OPEN_SOURCE ↗
REDDIT · REDDIT// 17d agoINFRASTRUCTURE
Stateless Memory hooks SentenceTransformer into SillyTavern
A Reddit project post outlines a local SillyTavern memory hook that uses SentenceTransformer cosine similarity to surface relevant lore without a separate vector database or Docker stack. It reads more like a proof of concept than a launch, but the design is a clean fit for fully local memory inside ST's existing Lorebook system.
// ANALYSIS
This is the right way to do ST memory: use embeddings for fuzzy recall, but keep the actual memory primitive in Lorebooks where users can inspect and prune it.
- –ST docs already describe World Info evaluation as stateless, so the project is aligning with the platform’s own memory model instead of fighting it. https://docs.sillytavern.app/usage/core-concepts/worldinfo/
- –SillyTavern Extras flags ChromaDB as part of the old Smart Context path, which is deprecated, so skipping a separate vector DB feels timely. https://github.com/SillyTavern/SillyTavern-Extras
- –SentenceTransformer similarity should beat keyword-only recall for fuzzy callbacks and character memory, but the quality bar will hinge on threshold tuning and how messy the Lorebook gets as chats grow.
- –The tradeoff is simplicity over scale: great for a self-hosted RP workflow, less compelling if you need a shared memory service across many chats or users.
// TAGS
stateless-memorysillytavernllmembeddingragself-hostedvector-db
DISCOVERED
17d ago
2026-03-25
PUBLISHED
17d ago
2026-03-25
RELEVANCE
7/ 10
AUTHOR
WillingnessSweaty282