Compact Settings
Configure auto-compact behavior for long conversations.
Compact Settings
Configure how BraceKit checkpoints long conversations: older turns become a structured summary, recent messages stay verbatim.
Enable Auto-Compact
| Setting | Behavior |
|---|---|
| On (default) | Compact when usage exceeds the Compact-at share of the model window |
| Off | Manual compact only via /compact |
Compact at
Percentage of the selected model’s context window. Auto-compact runs when estimated usage goes past this share (the remainder is reserved for the next reply).
| Value | Behavior |
|---|---|
| 50–70% | Compact earlier, more often |
| 87% (default) | Matches a ~16k token reserve on a 128k window |
| 95% | Compact only when nearly full |
The settings page shows the equivalent reserved token count for the current model.
Keep recent
Share of the model window left as verbatim recent messages (not summarized). Default 16% (~20k tokens on a 128k window).
BraceKit only cuts at safe points (never a tool result without its assistant call). If one turn is larger than this budget, the start of that turn is summarized separately.
Manual Compact
- Type
/compactin the chat - Optionally add extra instructions:
/compact preserve API error strings - Press Enter
/compact extra text is one-shot guidance for that summarization only.
Related
- AI Provider Settings (for Context Window configuration)
- Chat Settings
- Troubleshooting