Context Control
DesireCore keeps the visible conversation history, but each model call has a finite context window. Context control lets long conversations continue without deleting the visible history.
/new
/new creates a new context boundary in the same conversation.
- Visible history remains available
- Future model calls stop automatically reading the full content before the boundary
- Useful when switching topics or starting a fresh task
/compact
/compact summarizes older messages and uses that summary as part of future context.
- Original history remains visible
- The model mainly reads the summary plus recent turns
- Useful for long tasks, writing sessions, and multi-step reviews
Automatic Compaction
When a conversation approaches the model's context limit, DesireCore can compact old context automatically. This prevents sudden failures while retaining goals, decisions, completed steps, and open tasks.
CompactSession
Some agents can call CompactSession themselves when they detect a long-running task. It keeps the UI history intact and only changes what future model calls read.
Practical Recommendations
| Scenario | Recommended Action | Reason |
|---|---|---|
| Switching topics, such as from coding to the weather | /new | Prevent old task context from interfering with the new topic |
| Working on a long task, such as writing a 20-page document | /compact | Free up context space while retaining a task summary |
| The conversation is long but still about the same task | Wait for automatic compaction | The system compacts context at an appropriate time |
| The agent went in the wrong direction and you want to try again | Rewind | Return to a checkpoint before the mistake |
| You no longer need the conversation at all | Clear chat history | Irreversible; use with care |
/new and /compact/new: Breaks the context connection with earlier content. Use it when the next topic is unrelated to the previous discussion./compact: Keeps a summary of earlier content. Use it when you are continuing the same task but the conversation has become too long.
If you are unsure, prefer /compact: it preserves the task context in a summary.
A Companion's long-term memory (knowledge acquired through teaching or automatic learning) is stored in AgentFS and is unaffected by context compaction. Learned rules and preferences remain in effect after /new or /compact.