SillyTavern guide

SillyTavern Context Size: Why Long Chats Break and How to Summarize Them

Start with SillyTavern’s Context and Response settings, then inspect what was actually sent to your model. Older messages above the dotted context line are left out of the prompt. The built-in Summarize extension can help carry earlier events into an ongoing chat.

Last reviewed: 5 September 2026

Get 1 free summary

Check the native context settings first

SillyTavern builds a prompt from character information, instructions and available chat history. Response tokens take space from the overall context budget. Use your connected model and provider’s supported limit when choosing a context size. The official common settings guide explains the context boundary and Prompt Itemization.

  1. Check Context and Response in the generation settings. A larger reply allowance leaves less room for the prompt.
  2. After a reply, open its three-dot menu and choose Prompt Itemization, shown by the lined-square icon. Inspect which parts of the conversation reached the model.
  3. Look for the dotted context boundary in the chat. If an important event is above it, bring the relevant detail into a reviewed summary or concise note.

Use SillyTavern’s built-in Summarize

Open Extensions, the stacked-cubes panel, and find Summarize. It is included by default. Main API uses your connected model and needs a working API connection. Review the Current summary and correct omissions before relying on it.

Use Summarize now for a manual update, or choose an update interval. An interval of zero disables that automatic trigger. Revisit the intervals when changing models. Check the injection position so the summary appears where intended.

If concurrent requests fail, use a blocking mode. For llama.cpp-style backends, the documentation recommends Classic, blocking to support prompt reuse. All generated summaries can miss details or invent them. See the official Summarize settings.

Troubleshoot truncation before adding more tools

  • Older scenes disappear: use Prompt Itemization to check the input, then review the summary. A saved transcript and the text sent in the current prompt are different things.
  • A lorebook takes too much room: check World Info’s Context % or Budget and keep entries concise. Its Alert on overflow can flag when activated entries exceed the allowance. See World Info budgeting.
  • The backend trims a prompt near its limit: check token counting. For Text Completion APIs, the native tokenizer and token-padding controls can help account for estimates. Those padding controls do not apply to Chat Completion APIs. See SillyTavern’s tokenizer guidance.
  • The summary stops updating: check Pause, update intervals and the API connection. Restore Previous can recover an earlier summary if the latest output is unhelpful.

Keep a chat backup before starting over

In the bottom-left options menu, open Manage chat files. JSONL exports retain chat metadata for re-import, but exclude images and file attachments. Keep those separately. The plain-text download is useful for reading or supplying text to another tool; it cannot be re-imported as a complete SillyTavern chat. See Chat File Management.

Keep a private copy of your original chat before making changes. Check exported files for personal details before sharing them.

When a structured handover may help

If you deliberately want a new conversation or a different model, a short, reviewed account of the current scene can be useful. It can record established events, character relationships, unresolved threads and where to continue. You can write it yourself. Native Summarize may already meet your needs.

Before starting a fresh scene, check which branch and scene boundary your handover describes. SillyTavern’s Raw Summarize modes omit World Info and Author’s Note; Classic can also exclude them with No WI/AN. Check important lore separately rather than assuming it was included. For a manual source check, use the free fiction-canon handover template. Its original worked example separates approved events, discarded branches, proposed future scenes and what each character knows. You can copy it without an account.

thredly’s Chrome extension does not directly support SillyTavern. The supported workflow is manual: SillyTavern conversation text → thredly web app → structured handover → new conversation.

Copy text from your chat or open the plain-text export, supply the relevant text to the web app, and check the result against the original. Paste the handover into a new conversation with your next instruction. This does not synchronize chats, change SillyTavern’s context settings or guarantee that every detail will be retained.

Related guides

Frequently asked questions

What context size should I use in SillyTavern?

Use a size supported by your connected model and backend, with room for the reply and other prompt content. There is no single best size for every model or setup.

Does leaving the context window delete old messages?

The dotted line marks which chat messages are included in the prompt. Messages above it are not sent to the model; that is different from deleting them from your saved chat.

Do I need thredly to manage SillyTavern context?

No. Start with the native context settings and Summarize. If you choose a new conversation, you can write a handover yourself or review a summary created from supplied text.

Does thredly have a SillyTavern browser integration?

No. Its Chrome extension does not directly support SillyTavern. Supply conversation text to the thredly web app and copy the reviewed handover into a new conversation yourself.