"Everything your preset should have been: persistent memory, chain-of-thought reasoning, automated NPC tracking, and ComfyUI image generation in a single install."
Megumin Suite is a full-stack overhaul to how SillyTavern presets work. It replaces your preset, your memory system, your NPC management, and your image Generation — all in one extension. V8 introduces the new Obsidian, Spark, and Fusion engines with unprecedented dialogue realism, a completely redesigned NPC Dossier template, Inline Image Generation, and fully editable prompts across every subsystem.
Features • Installation • The V8 Engines • Memory Core • Image Gen
V8 marks a huge step forward in terms of story quality, NPC depth, and user controls.
- V8 Obsidian Engine: A complete rewrite that's all about obsession with psychology, dialogue, and story arcs. Has a strict hierarchy of priorities, story and sub-story arcs, and some of the most stringent dialogue rules against AI slop in the preset family.
- V8 Spark Engine: A much lighter version of the Obsidian engine that retains all the same rules about psychology, dialogue, and story momentum without the high token costs.
- V8 Fusion Engine: The best of both worlds – a mix of the deep psychology rules in V8 Obsidian and the V6 Dream Team framework, which involves having six specialized writer roles (NORA, ANVIL, OPUS, JULIA, MIKI).
- V7.5 Kismet Engine (Bonus): Initially planned for its own V7.5 release, the Kismet engine was instead integrated into V8. This engine is all about the story.
- Redesigned NPC Dossier Template: This is a completely revamped NPC Dossier that contains many new categories, including Role, Where to find them, Voice (the way they speak), Image Tags (for ComfyUI's Booru tags), Read on the PC, Tiered Secrets (semi-public → private → hidden), Canon Lock (immutably true facts about the character), and Orientation. The template also comes with trigger criteria to ensure that the AI will not create a dossier unless absolutely necessary.
- Inline Image Generation: Images are generated directly inline with the AI's response instead of posting them separately as part of a gallery. Includes individual retry buttons for each image.
- Fully Editable Prompts: All subsystems (Story Planner, Ban List, Image Gen, Memory Core, NPC Bank) now have an "Advanced: Edit Prompts" panel in which users can edit all the templates for prompts sent to the AI. The edits are enabled/disabled via a toggle and are saved individually per profile.
- Image Gen Overhaul: Instead of using style and perspective dropdowns, there's now a Prompt Template system (six preset options: Illustrious/Z Image × POV/Cinematic/Portrait). There are additional toggles for Include Examples, Better Booru Tags (direct NSFW language), Inject NPC Tags (inject the saved NPC image tags into the prompts), and multiple images per response (one to four).
- Auto-CoT Matching: Using a CoT framework other than V9 and selecting V7, V7.5, V8, or V8Fusion will automatically update your chosen framework accordingly.
- NPC/OOC Keyword Trigger: Avoid using unnecessary tokens by only injecting a new NPC dossier template if your input contains either of the key phrases "NPC" or "dossier". (Example: "Create an NPC dossier for the bartender")
- "Scan Story" Button for NPC Bank: You can manually scan your story and extract all your significant NPCs in one go — there's no need for the AI to do so in real time while chatting.
- Customizable Chunk Size for Memory Core: Customize how many messages are processed per chunk (between 10 and 40).
- Draggable Button: The wand button in the Megumin suite can be easily dragged and positioned to any location on the page, with persistent positions after session restarts.
- Live Token Counter Accuracy: The Token Counter now calculates tokens at a
4.8chars/token ratio (matching modern efficient tokenizers like Claude/GPT-4). It also now intelligently ignores highly variable dynamic blocks (like Memory Vaults and NPC lists) to give you a stable, accurate "Base Payload" estimation. - New Writing Style Interface Design: Updated the old stacked menu design to a more intuitive sidebar-based navigation panel, containing Precooked, My Library, and AI generators. Also removed the need for separate pages for DN Ratio & POV selectors.
- Min/Max Words Toggle: Switch between specifying either minimum or maximum target word count for writing prompts.
- Optimizations: Memory Core uses cached Set lookups to determine if a message is archived in O(1), paginates accordion/vaults (only loading 20 items at a time with a "load more" option), debounce visual updates, and caches token counts.
The V8 series constitutes the pinnacle of narrative engines in the Suite's history.
- V8 Obsidian: The king. No other narrative engine excels at portraying complex human psychology, realistic flawed dialogues, and multi-layered autonomous story plot generation. Equipped with an official 7-rule priority hierarchy, structured storyline (Main Arc + 3 Subplots + Micro-Tensions), foreshadowing methodology, causality and consequences, agenda-as-plot-motor of NPCs, and by far the most aggressive anti-slop dialogue.
- V8 Spark: All of the above rules condensed into a compact set of instructions taking up less tokens than the Obsidian version. For use on light models that find the Obsidian engine too cumbersome.
- V8 Fusion: Combines the best from V8 Obsidian (psychology, dialogue engine) with the V6 Dream Team approach of a 6-specialist writer room (NORA, ANVIL, OPUS, JULIA, MIKI). Each of the 6 specialists is responsible for a particular type of writing: NORA ensures continuity, ANVIL handles psychology, OPUS constructs the plot, JULIA writes the narration, and MIKI creates dialogues.
📝 Note: V8 engines lock the Persona & Toggles tab automatically. This is because V8 engines utilize their native persona and narrative settings system, which conflicts with any external injection.
A persistent character database that tracks every NPC accurately across sessions.
- Auto-Extraction: When a significant NPC is introduced, the AI writes a detailed dossier and saves it to the bank.
- Redesigned Dossier Template: The V8 dossier is dramatically more detailed than V7. NPCs now include Role, Where to Find Them, Voice (how they speak), Image Tags (Booru tags for ComfyUI), Read on the PC (what the NPC currently thinks of the player), Tiered Secrets (semi-public → private → buried), and Canon Lock (immutable facts that can never change). The template also includes strict trigger conditions — dossiers are only generated for characters who are Named, Voiced (more than a transactional line), and Staked (have a want or role that affects the story). No more dossiers for cashiers and bartenders.
- Dynamic Injection: Scans your last 4 messages and injects relevant NPC dossiers into the prompt so the AI remembers them accurately.
- Image Tags Only Mode: Per-NPC toggle to hide the text dossier from the AI (saving tokens) while still making their Booru image tags available to ComfyUI.
- OOC Trigger: Save tokens by only injecting the blank dossier template when you mention "NPC" or "dossier" in your message.
- Scan Story: Manually scan your entire chat history and extract all significant NPCs at once.
- AI Portrait Studio: Click a button to have ComfyUI automatically generate a character portrait based purely on the AI's physical description of them.
Manually control the AI’s internal thought process prior to text generation.
- Master Toggle: Enable/Disable CoT.
- Auto-matching: The selection of any V7/V7.5/V8 will automatically set your CoT to the correct version.
- The 5-Phase Audit (V7): Ground Truth -> Plot Engine -> Scene Design -> Active Draft -> Correction Loop.
- V8 CoT: A streamlined 7-step thinking process (INPUT -> STORY -> NPCs -> DRAFT DIALOGUE -> DIALOGUE KILL CHAIN -> NARRATION -> FINAL).
- V8 Fusion CoT: Brief thinking prompts specific to the writer room framework.
- V7.5 Kismet CoT: Story Engine-based thoughts that include vocabulary gate checks & arc phase.
Hook up your own personal ComfyUI instance to your bot to create images on the fly during roleplay.
- Inline Mode: Images are displayed directly in text with individual image-specific retries in the AI text reply.
- Gallery Mode: Images appear as individual galleries (the default way).
- Prompt Templates: Includes 6 templates (Illustrious/Z Image × POV/Cinematic/Portrait), complete with full rules and example images.
- Multi-Image Creation: Create from 1-4 images in one AI reply.
- Inject NPC Tags: Save and automatically insert NPC Booru image tags into prompts when relevant NPCs are involved.
- Improved Booru Tags: Ensures that the AI uses specific and unembellished Danbooru tags with all applicable NSFW tags.
- LoRA Lab & Parameters: Fine-tune Steps, CFG, Denoise, and 4 LoRA slots in SillyTavern.
Fed up with the AI repeating phrases like "a shiver ran down your spine" and "testament to..."?
- Select Analyze Chat and let the AI analyze your last 50 messages, finding the 5 most common crutch phrases that you use.
- Turns them into hard bans automatically, preventing the AI from ever using them again in the future.
- Import/Export ban list as JSON to transfer it between profiles.
- Story Planner: Generates ideas for future plot twists, following at least 10 upcoming milestones.
- World State Tracker: Inserts a dynamic collapsible panel displaying the world state including the date, weather, PC's current condition, NPC goals, planted seeds, consequences countdowns, and story progression phases.
- NPC Inner Chatter: Makes the AI generate an invisible block of conversation revealing the thoughts of your NPCs.
Keep track of the story and stop burning tokens on bloated context windows.
- Working Memory: The most recent conversation logs.
- Short-Term Memory: Automatically generated summaries by the background service of previous chunks.
- Long-Term Vault (Vector DB): Applies either TF-IDF Keyword Matching or SillyTavern's Semantic Embeddings to retrieve relevant archived memories and feed them back into the prompt without you knowing.
- Prompt Interceptor: Automatically strips out any archived messages from the prompt and saves thousands of tokens.
- Configurable Chunk Size: Control how many messages to sumarize at once ranging from 10 to 40 messages.
- "Every Reply" Auto-Trigger: You can enable memory retrieval to happen each time the AI replies automatically.
📝 Note: Memory Core is not for advanced users. This is an easy way for users with up to 1000 message conversations to save their context space without managing any manual summaries.
Every subsystem now includes an "Advanced: Edit Prompts" collapsible panel:
- Customize the system prompt, user task prompt, thinking instructions, and injection templates for Story Planner, Ban List, Image Generation, Memory Core, and NPC Bank.
- Each editor has an enable/disable toggle — disabled means the default prompts are used.
- Custom prompts are saved per-character/per-group profile.
- Open SillyTavern.
- Go to the Extensions menu (the block icon).
- Click Install Extension.
- Paste the repository URL:
https://github.com/Arif-salah/Megumin-Suite - Refresh SillyTavern.
- Download the JSONs files from this repo: https://github.com/Arif-salah/Megumin-Suite/tree/main/Presets
⚠️ Note: If you download these on your phone and your browser renames them to.json.txt, you must use a file manager to rename them and delete the.txtpart. Furthermore, make sure the Engine file is named EXACTLYMegumin Engine.jsonbefore you import it. The Suite file's name doesn't matter, but the Engine must be exact.
- Open SillyTavern, go to the Ai Response configuration tab.
- Click the Import Preset button (the little folder with an arrow) and upload the json files.
- Once imported, open your preset dropdown and make sure "Megumin Suite" is the active preset. The extension handles the Engine silently in the background.
or just watch the Install video: youtube Video
💡 Pro Tip: - Megumin Suite V8 DS4 is for Deepseek. - Megumin Suite V8 Gemini is for gemini models. - Megumin Suite V8 Claude+GLM you know for what. if you have model not here just try.
⚠️ Important: Megumin Suite ships with several Regex scripts that clean and format messages before they're sent to the AI. After installing, go to the Extensions → Regex panel and make sure all Megumin-related regex entries are enabled.
- Select an Engine: Open the Megumin Suite menu (wand icon) and pick a Core Engine. V8 Obsidian or V8 Fusion are recommended for the best experience.
- Set your Style: Go to the Writing Style tab. The V8 Default style (witty, opinionated observer) is applied automatically. Customize it or choose from precooked styles.
- Enable CoT: Go to the Chain of Thought tab — it's enabled by default and auto-matches your engine version.
- Chat! The extension will handle all prompt injection, formatting, and memory management silently in the background.
💡 Pro Tip: If you want to see exactly what Megumin Suite is sending to the AI under the hood, enable Prompt Payload Preview in the Global Settings tab.
- LLMs: Designed for highly capable instruction-following models (Claude 4.6 Sonnet/Opus, DeepSeek v4, Gemini 3.1 pro/flash, GLM 5.1). Smaller local models may struggle with the strict V8 CoT instructions. use v8 spark for small LLMs.
- Does this extension mess with my other presets? No — your other presets will work just fine. Megumin Suite only injects its rules into its own designated preset (Megumin Suite). Your existing presets remain completely untouched.
- Vector Storage (Optional): if you using Semantic Embeddings in the Memory Core, you can change the model its Cohee/jina-embeddings-v2-base-en by default if it heavy for your pc use Xenova/all-MiniLM-L6-v2 you can change it inside "sillytavern\config.yaml"
- Old Versions: Legacy docs are here: Megumin Suite v4 Legacy Readme Megumin Suite v5 Legacy Readme Megumin Suite v6 Legacy Readme Megumin Suite v7 Legacy Readme
- Built natively for the SillyTavern.
- MVU Compatibility integration inspired by KritBlade's MVU Game Maker.
Megumin Suite is free and always will be. If it saved you hours of prompt engineering or made your sessions better, consider tossing a few bucks it keeps development alive and the updates coming.
🪙 Crypto (LTC): LSjf1DczHxs3GEbkoMmi1UWH2GikmXDtis
⭐ Not in a position to donate? Starring the repo and sharing it helps just as much.