When the context is nearly full, the AI starts forgetting things. In most cases this is not a change in attitude, but a sign that earlier parts of the conversation have been truncated or compressed. The most reliable approach is to ask it to summarize the key points first, then start a new conversation, paste in the summary, and continue there, rather than repeatedly reminding it in the original conversation that it got something wrong.
Four signs the context is nearly full
The first sign is that it forgets the settings and names you gave at the start. For example, a character name, a form of address, or a formatting requirement you agreed on at the beginning suddenly changes later on, or disappears altogether. The second sign is that it brings up again a proposal you already rejected, as if that discussion never happened. The third sign is that its answers become noticeably shorter and more perfunctory: details and examples it could provide earlier gradually vanish, leaving only vague conclusions. The fourth sign is misattribution: it mixes up information from different paragraphs, different people, or different versions, and timelines and cause-and-effect start to fall out of alignment. If two or three of these appear, you can basically conclude that the problem is not that you were unclear, but that the conversation has simply become too long.
Why it starts forgetting after a long chat
The reason is that the context window is limited. Only a certain amount of text in a conversation can be visible to the model at the same time. The longer you chat, the more likely early content is to be pushed out, or compressed by the system into a shortened version, so details naturally cannot be preserved. Even if the content has not been pushed out completely, attention is diluted by the large amount of intervening conversation: information that is further away and mentioned less often is more easily covered over by what comes later. That is also why quality can still drop in the later stages of a long conversation even when the window is advertised as very large. For more on this, see A Bigger Context Window Is Not Always Better: The Attention Cost Behind 1M Tokens and Three Misconceptions.
Once it starts forgetting, this is the most reliable way to continue
First, ask the AI to produce a continuation summary, and explicitly require it to include four things: the current goal, the conclusions already decided, the questions still unresolved, and the key original wording that must be preserved, such as names, numbers, and fixed phrasing. Second, start a new conversation, paste in the summary first, add a line such as "Please continue based on this summary and do not overturn the conclusions already decided," and only then ask your new question. After pasting the summary, ask it to restate the goal and the decided conclusions first, and confirm there is no misunderstanding before moving on. That takes less effort than jumping straight to the next question, and it prevents it from continuing in the wrong direction based on a mistaken understanding. You can also save names, wording, and taboos that you use repeatedly as a separate piece of text, and paste it in whenever you start a new conversation, instead of relying on it to remember on its own. Third, split long tasks into stages wherever possible, give each stage its own conversation, and carry over only the summary and results from the previous stage, rather than cramming the whole job into a single conversation and dragging it out to the end. Fourth, put important settings at the start of each conversation, and restate them in good time when you reach key points. Fifth, do not mix several unrelated tasks into the same conversation: the more topics there are, the more the information interferes with itself, and forgetting and confusion will only arrive earlier.
Two boundaries that are easy to confuse
Paid versions usually have a larger context window and can last longer, but they still have a limit, and a long enough conversation will fill them too, so a subscription is not a once-and-for-all solution. In addition, the "memory" feature in many products is not the same thing as the context of a single conversation: memory stores a small number of preferences and facts across conversations, while context is everything in the current conversation that it can see right now. If this conversation forgets a setting you just gave it, that is usually a context problem, and toggling memory in the settings will not fix it.
