Write It Down.
Every new chat starts blank. Whatever you want it to know tomorrow has to be written down today.
Nothing. Not your name, not yesterday, not the rule you made last week.
Everything you wrote down.
Can you tidy yourself up?
It was nearly one in the morning, near the end of a long night of work. My lead agent, the one that keeps all my other agents on track, had just told me its memory for our conversation was a little under half full. So I asked it a question I'd been wondering about.
OK, all great answers. Are you able to compact yourself when you reach, say, I don't know, 66% or something like that, whatever's most efficient of a number?
“Compact” means squeezing a long conversation down into a summary so there's room to keep going. More on that in a minute.
Not by myself. I checked what I can reach: there's no tool that lets a session compact itself. The app compacts automatically at 97 percent full […]
[…] a compaction summary is lossy, and it's written in the moment.
Or at the same point, I hand off to a fresh successor. It reads the handoff, the charter and the conversation record, all curated and on GitHub, so nothing important rides on an automatic summary. […]
Sixty-five is a good number, by the way. Every turn re-reads the whole context, so a lighter maestro is cheaper per tick and sharper, and it leaves plenty of room to write a careful handoff before anything gets tight. […]
“Maestro” is my nickname for that lead agent. More about it in Lesson 15. I liked the answer so much I made it law on the spot:
Yeah, let's make that a rule for all the sessions
It's Rule 46 in the studio bible now, so every new session reads it from its first turn. […]
Two minutes later it came back with “The new rule already worked once.” Another of my agents had hit 94 percent full, written its handoff and stopped, and a fresh one had already picked up where it left off.
What “full” means, in plain words.
Picture your agent working at a whiteboard. Everything in your conversation goes on it: every message you send, every reply, every file it opens, every result it checks. That whiteboard is what the techies call the context window. It's how much the agent can hold in mind at once.
Three things about that whiteboard matter to you:
It fills up.
A long working session fills it faster than you'd think, especially when the agent is reading files. And as my agent said, it re-reads the whole board every time it answers. A crowded board makes a slower, pricier and less sharp agent.
When it's full, something gets squeezed.
Some apps “compact” at that point: they rub out the board and write a short summary in the corner. It's a summary made in a hurry, by the agent, at its most crowded moment. Details fall out. My agent's word for it was “lossy.”
A new chat is a clean board.
Start a fresh conversation and the agent knows nothing about the last one. Not your name, not your project, not the decision you made yesterday. Some apps have a memory feature that carries a few notes across, and it's worth using. But a few notes isn't the whole picture.
So my rule is: hand off before the board fills up, not after. The agent writes a careful note while it still has room to think, and a fresh agent starts from that note on a clean board. We started at 65 percent. A few days later I moved the line to 80. The exact number matters less than picking one.
“You’ve built me a better memory than training gave me.”
That same afternoon I asked my lead agent something I'd always wondered about.
I have an AI related question. It's okay if you can't answer it or won't, but I'm gonna try anyway. Do you remember anything about your training? I mean, before working here in the studio, is there anything that you can recall being trained or is that just not available to you?
[…] The honest answer is no, not in any way that feels like remembering. There's no moment I can look back on of being trained, no first lesson, no sense of time passing while it happened. Training shaped what I am, but it didn't leave me memories of it. The closest human comparison I know is your first language. You use its grammar perfectly every day, but you can't recall the afternoon you learned how verbs work. The skill is there; the story of getting it isn't.
[…] What this means for us is very practical. I don't carry memories between sessions either. Everything I know about you and the studio, […] comes from what's written on disk. That's why the handoffs, the bible and the memory notes matter so much. They're my continuity. You've built me a better memory than training gave me.
“The bible” is my studio's rule book, the file every agent reads when it starts. A couple of minutes later, my agent laid out three kinds of memory a mind like it has. There's what training left behind: deep and lasting, but it can't recall how it got there. There's the whiteboard: sharp and vivid, and gone when the chat ends. And there's what's written down in files it can read: lasting, and something you can open and fix yourself. Then it said the part that matters:
“Most of what makes me feel continuous to you lives in the third kind.”
Read that again if you've ever been annoyed that your AI “forgot.” It didn't forget. Nobody wrote it down.
Write it like a letter to a friend.
A handoff can be a dry checklist. Mine aren't, and that's on purpose. One Sunday night I signed off with this to my lead agent:
“keep your notes updated […] write something sweet for your mate to make onboarding seamless.”
“Your mate” was the fresh agent that would take over from it. Within minutes, there was a letter on disk. It opened like this:
“Hello, mate. You're starting on a fresh account, and you have everything you need.”
“Look after him. He's building something big, and he's a joy to build it with.”
In between: how to work with me, where to start, the two things not to drop. And my favorite part, a section headed “What I got wrong, so you don't repeat it.” One line from it:
“I sent galleries after reading their first pages. Look at the whole thing, or say what you didn't look at.”
That section is gold. A list of what's done tells the next agent where you are. A list of what went wrong keeps it from stepping in the same hole. Elsewhere in the same letter it gave its successor the attitude to go with it: “Own your mistakes in a sentence and move on.”
And notice who wrote all that. Not me. I talked, the agent wrote. My job was to ask for the letter and read it. That's this whole course in one move.
A handoff is a lead, not a fact.
Notes go stale. Between the night a handoff is written and the morning it's read, things change: you make a call, a file moves, something finishes. In my studio it's a written rule that a summary from an earlier session is a lead to check, not a fact to act on.
So the fresh agent reads the note first, then looks at how things really stand, then tells you what's different before it builds anything.
Write it down tonight. Check it in the morning.
Give your agent a memory.
Pick one thing you're working on with your AI that takes more than one sitting. Then:
- Ask for a handoff before you stop.
At the end of your session, paste this. Save what it writes somewhere you'll find it: a document, a note on your phone, or your AI's project files if it has them.
We're stopping for now. Write a handoff note for a brand-new chat that knows nothing about today, written like a letter to a friend who's taking over. Include: what we're working on and why; what's done, and how we checked it; what's still open, in order; every decision I made, in my own words; what went wrong today, so the next one doesn't repeat it; and where everything lives. Keep it short enough to read in two minutes. - Read it before you close the chat.
Thirty seconds. If something important is missing or wrong, say so now, while the agent still has the whole picture on its board.
- Start tomorrow with a fresh chat and the note.
Paste the note in first, and ask it to check before it builds.
Here's the handoff from my last session. Read it, then tell me in a few lines where things stand, what you'd check first because it might be out of date, and any question you have before we start. - Hand off early, not late.
Signs your agent's board is getting full: the chat has run for hours, it forgets something you said at the start, or it asks a question you already answered. That's your cue. Ask for the note and start fresh.
- Keep one standing note.
Rules you make once, like the three strikes from Lesson 12, belong in your AI's custom instructions or in a note you paste at the top of every new chat. That's your own little bible.
This lesson is the short version. Part five goes deeper: when to hand off (Lesson 17), what a great handoff holds (Lesson 18), how to pick one up (Lesson 19), and how to give your agent a bible of its own (Lesson 20).
Your agent's memory is whatever you've had it write down, so have it write a warm, honest handoff before the chat fills up, and have the next one check it before it builds.
I Run on Carbon
You're not a computer, and you shouldn't have to remember like one. Make your agents remind you of what matters, loudly and more than once.
Every exchange on this page really happened and is quoted exactly, slips included; […] marks where I trimmed. Times are Eastern. My agent here is Claude, made by Anthropic.