SillyTavern Summarize and Vector Storage for long memory

Do you need them yet?

Both extensions solve one problem: the oldest messages falling out of the context window. If your story still fits, they add cost without adding memory. At a context setting of 56,000 tokens, which is what context size for roleplay recommends for UnsavedAI's Gemma 4 models, something like 40,000 words of chat stays in view at once. Many stories never get there; long campaigns do.

For the general picture of memory across apps, see making AI remember your story. This page is the SillyTavern machinery.

Setting up Summarize

Open the Extensions panel and expand Summarize. Choose the summary source; the main API option uses whatever model you are connected to. Then set:

• Update interval: how many messages (or words) between automatic summaries. With a long context you can go every 30 to 50 messages instead of every 10.
• Injection position: before or after the main prompt, or in the chat at a depth. Above the history suits a synopsis; a moderate depth makes it stronger.
• The summary prompt itself, which you can rewrite.
• Summarize now, for a manual run after a big scene.

A summary prompt that keeps the useful parts:

[Pause the roleplay. Update the story summary in no more than ten bullets: where we are, who is present, how {{char}} feels about {{user}} now, promises made, secrets known, unresolved threads. Past tense, plain facts, no prose.]

The current summary sits in an editable box. Read it after each update and fix anything wrong, because a mistaken summary is repeated on every later reply.

Setting up Vector Storage

Vector Storage turns each message into an embedding, a list of numbers describing its meaning, then retrieves the closest old messages when you write. In its panel, pick a vectorization source. The local option runs on the machine running SillyTavern and needs no extra key; remote embedding sources need their own provider's key. Your UnsavedAI key powers the chat model, so the local source is the simplest pairing.

Turn on vectors for chat messages, then set how many recent messages form the search query, how many old messages to retrieve, and where to insert them. Start with a query of 2, three retrieved messages, inserted a few messages deep. For an existing chat, vectorize everything once so old messages are searchable.

Vector Storage shines at "the name of the ferryman from week one" and stumbles on "how has our relationship changed", which is the summary's job.

Running both, and what it costs

They combine well: the summary carries the arc, vectors recall a specific line when the story circles back to it. Watch the cost of summaries made through the main API. Each update is an extra generation, and on UnsavedAI Free writes one reply at a time, so a summary running right after a reply makes your next message wait for it. Longer intervals keep that rare; Priority writes two at a time, which hides most of the wait.

Long sagas are where this pays off. A caravan road like Rokhshana's journey to Chang'an or a salvage mystery with Captain Amos Tregarth builds up names, debts and clues over hundreds of messages, exactly what a tidy summary preserves. When the summary is getting crowded too, continuing a long roleplay covers starting a fresh chapter.

Questions people ask