Long roleplays fall apart because context windows fill up. Characters forget plot points. Manual lorebooks become a chore. And vector databases make characters weirdly omniscient-they somehow know secrets from conversations they were never part of
OpenVault fixes this. It runs in the background while you chat, tracking what actually happened, who was there, and how relationships evolve. No external databases. No Docker. No Python servers. Everything lives inside SillyTaverns native chat storage. Can be used with free models
Heres what breaks with most memory systems: you have a secret conversation with Character A. Later you're alone with Character B. You ask "what should we do about that secret?" and Character B responds like they were there
Thats not how memory works
OpenVault tracks witnesses. Every extracted event records who was present. Characters only recall what they actually experienced, were told directly, or overheard. No more accidental omniscience
Event Extraction. As you chat, OpenVault identifies what happened-actions, emotional shifts, revelations. Each event gets an importance rating (1-5 stars) and a witness list
Knowledge Graph. People, places, factions, objects, and concepts get tracked as nodes. Relationships between them are edges that update over time. When two characters go from enemies to allies, the graph knows
Reflections. After enough significant events pile up for a character, OpenVault pauses to reflect. It synthesizes raw memories into psychological insights-shifting motivations, subconscious drives, evolving relationship dynamics. These are internal truths, not things the character says out loud
GraphRAG Communities. Every 50 messages, it analyzes the relationship web to detect social circles and factions. This produces a running "world state" summary so macro-level plots don't get lost
Smart Retrieval. Before the AI generates a response, OpenVault scores all candidate memories using a blend of:
- Exponential forgetfulness (old trivial stuff fades, critical memories stick)
- BM25 keyword matching
- Vector similarity against recent context
Memories get injected into the prompt in chronological buckets: The Story So Far, Leading Up To This Moment, Current Scene, and hidden Subconscious Drives that influence behavior without being spoken
Requirements:
- SillyTavern 1.13.0+
- An extraction-capable LLM (mid-tier models work fine)
Embeddings: By default, OpenVault downloads a lightweight multilingual model (multilingual-e5-small) that runs in your browser via WebGPU/WASM. Or point it at your local Ollama instance
Thats it. No ChromaDB. No Docker. No cloud vector service
OpenVault adds a panel to SillyTaverns Extensions menu. The layout is intentionally progressive-common stuff up front, tuning buried in collapsible sections
Dashboard. Status, health, and a live Payload Calculator showing exactly how many tokens your extraction model needs
Memories. Searchable memory bank. Filter events vs reflections. Edit importance or summaries manually. View character emotional states
World. Read-only viewer for the knowledge graph. Browse communities, factions, and tracked entities
Advanced. Tuning knobs: alpha-blend weights, decay rates, similarity thresholds. Most users should leave these alone
Perf. Real-time metrics so you know if background processing is bottlenecking your browser
You control where memories appear in the prompt:
| Position | Where it goes |
|---|---|
| ↑Char | Before character card |
| ↓Char | After character card (default, recommended) |
| ↑AN | Top of authors note |
| ↓AN | Bottom of authors note |
| In-chat | At specific message depth |
| Custom | Manual placement via macros |
Manual macros (when set to Custom):
{{openvault_memory}}- Memory context{{openvault_world}}- World/faction context
The main panel shows current positions as badges like [↓Char | ↑AN]. Click a macro badge to copy it
OpenVault handles non-English roleplay without mangling JSON. It detects script (Latin, Cyrillic, etc.), uses appropriate stemming, and deduplicates characters across scripts
Everything stays on your machine. The in-browser embedding model means no text goes to third-party services. The only network calls are to your configured LLM endpoint (local or API-your choice)
-
GraphRAG - Community detection and hierarchical summarization. Park et al., "From Local to Global: A Graph RAG Approach to Query-Focused Summarization" (arXiv:2404.16130)
-
Generative Agents - Importance-weighted memory, reflection triggers, observation loops. Park et al., "Generative Agents: Interactive Simulacra of Human Behavior" (arXiv:2304.03442)
- 16.00 - Improved LLM answer parsing
- 15.50 - Added data schema versioning for easier migrations
- 15.00 - Major refactoring, bug fixes
- 14.00 - Emergency Cut button (prune chat history while keeping key events)
- 13.50 - Compatibility with other chat-altering extensions
- 13.00 - Support for all ST vectorization sources
- 12.50 - Customizable injection positions
- 12.00 - UI revamp (protecting users from over-tuning)
- 11.50 - Fixed memory balance distribution
- 11.00 - Test refactoring, prep for additional languages
- 10.50 - Fixed hairball cluster bug
- 10.00 - Parallel request support
- 9.50 - Group play PoV fixes
- 9.00 - Stable single-character RP release
Check my other project https://github.com/vadash/litellm_loader. Its a bit Do It Yourself, but should provide nice start. You can setup complex fallbacks. For example, try (Kimi K2 and Kimi K2 0905 at Nvidia NIM), if both down then use long cat chat
GNU AGPL v3.0