Skip to content
LyraDocs

Context compaction

Long conversations eventually outgrow the model’s context window. Lyra summarizes the older part for you, and the original stays in the session log.

See what’s using it

Section titled “See what’s using it”

The ring to the left of the model name in the composer shows how much of the context window this chat fills. Click it for a breakdown: messages, built-in tools, MCP tools, the skill list, the system prompt, project instructions, project memory, and what’s left.

The numbers come from the last reply’s actual usage, estimated until there’s been a reply. Past 80% the ring turns red.

Automatic compaction

Section titled “Automatic compaction”

Three steps, cheapest first:

  1. Trim oversized tool results. No model call needed, and often enough on its own.
  2. Summarize. At 80% of the window, older turns are replaced by a structured summary and the most recent ones are kept word for word.
  3. Recall on demand. The recall tool reads the original log back whenever it’s needed, so nothing summarized is really lost.

Compacting by hand

Section titled “Compacting by hand”
  • Type /compact, optionally followed by what the summary should keep.
  • Or pick compact under Actions in the @ menu.

Doing this between tasks frees up room without losing what you concluded.

Choosing the summary model

Section titled “Choosing the summary model”

By default the summary is written by the conversation’s own model. To use a cheaper one, go to Settings → Subagents and, under Session, pick a model for Compact the context. It only writes summaries; answers still come from your chosen model.