Conversations: chats, threads and compaction

You have been in one chat since breakfast: a renewal letter, three drafts of it, a spreadsheet the assistant read on the way, and a supplier's website it looked at to check a price. Somewhere in the afternoon the small ring beside the model name in the composer fills up and changes colour. Later, a line appears across the conversation reading Context compacted, and the assistant carries on as though nothing had happened.

That ring and that line are the two things worth understanding about conversations in Player One. Most of the advice about how to work with it follows from them.

A conversation is what your assistant can see at once

Every message you send goes to the model with the whole conversation attached. That is why it can pick up something you said an hour ago without being reminded, and why it never asks you to repeat yourself inside one chat: it reads the lot, every time.

There is a limit to how much it can read at once. That limit is the context window, its size depends on which model the chat is running, and the models differ a great deal. Hover the ring in the composer and you get the numbers, the model, and a one-line verdict: plenty of room remains, context is filling up, or context is nearly full.

The part people get wrong is assuming a conversation is only what they typed. It is not. Files the assistant read, pages it fetched, results a tool handed back and its own replies all sit in the same window. A chat with three short questions in it can be far heavier than it looks, if answering them meant reading a long document.

What compaction does

When a conversation stops fitting, Player One summarises the older part of it and carries on in the same chat. That is compaction. It happens on its own, at a safe point once whatever the assistant was doing has finished, so it does not interrupt a reply half way through.

What you see is a divider in the thread reading Context compacted, with a rough figure for what it saved. Click it and it opens out: the size of the older section before and after, the saving, and the summary itself, which you can copy. While the work is running, a pill in the thread says Compacting the conversation.

What changes is what your assistant is working from. The recent part of the conversation stays exactly as it was, word for word. Everything before the divider is now that summary, so a small detail from three hours ago may come back paraphrased, or not at all. The figures on the divider describe only the section that was replaced, not the whole chat.

What does not change is the chat itself. Nothing is deleted, the whole thread stays on screen for you to scroll through, and search still finds messages inside the compacted part. Two consequences follow. You cannot delete a message inside a compacted section, because it is one of the things the summary was written from; to get behind it, rewind the chat to a point before the divider. And on a long chat the model picker greys out any model whose window is too small to hold the conversation, which is usually the first sign that a thread has grown large.

Compacting on purpose

Click the ring and Player One asks whether to compact now, and is honest about the trade: it frees room, but that summary becomes the working memory of everything before it. You can answer Compact now, New chat or Not now. Typing /compact does the same job without the dialog, and /context tells you where you stand.

Since it happens automatically anyway, compacting by hand is only worth it when you want a clean break, or you are about to move the chat to a model with a smaller window.

The better move is often the one the dialog itself suggests. If the thread has taught your assistant something worth keeping, ask it to write that up as a skill before you go, then start fresh. The lesson outlives the conversation, and the new chat begins with room to work. See Teaching it: from a conversation to a skill.

One chat, one piece of work

Given all of that, the practice is straightforward.

Start a new chat when the subject changes. A chat that has carried the renewal, then a payroll question, then a birthday message is expensive to run, awkward to find later, and its summary will be a muddle of three unrelated things. New chat in the sidebar (/) opens one with whichever agent the sidebar is pointed at.

One chat per piece of work, not per day. Returning to yesterday's chat about the same job is exactly right: everything it learned is still in front of it, and any unsent text you left in the box is still waiting for you.

Push side work off the main thread. Anything that means reading a great deal to produce a little, such as research, a sweep through files, or a long build, is better handed to a helper or a specialist agent. It does the reading in its own conversation and brings back the answer, and your chat pays for the answer rather than the reading. This is the single most useful habit for keeping a main chat healthy, and it is covered in Agents, delegation and team chat.

Do not use chats as storage. Anything you want back reliably belongs in memory rather than in a thread you hope to find again. See Memory, explained.

Finding a conversation again

The sidebar keeps folders first, then pinned chats, then everything else grouped by day. Pin the two or three you are living in. File the rest into folders, which nest as deeply as you like and are saved for you rather than for one browser, so they follow you to another machine. The cog at the top of the list chooses what it shows: under Chats, Lead by default, the chats you started, with the ones your agents started for them kept out of the way, or Helpers, Scheduled runs, Flow runs, API chats or All; under Status, Unread, Read or Any. Mark all read clears the unread marks on what is on screen, and Reset puts the list back to its default.

When a chat is done but you would rather not lose it, Hide from sidebar in the chat's menu removes the row and nothing else: the chat still opens, and search still finds it. Delete is a different thing, and moves the chat, along with any conversations it spawned, to Trash (/trash), where you can preview it, restore it, or remove it for good.

To find something you have mislaid, Search (/search) covers chats, jobs, agents and files together, and can be narrowed to chats alone. Failing that, just ask: your assistant can search your past conversations and tell you which one had the answer.

Where this goes next

Chat describes the surface itself and what each control does. Tokens, cost and caching explains why a long conversation costs more to run than a short one, and what the platform does about it. Organising with projects covers the scope that folders, search and Trash all sit inside.