Knowledge

What just happened

You scrolled back through a long chat and hit a dashed rose line with a small chip that reads Context compacted. Nothing is broken, nothing was deleted, and nothing needs your attention. Here is what it means.

Why conversations get compacted

A model answers by reading the conversation so far, and there is a limit to how much text it can take in at once — its context window.

When a conversation grows long enough, Lexika distills its older part into a summary and hands the model that summary in place of those messages. The conversation keeps going, and the model keeps answering with an understanding of what came before — just in condensed form for the oldest stretch.

This is something Lexika does on its own. There is no button that starts it, no setting that configures it, and no switch that turns it off.

What is preserved, and what the model reads

The dialog behind the marker puts it in one sentence:

Messages up to March 14 at 09:20 were distilled into this summary — the model reads it in place of the original messages, which stay visible in the transcript.

The date is the moment the summary reaches up to, shown in your own time format.

Two separate things, worth separating clearly:

	You see	The model reads
Above the seam (older messages)	Every original message, in full	The summary
Below the seam (newer messages)	Every original message, in full	Every message, verbatim

Your transcript is untouched. Scroll up as far as you like — every question you asked and every answer you got is still there, word for word. Compaction changes what gets handed to the model, not what the conversation still shows you.

The cut-off point is a moment in time: the timestamp of the last message included in the compacted context. Messages sent at or before that moment are covered by the summary. Everything after it reaches the model as written.

A conversation can be compacted more than once as it grows, and the boundary moves forward when that happens.

The signals in the interface

All of them share the same rose accent, and the seam chip, the floating button and the summary dialog carry the same history glyph — a clock with a backwards arrow.

• The seam row. A dashed rose line running in from both edges to a small centred pill reading Context compacted. It sits directly after the last message covered by the summary. Click the pill to read the summary.
• The dashed edge line. Covered messages carry a short dashed rose rule up the transcript's end edge — the right side in English, the left in Arabic. Stacked together they read as one continuous line, showing how far back the summary reaches. Hover it and a small tooltip reads Context compacted. It is a passive marker: it names itself and nothing more.
• The floating indicator. When your whole view is above the seam, a small round button with a rose outline and the history glyph appears at the bottom of the transcript, beside the jump-to-latest arrow. Its label reads You are reading summarized context. Clicking it opens the summary.
• Rose in the question index. In the index along the transcript's end edge, questions inside the summarized region are drawn in rose instead of the usual neutral — held back at rest, brighter as you point at them — so a glance tells you how far back the summary reaches.

Worth knowing

You will often see no compaction marker at all, even in a compacted conversation. A chat opens on its most recent messages, and if none of them is covered by the summary there is nothing to mark. Every one of these marks only appears once you have scrolled back far enough to reach a covered message.

Reading the summary

The summary opens from the seam chip or the floating indicator — not from the dashed edge line.

The dialog is titled Compacted context. It shows the sentence quoted above with the cut-off date filled in, then the summary itself as plain text in a scrollable box, then a Close button. That box is the entire thing the model reads in place of your older messages, shown exactly as it was written.

Getting good results in a long conversation

• Read the summary if an answer feels off. If the model seems to have lost a detail from early on, open the dialog. If the detail is not in there, restate it in your next message.
• Restate, don't assume. Anything defined far back in a compacted conversation — a constraint, a file name, a decision — is worth repeating once when it matters again.
• Say the durable things out loud. A fact that matters beyond one thread is worth stating plainly rather than leaving implied, so it has a chance to reach the project's memory instead of living only in an old message.
• Start a new conversation when the topic changes. A new conversation in the same project starts with nothing to summarize, and still has the project's Brain behind it. Long single threads are best kept to a single line of work.
• Recent detail is safest. Compaction always covers the oldest part of a conversation. The stretch you are actively working in is read verbatim.

In short

The rose marker is Lexika telling you where the conversation switched from being read verbatim to being read as a summary. Your messages are all still there. The model is still working from what came before — as a summary above the seam, and word for word below it.