SillyTavern vector memory setup with Ollama

This tutorial explains how to configure vector storage in SillyTavern for lorebook retrieval and chat memory. The speaker starts with Local (Transformers), which they say requires no extra installations, and enables world info scanning for an active lorebook whose entries have no keywords.

The settings walkthrough covers how many recent messages form the search query, the similarity threshold, and the maximum number of entries to include. It also explains the choice between enabling all eligible entries and selecting entries individually. In the terminal demonstration, reducing the maximum from five to two changes how many lorebook entries enter the prompt. The speaker enables chat vectorization too, but the conversation is too short to demonstrate retrieval of older messages.

Ollama provides another local AI embedding source with a selectable model. The tutorial shows downloading an embedding model, listing installed models, and removing one. After removal, vectorization fails; response generation continues, but the retrieved lorebook entries are absent. The demonstration uses OpenRouter for responses, so local embeddings do not make the whole chat local.

For Spanish chats, the speaker downloads a Spanish embeddings v2 model through Ollama and selects it in SillyTavern. They recommend language-specific embeddings over the default English model. File vectorization appears briefly at the end, but its setup is outside this tutorial.