When the Chat Closes, What's Still True?

A session can end with every word saved and the status of the work lost. By "session" we mean any working session with AI: a chat, a project with its own files and memory, or an agent run. What was decided, what's still open and who moves next need to live somewhere outside it. If those three facts only exist in the session, they aren't recorded. They're remembered, and only by the person who was there.

Find last month's decision

You can test this on your own work.

1. Pick an ordinary decision from last month: a scope change, a hiring call, a page that would or wouldn't ship.
2. Find where the decision is written down. Not the discussion: the decision.
3. Find what's still open, and who was supposed to move next.

If you can point to a line a colleague could read without asking you, that work is in good shape. If you find the path but not the landing, you're looking at the cost.

Nobody was careless. The sessions did their job: real decisions came out of them and the team acted on them. The missing line stayed invisible because the person who remembered was always around. The cost only shows when they're not.

Why the transcript isn't the record

A session is ordered by conversation. It records how the thinking moved: the question, the comparison, the aside, the next question. Everything stays equally present. The option you rejected sits in the same place, with the same apparent confidence, as the one you kept. Later lines don't overwrite earlier ones; they just come after them. Reading it back, you see movement, not state.

Status is different. It's three short facts, often one sentence, and that sentence is often not the last one. A colleague asking about the pricing page doesn't need the comparison. They need the chosen option, the open item attached to it, and the name of whoever does the next thing.

Long sessions have a second, more technical problem. Research on how language models handle long inputs finds that their reliability tends to drop as the amount of text in front of them grows, an effect often called "context rot" (Chroma, 2025; Anthropic, 2025). Anthropic's guidance recommends keeping structured notes outside the model's working context for long-running work, for this reason. A short, current statement of where things stand helps the assistant as much as it helps your team.

Projects and memory features help an assistant pick up where you left off. They don't, on their own, give a colleague a plain answer to "what did we decide?"

Three lines, where the team already looks

After a session produces a decision, write three lines:

• Decided: what is now true.
• Open: what's still unresolved.
• Next: who moves next, and on what.

Put them somewhere the team already opens: a shared document, a short standing note, or the top of the brief or plan. If there's already a brief for the pricing page, the decision belongs at the top of that brief.

Don't build a new home for this. The format matters much less than whether someone who wasn't in the session can find it without asking you. Write it as if you won't be in the room.

Open items and owners are the parts most often left behind. A decision without an owner gets rediscovered next month as if it were new. Put the open item next to the decision, and a name next to the open item.

Let the assistant keep the record

This is where a connected assistant earns its place. If your assistant can read and write files in a shared folder or project, you can ask it to finish every session by updating a decisions file with those three lines.

• The file sits in a shared place, so the whole team can read it.
• It's versioned, so you can see what changed and when.
• The next session, yours or a colleague's, starts by reading it, so the assistant begins from the current state rather than an old transcript.

That's persistent state: a small, current record that outlives any single session. A person still owns the decision and checks the lines are right. The assistant does the writing and the keeping.

Start with one decision

If last month's decision is already hard to find, start there. Write the three lines for that one piece of work, in the place the team will look. Then ask your assistant to keep them current from now on.

Keep thinking out loud in sessions; it's useful. It just isn't the record.

If you'd like to set this up once, on a live piece of work, that's what The Catalyst is: one real problem, built with you, and the keys handed over at the end.