Why does the AI forget your opening conditions as the chat grows long?

#The condition you set an hour ago quietly drops out

You have spent an hour going through quarterly results with an AI. At the start you agreed to "leave out returns and look at domestic branches only," but at some point returns start showing up in its numbers. Point it out and it corrects itself; a little later it forgets again. Anyone who has had a long conversation with an AI has probably run into this.

It happens because the conversation is approaching the amount the AI can read at once, and the opening, where you set the condition, has been cut off. When a conversation reaches that point, OwlAI condenses the earlier part into a summary instead of silently cutting it.

#The AI rereads the whole conversation for every question

An AI doesn't remember a conversation. Each time you ask something, it rereads everything said so far, from the beginning, and then answers. But there is a cap on how much it can read at once. As a conversation nears that cap, something has to go, and if the oldest part is dropped without any notice, the conditions set at the start are the first to disappear. This is one reason an AI loses track of what was said earlier in a long conversation.

#The earlier part becomes a one-page summary, like meeting minutes

Much as a long meeting ends with a summary of the minutes, OwlAI condenses the earlier part of a conversation once it reaches a set point. There are two ways this happens, depending on the provider serving the model. If the provider offers server-side compaction, OwlAI uses it as is. If it does not, OwlAI compacts the conversation itself with its own mechanism.

When OwlAI compacts the conversation itself, what the summary must keep is fixed: the goal you originally had, the conditions agreed and facts confirmed along the way, the queries run and their key results, approaches that were tried and failed and why, and the work in progress with its next step. What goes first is anything that can be regenerated, such as raw rows from query results or repeated explanations. The most recent exchanges and results are left untouched. However many times compaction runs, the goal and conditions are carried into each new summary.

When compaction happens, a Conversation cleanup step appears in the thinking process above the answer, so you can see on screen that the AI has summarized the earlier part and is carrying on.

#A gauge shows when compaction is coming

As a conversation grows, a context gauge appears next to the input box. Click it to see the current usage and the maximum as numbers. That maximum is exactly where compaction happens, and the gauge appears once the conversation reaches half of it.

Because the screen and the internal behavior read the same value, the gauge never fills up without compaction happening, and compaction never kicks in out of nowhere at the halfway mark. When the gauge is nearly full, expect compaction soon.

#When the topic changes, start a new conversation

Compaction runs automatically, and the gauge tells you in advance when it will happen, so there is nothing you need to do about it. OwlAI's own compaction is built to trim details such as raw rows first and to hold on to the goal and conditions to the end.

Two things are still worth knowing. If the topic has changed completely, starting a new conversation is cleaner. And if you switch models in a conversation that has been summarized, the new model rereads the conversation from the beginning, which increases usage. That is why a confirmation dialog appears before the switch.