Preserve Thinking
Preserve Thinking refers to the setting whereby an AI chat program retains its internal intermediate reasoning from earlier responses and reads it back in when processing the next question. This keeps the chain of thought intact across a longer conversation instead of starting fresh with every question.
Modern chat programs like ChatGPT or Claude often don’t write their answer immediately. Beforehand, they generate an internal text in which they break down the problem and jot down intermediate steps. This intermediate text is usually not displayed at all, or only shown in summarized form. Preserve Thinking means: this intermediate text is not discarded after the answer, but saved. If a follow-up question is then asked, the program reads the old reasoning back in. It thereby knows not only what it answered, but also why.
Why a discarded chain of thought becomes costly
Without this setting, the same thing happens with every new question: the intermediate reasoning from the last round has vanished. The program only sees its own finished answer. Everything it checked along the way and then discarded is gone. It can therefore walk into the same dead end a second time.
An example from programming makes this clear. The AI is supposed to find a bug in a program and, along the way, rules out three possible causes. In the end, the answer only states the fourth cause. If you now ask a follow-up question, without Preserve Thinking it starts over at cause one. With the chain of thought preserved, it picks up right where it left off.
The price for this is space. Every model has a limited short-term memory, the context window. Only a fixed amount of text fits in there. Saved reasoning is often longer than the visible answers and fills up this window quickly. Whoever keeps it has less room for the actual conversation history.
What exactly gets carried along
Technically speaking, a conversation with an AI is a long list of text. With every question, the model is presented with this entire list again. It has no memory in the human sense. It simply reads everything it’s given anew each time.
By default, many providers trim this list. The internal reasoning of older rounds gets cut out, while the answers remain. If Preserve Thinking is active, the reasoning stays in the list as its own section. Developer interfaces often let you configure how many rounds back this should apply to.
A compromise is frequently used. The reasoning from the last one to two rounds is kept in full. Older reasoning is condensed into a few sentences or deleted entirely. This way, the current thread of thought stays intact without overflowing the context window.
Where the setting shows up
In the normal chat window, Preserve Thinking is rarely visible as a toggle. There, the provider decides what gets saved. It becomes visible in developer tools, for instance in coding assistants like Cursor or in the interfaces from OpenAI and Anthropic. There, the option often sits right next to settings for the length of the chain of thought.
It is especially important for agents. These are AI systems that independently work through a task over many steps, such as searching through files and then editing them. Such an agent makes decisions in step three that still need to hold in step twelve. Without preserved reasoning, it contradicts itself more often along the way.
The term appears in billing statements and technical articles because of the costs involved. Every text block that gets included is paid for, including the old reasoning. Whoever keeps it permanently pays noticeably more per question. A common misconception, moreover, is that Preserve Thinking is a long-term memory spanning different conversations. It only works within a single ongoing conversation.