Conversation Context
A new chat starts with no history of your earlier conversations. The model sees what is written in this chat and whatever the service added to it on its own.
ChatGPT, Claude, and Gemini have features that pull details out of past chats. When they are on, the service inserts selected details from old conversations into a new one, and it decides which ones.
These features are not on every plan, and they can be turned off in the settings.
Yesterday's work is safer to continue in the same chat: there the model sees the whole conversation.
Besides the conversation itself, the service has two more layers, and each gets its own chapter: what it remembers about you, and the shared materials of a project.
The length of a conversation eventually starts getting in the way. That is the next chapter.
One chat, one task
Everything you said inside a chat stays available until the end of it: conditions, corrections, the options you turned down. The model takes them into account in every answer that follows.
One recommendation comes out of that: one task, one chat. If you discussed the deck on Monday, continue in that conversation rather than a new one.
Instructions you gave at the start of a conversation hold to the end of it. In the deck chat you asked for short phrasing and a business tone. Ask there how to request vacation, and the answer arrives in the same clipped bullets, even though you never meant that for vacation.
The service keeps your history. You can come back to an earlier discussion and pick it up where you left off.