SillyTavern guide
If your SillyTavern character has stopped remembering earlier scenes, started contradicting established events, or your responses have slowed to a crawl in a long roleplay, you've hit the context window. Every long-running SillyTavern chat gets there eventually. This page explains how context actually works in SillyTavern, what the built-in tools do (and don't do), and how to compress a long chat into a summary that keeps your story going.
SillyTavern doesn't send your whole chat to the model. On every message, it assembles a prompt from a fixed budget of tokens, your context size setting, and fills it in priority order: system prompt and character card (description, personality, scenario), World Info / lorebook entries triggered by keywords, Author's Note and summaries if you use them, and as much recent chat history as fits in whatever's left.
That last part is the problem. The character card, lorebook, and prompts are permanent residents of the context; chat history is what gets evicted. In a long roleplay, everything before the cutoff simply isn't sent, the model has never seen it. Your character isn't 'forgetting', it was never shown those messages in the first place.
SillyTavern ships with a Summarize extension that periodically asks a model to condense your chat history and injects the result into the prompt. Best for hands-off, incremental memory in an ongoing chat.
Limitation: quality is inconsistent, it summarizes in small windows as it goes, so it compounds its own errors, loses emotional beats and minor-but-important details, and with local models the summary passes eat generation time. Most long-RP users end up editing its output by hand or turning it off.
The Vector Storage extension embeds old messages and retrieves 'relevant' ones by similarity when they seem related to the current message. Best for resurfacing specific facts ("what was the name of that tavern").
Limitation: retrieval is keyword/similarity-driven and hit-or-miss for narrative continuity. It can pull back a relevant fragment; it can't maintain a coherent picture of how the story developed.
Manually recording key facts in World Info entries or a running Author's Note is the most reliable native method, and the most labor-intensive. Best for permanent facts: character traits, world rules, established relationships.
Limitation: you're doing the summarizing yourself, after every significant scene, forever.
If you have a long chat, or a whole context window's worth of story, the practical move is to compress it once, well, and inject the result back in.
Export or copy your chat. SillyTavern stores chats as files you can export, or select and copy the conversation from the chat window. Open thredly.io and paste it in. thredly compresses the whole thread into a structured summary: what happened, key events in order, character and relationship developments, established facts, and where the story currently stands. Paste the summary into your Author's Note (or a high-priority lorebook entry), then trim your context back to recent messages.
Because thredly summarizes the entire thread in one pass rather than incrementally, it keeps the through-line of the story, the arcs and turning points, not just scattered facts. Long inputs are handled properly: higher tiers process very large pastes (up to 750,000 characters on Ultimate), so even a marathon RP fits in one thread.
One more thing worth knowing: thredly deletes your source text after summarizing. The raw chat never persists on our servers, only the summary you keep. For roleplay chats, which tend to be personal, that matters.
For long-running roleplays, the setup that works is layered. This is the same structure SillyTavern's prompt assembly already wants: stable facts, condensed history, fresh dialogue. You're just making the middle layer actually good.
Match it to your model's real limit (and your hardware, for local models), but don't treat a bigger window as a memory strategy. 8kâ16k with a solid summary in the Author's Note beats 32k of raw history that still eventually overflows.
Prompt processing scales with context. If you're running a local model with a large context size and a long chat, every message reprocesses a huge prompt. Summarizing the history and trimming context is usually a bigger speedup than any sampler tweak.
It works, but incrementally, each pass only sees a window of the chat, so errors and omissions accumulate. For a chat you care about, a single full-thread summary (manual or via thredly) is consistently more faithful.
Open the chat management menu for the current chat and use the export option, or copy the messages directly from the chat window. Either output can be pasted into thredly.
Yes, that's one of the best uses of a compressed summary. A thredly summary of your RP is portable text: paste it into a new SillyTavern chat, a different character card, or an entirely different platform, and the story continues with full context.
No. thredly processes your pasted text, produces the summary, and deletes the source. Your raw chat is never stored.