SillyTavern-CheckpointSummarize

Manual checkpoint memory for long SillyTavern chats.

SillyTavern-CheckpointSummarize helps keep long roleplay and chat sessions usable by turning old chat ranges into locked checkpoint summaries. Each checkpoint covers a specific message range, can be reviewed or edited, and can be injected back into the prompt as long-term memory.

Why This Exists

Long chats eventually become too large for a model context window. A single global summary can drift, overwrite details, or blur the timeline.

CheckpointSummarize uses a more controlled approach:

• summarize one raw message range at a time;
• keep earlier checkpoint summaries immutable by default;
• use previous checkpoints only as read-only context for the next draft;
• inject locked checkpoints as stable memory;
• reopen deleted checkpoint ranges as gaps that can be rebuilt.

The result is a modular memory timeline instead of one fragile rolling summary.

Highlights

• Range-based checkpoint summaries.
• Manual review and editing before locking.
• Auto mode for unattended checkpoint generation.
• Easy mode for "turn it on and let it maintain itself" workflows.
• Prompt injection with an After World info option.
• Per-checkpoint injection toggle.
• Import and export for backups or migration.
• Range hash warnings when the underlying chat messages changed.
• Connection Profile selection for dedicated summary backends.

Quick Start

1. Open SillyTavern.

2. Go to Extensions.

3. Install from:

   https://github.com/NovNovikov/SillyTavern-CheckpointSummarize
   

4. Enable Enable extension.

5. Enable Easy mode (recommended).

6. Select a Connection Profile suitable for summarization.

7. Watch Current Summary for generated checkpoints.

Easy mode applies the recommended automation preset, including prompt injection.

How It Works

1. The extension finds the next unsummarized message gap.
2. It builds a draft prompt for that raw range only.
3. The model generates a checkpoint summary.
4. The summary is reviewed or auto-approved.
5. Locked checkpoints become prompt memory.

Checkpoints are stored in the current chat metadata under:

chatMetadata.checkpoint_summarize

The extension does not mutate or delete raw chat messages.

Modes

Manual Mode

Use this when you want full control.

• Choose a range.
• Generate a draft.
• Edit the text.
• Lock it as a checkpoint.

Auto Mode

Use this when you want automation with your own settings.

• Auto mode waits until an unsummarized tail reaches the configured raw block target.
• Middle gaps, such as deleted checkpoint ranges, are rebuilt automatically.
• Auto approve can lock generated drafts without manual review.

Easy Mode

Use this when you want the extension to manage the workflow.

• Applies recommended automation settings.
• Recalculates context limits from current prompt conditions.
• Keeps the checkpoint timeline maintained.
• Hides or locks controls that could break the automated flow.

Recommended Workflow

For long roleplay chats:

1. Use a non-reasoning summary profile if possible.
2. Keep prompt injection enabled.
3. Use Easy mode.
4. Periodically review Current Summary.
5. Edit only checkpoints that need human correction.
6. Export checkpoints as backups.

Import And Export

The extension supports:

• Export checkpoints
• Export current summary
• Import checkpoints
• Import current summary

When imported checkpoints do not match the current chat message hashes, they are kept as memory-only records. They can still be injected, but they do not count as range coverage.

Documentation

• English manual
• Russian manual
• Architecture notes
• Changelog

Status

This is a third-party SillyTavern extension focused on practical long-chat memory management. It is designed to work without patching SillyTavern core.

License

MIT. See LICENSE.