understand it and will need it again, that is a note — and write it in your own words, so you find out now whether you understood it.

Anything you could look up in five seconds. A record that starts accumulating general knowledge has begun to rot, because the general knowledge crowds out the specific things that exist nowhere else.

Anything you could reconstruct exactly. Keep what you could not reproduce: what you were quoted, what somebody told you, what you judged, what you chose.

And one more, which is the one worth arguing about. Anything that keeps everything for you has made the decision you were supposed to make. Choosing what matters is not overhead on the way to a record — it is the work that produces one. It is four minutes at the end of a conversation in which you decide what actually happened. Skip it and what you have is another transcript, and you already have more of those than you will ever open.

Everything you keep is written for someone holding none of the context you have at this moment. That person is you, in eight months, and they will be a stranger. They will not remember the vendor’s name, which of the two quotes was the low one, or why any of it was urgent.

In practice:

• No pointing words. “It came in high.” “The new one.” “That approach.” Pointing works in a conversation because the thing pointed at is on the screen. In a record there is nothing to point at.
• No relative time. “Last week” is meaningless at any distance. Write the date.
• A name gets its job attached the first time. Raj is not a person to you in a year. Raj, in operations, who took the budget over from Elena in June is.
• No words that only exist inside the conversation. “Option B”, “as discussed”, “the above”. Option B has no meaning outside the message that listed A, B and C.
• Every number gets a unit and a subject. Not 38. Not $38k. Kitchen budget, $38,000, all in — or, once it has a name of its own to sit under, Budget — $38,000.

The question at the end is always the same one: could a competent colleague, given only this, walk into the room? If not, what you have written is a reminder — it points at knowledge you are still carrying in your head — rather than a record, which contains it. Reminders work for about a fortnight.

Facts want to be a name and a value, not a sentence.

Instead of “we’ve settled on quartz for the countertops, and the quote is in”:

Three things follow from naming.

The name is what you go looking for. A year from now you will not remember the number. You will remember that there was a number, and the name is the only handle you have on it. So use the word you would say out loud. Budget, not Financial parameters. Renews, not Contract renewal date.

A named fact can change without disturbing anything around it. A fact buried in a paragraph has to be found and the paragraph rewritten, and rewriting a paragraph is how a record silently loses the parts you did not mean to touch.

One name should hold one live value. This is the rule that improves your thinking rather than your filing. If you find yourself wanting to put two values under one name, you do not have a fact — you have either two facts that need separating, or a question you have not admitted is open. Both are much better discovered four minutes after a conversation than four months later.

The adjective in current fact is not decoration. It is a promise you are making to yourself about that value.

In a year you will not doubt what you decided. You will doubt whether you decided it well — and a record that holds only the choice has no answer to that. So the doubt sends you back to reopen the whole question, and you spend the hours a second time and land, usually, in the same place.

A decision that survives has four parts.

1. What you chose.
2. The one option you nearly chose.
3. The reason, with the condition that made it decisive.
4. The date.

The second version costs about twenty seconds more and does something the first cannot. Look at the condition: available now against a six-week wait. Conditions expire. A reason with a condition inside it tells you when the decision should be reopened — and, just as usefully, when it should not. Six months later, when somebody asks why you did not go electric, the answer is on the page and you do not have to re-argue anything to be sure of it.

One further rule, harder than it sounds: write the reason you would give to someone who disagreed with you. The reason you would give to someone who agrees is always shorter, and always missing the part you will need.

This is the practice people skip, and it costs more than any of the others.

A task is a thing you must do. A question is a thing you do not know. They feel alike in the moment and behave completely differently over time. A task is closed by effort, has an owner, and generates a small quantity of guilt while it sits there. A question is closed by an answer, which may arrive on its own — from a phone call, from somebody else, from a source you had not thought of. Much of what is genuinely open in any project is not yours to do. Waiting on the credit union’s rate is not a task, and filing it as one makes it look like something you are failing at rather than something you are waiting for.

So the rule is literal: write it with a question mark, and if it will not take one, it is a task and belongs wherever your tasks live.

Written that way, an open question can be closed by anyone who happens to find out — a colleague, a phone call, or an assistant you hand the project to later.

And then the reason this matters more than it appears to. An unknown you failed to write down is indistinguishable, later, from a thing you knew and forgot. That is the exact state that sends people back to redo work they have already done. Not I don’t know this, which is cheap. I don’t know whether I knew this, which is expensive, and can only be settled by starting over.

The moment to catch them is the end of a session, because that is the only moment you will ever hold a complete view of what you do not know about this thing. Ten minutes later it is gone, and you will not even remember having had it. So the last question at the end of a conversation is always the same one: what is still open?

The renewal comes back at $48,000 a year. The record says $56,000. The obvious move is to type over it.

Don’t. What you have afterwards is a record that is true about the world and false about your own history. It says the number was always $48,000, and that is not what happened.

The old value carries three things the new one cannot.

That it moved at all. A figure that dropped fourteen percent in a month is a different sort of figure from one that has not moved since January, and you cannot tell them apart by looking at today’s value.

Which way, and how fast. A trend needs at least two points, and you have just deleted one of them.

Whether to trust it today. A number revised three times is a number to check before you say it out loud in a room. A number that has stood still for a year is one you can quote.

There is also the plain practical case. At some point somebody who saw an earlier version will say wasn’t it fifty-six?, and you need to answer that in four seconds without doubting yourself.

The word for this is superseded, and it is a better word than old, because it says how the value ended. It was not wrong. It was replaced, on a date, by something that replaced it. So: the new value goes on top, the old one goes behind it with the date it stopped being true, and nothing is deleted for being out of date. Out of date is a status, not a reason for removal.

The same holds for a reversed decision, and more sharply. We are not doing X after all is meaningless if X is not still on the page. Delete the original and the reversal goes with it, and six months later nobody — including you — can tell whether the option was considered and rejected, or never considered at all.

Version history is not a new idea. A document keeps its revisions; anyone who writes code has a command that prints who changed a line and when. What almost nobody keeps is the same thing for the five sentences that actually matter, and that is the part you eventually want.

It buys three things.

Trust. You read a figure differently depending on whether you put it there yourself after a phone call, or it arrived out of a document you skim-read at an airport. Same figure, different weight — and you can only apply the weight if the record tells you which it was.

Repair. When something turns out to be wrong, the only useful question is where it came from, and the answer is a time and an author.

A short read. The least interesting of the three and the most used: knowing whether anything has moved since the last time you looked, so you can read four lines instead of the whole thing.

What is new is that who is not always a person. If you have ever pasted an assistant’s summary into something and kept it, a thing you did not write is now sitting in your record under your name. A record that cannot tell I wrote this after the call from this came out of a PDF I handed to an assistant is giving both the same standing, and they should not have the same standing. Not because assistants are wrong more often — because you check them differently. You know what you saw. You do not know what it read.

So the field has to be able to hold something that is not a person, and it has to be filled in every time, including when you are the one making the change. A record where only some changes are attributed is worse than one where none of them are, because it invites you to guess about the rest.

One small thing worth getting right: the date on a fact is the date it last moved, not the date you first wrote it down. Those are different facts and people conflate them constantly. Budget — $38,000 — 12 February should mean the budget became $38,000 in February, not that you started the project in February and the number may or may not have shifted since.

The instinct is to attach a source to an item: Budget — $38,000 (from the Ramirez quote). It works for about a month.

Then the budget becomes $41,000 and the citation stays where it is, and now the record asserts $41,000 on the authority of a quote for $38,000. Nobody did anything wrong. The source was attached to a thing that moves, and it did not move with it. Sources attached to items rot quietly, which is the worst way for anything in a record to fail, because there is no moment at which you notice.

The source belongs to the act of writing, not to the thing written. Every change had exactly one origin — this conversation, this document, this phone call, this email — and if you attach it there, the item accumulates a chain of origins that reads, in order, as the actual history of how you came to know something.

Two practical notes. A source is where it came from and when; that is the entire requirement. And a source that cannot be clicked is still a source. Conversations with assistants frequently have no address you can point at, and a phone call never does. The 14 March call with Ramirez is a perfectly good source, and it is worth a great deal more than the blank you would otherwise have left because there was no link to paste.

Write it in the same breath as the thing itself. Nobody in the history of records has ever successfully gone back and added the sources later.