Checkpoints
Checkpoints keep long chats coherent by summarizing the story as chapter-like memory. Instead of relying on the model to reread hundreds of messages forever, Pyre injects a compact recap into the prompt.
How they work
As a chat grows, Pyre periodically summarizes only the span since the previous checkpoint. The chain reads like chapters: each checkpoint adds the next piece of the story instead of retelling everything from the beginning.
Automatic checkpoints trigger by character-reply count, not by every stored note. You can also create one manually.
Branch-aware memory
Checkpoints are tied to the branch they were made on. If you rewrite a past turn, the old branch keeps its checkpoints and the new branch builds its own. If a checkpoint belongs to another branch, Pyre surfaces that honestly instead of pretending it applies to the current path.
Controls
| Action | What it does |
|---|---|
| Summarize now | Create a checkpoint manually. |
| Retry | Regenerate a failed or weak checkpoint. |
| Edit | Adjust checkpoint text by hand. |
| Delete | Remove one checkpoint or clear all. |
| Per-chat toggle | Turn Checkpoints on or off for a chat. |
| New-chat default | Decide whether new chats start with Checkpoints enabled. |
Set the automatic frequency to 0 to disable automatic checkpoints. Manual checkpoints remain available.
Prompt placement
Use {{summary}} in a preset to place the recap exactly where you want it. If the token is absent, Pyre injects the recap at its default location.
Failure handling
Checkpoint generation is a model call. If it times out, fails, or returns empty text, Pyre reports the issue and can retry instead of silently letting memory stop.
Checkpoints vs Live Sheet vs Script
- Checkpoints: narrative memory of what happened.
- Live Sheet: structured current state.
- Script: future direction and planned beats.