Where the memory of a conversation lives
The model remembers nothing. Each request carries the earlier messages along with the new message, and a limit decides how many fit. When the conversation grows past the limit, the oldest messages are dropped and the assistant answers without them.
Forgetting reads as a defect in attention, and it is a defect in arithmetic. Knowing which one it is changes what you do about it: repeating yourself louder does not help, and a short summary does.
Each request is a fresh start
The model is given text and returns a continuation. Between 2 requests it holds nothing. Everything that looks like memory is the product resending your earlier messages, which is why a long conversation costs more to run than a short conversation.
The limit is measured in tokens, not messages
A token is roughly a short word or a piece of one. The budget covers the instructions, the conversation, your new question and the reply together. This is why a long question can shorten an answer, and why a conversation full of long pasted text runs out sooner than a conversation of the same message count.
What falls out first
The oldest messages are dropped first on most services, which means the setup you gave at the start is the part most likely to disappear. That setup is usually the part that mattered, so the failure is not random: it removes exactly the context you would have wanted kept.
Working with the limit
Start a new conversation and open it with a 3-sentence summary of what matters. This costs less than arguing with a conversation that has already lost the beginning, and it puts the important context at the newest end where it survives. Keeping 1 conversation per task also helps, because the budget is spent on relevant text rather than on unrelated history.
Before you paste it in
Do some chatbots remember between sessions?
Some store notes about you and attach them to later conversations. That is a stored profile rather than memory in the model, and it is a separate feature with separate settings.
Does asking it to remember work?
Inside 1 conversation it makes no difference, because everything is resent anyway. Across conversations it works only when the product has a store, and then the store does the work.
Why does it sometimes recall the start of a long chat?
Because the conversation had not passed the limit yet, or the product summarises older messages instead of dropping them. Both behaviours exist and neither is announced.
Start a conversation, give it 3 sentences of context, and see how far it holds.
Open uncensored chatbot