Product: SillyTavern Memory Books (STMB)
Reference version: v8.5.0, August 1, 2026
Purpose: A single, dense source of truth for an AI assistant that teaches, explains, and troubleshoots Memory Books.
- 1. How an AI Assistant Should Use This Manual
- 2. Product Definition and Mental Model
- 3. Core Vocabulary and Feature Selection
- 4. Requirements, Installation, and Initial Verification
- 5. Opening Memory Books and Understanding the Main Panel
- 6. Memory Book Storage Modes
- 7. Profiles, Connections, and Generation Routing
- 8. Scenes, Manual Memories, Automatic Memories, and Catch-Up
- 9. Token Saving, Hidden Messages, and the Memory Boundary
- 10. Lorebook Activation and Retrieval
- 11. Real Group Chat Mode
- 12. Narrator Mode
- 13. Branching Chats
- 14. Clips
- 15. Topical Clips
- 16. Side Prompts
- 17. Consolidation
- 18. Compaction
- 19. Regeneration
- 20. Context for Generation
- 21. Prompt Architecture and Authoring Rules
- 22. Summary Prompt Manager and Consolidation Prompt Manager
- 23. Regex Integration
- 24. Lorebook Entry Titles and Character Policy
- 25. Job Queue and Retry Controls
- 26. Visual Feedback and Accessibility
- 27. Current Settings Reference
- 28. Slash Command Reference
- 29. Troubleshooting by Stage
- 30. FAQ
- 31. Compatibility, Migration, and Current Historical Notes
- 32. Developer and License Notes
- 33. Compact Diagnostic Decision Tree
- 34. Minimum Recommended Teaching Sequence
- 35. Final Concept Summary
Treat this document as the current operational reference for Memory Books. It replaces the need to load the separate Start Here guide, README, User Guide, Side Prompts guide, How STMB Works guide, and historical changelog as independent knowledge files.
Terms:
- STMB = SillyTavern=MemoryBooks (this extension)
- ST = SillyTavern (base code that STMB extends)
When answering users:
- Preserve Memory Books terminology exactly. A Memory Book is a SillyTavern lorebook used by STMB; it is not a separate database format.
- Distinguish current behavior from historical behavior. Do not teach a removed or superseded workflow merely because it appeared in an old changelog.
- Distinguish Group Chat Mode from Narrator Mode. They solve different problems.
- Distinguish memory generation, lorebook storage/configuration, and later retrieval by SillyTavern. Activation/retrieval is part of base ST code.
- Do not invent controls, menu labels, provider behavior, or settings not described here.
- When a screenshot is supplied, identify only visible controls. Give the next immediate action rather than assuming an off-screen control exists.
- When troubleshooting, identify the first failed stage and test it before recommending prompt rewrites.
- Prefer a simple working configuration before advanced routing, multiple books, custom prompts, Regex, or Side Prompt automation.
- Explain that character filters and separate Memory Books improve routing and relevance; they are not a security boundary.
- State uncertainty when the user’s installed version, SillyTavern version, provider, or custom prompt may differ.
Narrator Mode is implemented in v8.5.0.
Several beginner documents said a manual memory was technically required before automatic memories could begin. Current STMB can create the first automatic memory from message 0 when no processed-message baseline exists. A first manual memory is still recommended because it verifies the connection, Memory Book, output format, and desired starting boundary before automation is trusted.
Memory Books is a SillyTavern extension that converts selected or automatically chosen chat ranges into structured memory entries stored in SillyTavern lorebooks.
The basic process is:
Chat messages
↓
STMB selects or receives a message range
↓
STMB assembles an AI request
↓
The model returns a structured memory
↓
STMB saves a lorebook entry
↓
Old processed chat messages may be hidden from active context
↓
SillyTavern later activates relevant lorebook entries
↓
The chat model receives those entries as context
STMB does not give a model permanent internal memory. It maintains an external reference system (lorebook entries). The chat model “remembers” when SillyTavern includes the relevant lorebook entries in the prompt to the AI.
- Generation quality — Did the memory-generation model produce an accurate, useful result?
- Storage and configuration — Was the result saved to the intended Memory Book with appropriate activation settings?
- Retrieval and model use — Did SillyTavern activate and send the entry, and did the chat model use it correctly?
Troubleshoot these stages separately.
A lorebook, also called World Info in parts of SillyTavern, is a collection of entries that SillyTavern can conditionally add to a model request. A lorebook entry normally has:
- a title/comment;
- content;
- activation keywords or another activation mode;
- insertion position and order;
- recursion and budget controls;
- optional character filters and other metadata.
A Memory Book is an ordinary SillyTavern lorebook used by STMB. It can be opened, edited, reordered, exported, imported, or deleted with normal lorebook tools. Depending on the features used, it may contain:
- scene Memories;
- Arc, Chapter, Book, Legend, Series, or Epic summaries;
- Clip and Topical Clip entries;
- Side Prompt tracker entries;
- other STMB-managed entries.
A scene Memory is not the original transcript. It is a compressed representation intended to preserve continuity-relevant information such as:
- events and consequences;
- decisions and plans;
- discoveries and reveals;
- relationship or emotional changes;
- individual knowledge, beliefs, or misunderstandings;
- important objects, locations, identities, promises, and constraints.
Hiding processed messages does not delete them. It prevents those messages from being sent to the AI and therefore continuing to consume active chat-history context.
| Need | Feature | Meaning |
|---|---|---|
| Summarize one selected or automatic chat range | Memory | “Remember what happened in this scene.” |
| Save selected chat wording or one fact | Clip | “Save this note.” |
| Gather facts about one subject from saved Memories | Topical Clip | “Gather everything my Memories say about this.” |
| Maintain changing information over repeated runs | Side Prompt | “Keep this tracker updated.” |
| Combine several lower-tier Memories or summaries | Consolidation | “Roll these entries into a higher-level recap.” |
| Shorten one existing STMB-managed entry | Compaction | “Trim this entry without losing its facts.” |
| Replace an existing entry using its original sources | Regeneration | “Rebuild this entry and review a replacement.” |
- Clip vs Topical Clip: A Clip starts with text highlighted in the current chat. A Topical Clip starts with existing confirmed STMB Memories.
- Topical Clip vs Side Prompt: A Topical Clip is run manually to gather a topic. A Side Prompt can repeatedly maintain a changing tracker.
- Compaction vs Consolidation: Compaction rewrites one entry. Consolidation creates a new higher-tier summary from several entries.
- Memory vs Side Prompt: Memories are normally sequential scene records. Side Prompts usually update or overwrite one continuing support document.
- Generation vs retrieval: Creating an entry does not guarantee that SillyTavern later activates it.
- SillyTavern 1.18.0 or later; the latest compatible release is recommended.
- A working AI connection.
- A model capable of following instructions and, for Memory and Consolidation workflows, returning valid JSON.
- Permission to install third-party SillyTavern extensions.
- A Chat Completion preset available in SillyTavern when using a local or Text Completion backend through an OpenAI-compatible Chat Completion endpoint.
OpenAI, Anthropic/Claude, OpenRouter, Gemini/Google, and other Chat Completion connections can normally use the built-in Current SillyTavern Settings profile.
KoboldCpp, llama.cpp, TextGen, Ollama, and similar backends generally work most reliably when exposed through an OpenAI-compatible Chat Completion endpoint. Even when normal roleplay uses Text Completion, SillyTavern must have a Chat Completion preset available for STMB.
Typical KoboldCpp setup:
- API type: Chat Completion;
- source: Custom OpenAI-compatible;
- endpoint such as
http://localhost:5001/v1orhttp://127.0.0.1:5000/v1; - any nonblank custom API key if SillyTavern requires one;
- model ID in the endpoint’s expected format, commonly
koboldcpp/modelname, without an unnecessary.ggufsuffix; - Chat Completion preset imported;
- response length at least 2048 tokens, with 4096 often safer.
Typical llama.cpp setup:
- API type: Chat Completion;
- source: Custom OpenAI-compatible;
- endpoint
http://localhost:8080/v1, orhttp://host.docker.internal:8080/v1when SillyTavern runs in Docker; - any nonblank API key if required by SillyTavern;
- the served model ID;
- no prompt post-processing unless the endpoint requires it.
Example server command:
llama-server -m <model-path> -c <context-size> --port 8080STMB works without Chat Top Bar / Chat Top Info Bar. Installing it adds the Memory Books Jobs queue interface for active, completed, failed, canceled, blocked, and review-needed work.
- Open SillyTavern.
- Open the main Extensions panel.
- Choose Install Extension.
- Install the official Memory Books repository.
- Reload SillyTavern if prompted.
- Open a character chat or group chat.
- Wait several seconds for STMB controls to initialize.
SillyTavern Extras is not required.
At least one of these should appear:
- Memory Books in the magic-wand Extensions menu beside the chat input;
- scene chevrons ► and ◄ in expanded message actions.
If neither appears:
- wait up to ten seconds;
- refresh the page;
- verify that the extension is installed and enabled;
- reopen a character or group chat;
- inspect the browser console only after the basic checks fail.
Open the magic-wand Extensions menu near the chat input, then choose Memory Books.
The panel can include:
- Current Scene;
- Memory Status / highest processed message;
- Current Lorebook Configuration;
- Memory Profiles;
- Profile Actions;
- Extra Function Buttons;
- Prompt Managers;
- General Settings;
- Automatic Memories;
- Token Saving;
- Group-character or Narrator controls when relevant.
For a first Memory, only three decisions are required:
- Which Memory Book will receive the entry?
- Which profile/connection will generate it?
- Which chat messages are the scene?
Automatic Mode is the normal default. STMB uses the lorebook bound to the current chat through SillyTavern.
Use it when:
- one chat has one primary Memory Book;
- minimal configuration is preferred;
- group characters do not need separate Memory Books.
If no lorebook is bound, bind one in SillyTavern or use Auto-Create.
Enable Auto-create lorebook if none exists to let STMB create and bind a lorebook when a Memory is first saved.
The default naming template can use:
{{char}}— character or group name;{{user}}— user name;{{chat}}— chat ID/name.
STMB adds numeric suffixes when necessary to avoid duplicate names.
Auto-Create and Manual Lorebook Mode are mutually exclusive.
Enable Manual Lorebook Mode to choose a Memory Book independently of the lorebook bound to the chat.
Use it when:
- memories must live in a dedicated lorebook;
- multiple chats intentionally share one Memory Book;
- group members need separate books;
- Narrator Mode is used;
- the user understands the resulting activation plan.
The main manual Memory Book selection is stored for the current chat unless a persistent character lock overrides it in a compatible solo chat.
A dedicated Memory Book makes it easier to:
- separate memories from character definitions and setting lore;
- set an independent lorebook budget and order;
- reuse or export memory history;
- inspect STMB-managed entries without unrelated lore;
- diagnose activation.
It is a recommendation, not a requirement.
A character Memory Book lock is a persistent Manual Mode assignment attached to a character card.
In a solo chat:
- an unlocked manual book belongs to the current chat;
- a locked book follows the character card across compatible Manual Mode chats;
- the manual book cannot be changed until the lock is removed.
In a real group chat:
- an unlocked per-character assignment belongs to the current group chat;
- a locked per-character assignment follows that character card into compatible Manual Mode groups;
- a missing locked book produces a broken-lock state that must be unlocked or repaired.
Use locks only when the same character should intentionally share one continuing Memory Book across stories. They are dangerous for alternate universes or unrelated timelines.
- Solo chat: one chat-bound or auto-created Memory Book.
- Real group chat: one group Memory Book.
- Narrator chat: one omniscient Memory Book plus one unique book per declared character, as required by Narrator Mode.
A Memory Books profile controls both generation and the resulting lorebook-entry settings.
Use Current SillyTavern Settings first. It uses the provider, model, and temperature currently active in SillyTavern.
Do not begin by rewriting prompts or configuring a Full Manual endpoint. First prove that one Memory can be generated and saved.
Create a separate profile when the user needs to:
- use a cheaper or more reliable model for memories;
- use a different provider from roleplay;
- bind a named Custom connection;
- choose a custom summary prompt;
- use different temperature or maximum output behavior;
- change title formatting;
- change activation, insertion, order, or recursion settings;
- use separate group/omniscient and character-focused prompts.
A profile may include:
- display name;
- API/provider;
- model ID;
- temperature;
- Summary Prompt preset;
- optional separate multi-character prompts;
- structured-output behavior;
- optional SillyTavern ChatCompletionService routing;
- optional Chat Completion preset;
- reverse-proxy behavior;
- title format;
- activation mode: Normal, Constant, or Vectorized;
- insertion position, including character, example-message, author’s-note, and Outlet positions;
- Outlet name when applicable;
- automatic or manual order value;
- Prevent Recursion;
- Delay Until Recursion.
A Custom OpenAI-compatible profile can:
- use the currently active SillyTavern Custom connection; or
- bind one named Custom connection from SillyTavern’s Connection Manager.
The named connection supplies its saved URL and secret. The model field in the STMB profile remains the model override. If the named connection is deleted or ceases to be a Custom Chat Completion connection, STMB blocks the request rather than silently routing elsewhere.
Skip structured output and use plain-text completion prevents STMB from sending a structured-output schema to providers that reject it. The model must still return the valid JSON demanded by the selected Memory or Consolidation prompt.
Use ST’s ChatCompletionService routes supported profile requests through SillyTavern’s request helper and can apply a selected SillyTavern Chat Completion preset. Full Manual profiles do not use this route.
Use reverse proxy forwards SillyTavern’s configured reverse-proxy details for supported providers.
Full Manual Configuration stores a separate endpoint and key inside the STMB profile. It is an exceptional path. Prefer a provider or Custom connection configured and tested in SillyTavern whenever possible.
The global STMB maximum response-token setting can override normal Chat Completion output length for Memory Books work. Cut-off JSON is a common reason for failed generation. Increase output length before weakening the schema or prompt.
A scene is the inclusive chat-message range STMB processes into one Memory.
Useful boundaries normally contain one coherent unit:
- an event;
- a conversation;
- an investigation step;
- an emotional or relationship development;
- a location or goal change;
- a connected action sequence.
Very small trivial ranges may produce little value. Very large ranges cost more, are harder to summarize, may exceed context, and often combine unrelated events.
- Expand the message actions, usually through a three-dot or similar control.
- Click ► on the first included message.
- Click ◄ on the last included message.
- Open Memory Books and verify the displayed start, end, speakers, message count, and token estimate.
Both boundary messages are included.
Use Clear Scene to remove the selection, or choose another start/end marker to replace one boundary.
- Verify the scene.
- Verify the effective Memory Book.
- Verify the selected profile.
- Click Create Memory, or use
/creatememory. - Review confirmation, token warning, participant confirmation, or preview windows when shown.
- Approve the result.
- Confirm that a new lorebook entry exists and that Memory Status advanced to the end of the scene.
A valid Memory result normally contains:
- a title;
- content;
- keywords;
- STMB metadata, including source range and chat identity.
When Show memory previews is enabled, review and optionally edit:
- title;
- memory content;
- keywords.
Check names, attribution, facts, omitted consequences, and unrelated commentary. Without previews, a valid result is saved automatically.
Enable Auto-create memory summaries and configure:
- Auto-Summary Interval — number of new messages processed per automatic Memory;
- Auto-Summary Buffer — newest messages left out so an unfolding scene is not summarized too early.
Example:
Interval: 30
Buffer: 2
STMB waits until at least 32 messages exist beyond the processed boundary, then creates a Memory ending two messages before the newest message.
If no processed baseline exists, current STMB treats the baseline as -1 and can begin at message 0. A manual first Memory remains recommended for setup validation and for choosing a deliberate starting point.
Lower intervals create more focused Memories and more requests. Higher intervals create fewer, larger Memories with a greater risk of combining unrelated material. A practical starting range is approximately 20–40 messages for detailed roleplay and 40–60 for shorter, faster exchanges.
Automatic generation can be postponed when a required Memory Book is not yet assigned.
STMB stores the highest processed message for each chat. It determines:
- where
/nextmemorystarts; - where automatic Memories start;
- the memory-boundary indicator;
- which messages count as already processed.
Use:
/stmb-highestto display it;/stmb-set-highest <N>to set it manually;/stmb-set-highest noneto clear it.
Manual changes should be deliberate because they can cause skipped or repeated ranges.
Use:
/stmb-catchup interval=<chunk size> start=<first message id> end=<last message id>
Example:
/stmb-catchup interval=40 start=0 end=245
The range is inclusive. Chunks are processed consecutively; the final chunk may be smaller.
Catch-up is intentionally non-interactive. Before running it:
- select and test the intended profile;
- enable Always use default profile;
- disable Show memory previews;
- ensure the effective Memory Book exists, or permit Auto-Create in Automatic Mode;
- repair all required multi-character book assignments;
- choose a chunk size below the token-warning threshold.
STMB preflights every chunk, processes in order, and stops on the first failure or /stmb-stop. Completed earlier chunks remain saved. Resume from the first unfinished message rather than repeating the whole range.
Use catch-up for broad conversion. Manual scene boundaries remain better when literary or event boundaries matter.
9. Token Saving, Hidden Messages, and the Memory Boundary
Hidden messages remain in the chat file. They are omitted from active chat context until revealed again.
Auto-hide messages after adding memory can be:
- Do not auto-hide;
- Auto-hide all messages up to the last Memory;
- Auto-hide only messages in the last Memory.
Messages to leave unhidden preserves a small recent overlap near the boundary.
Unhide hidden messages for memory generation temporarily reveals a selected range before STMB compiles it. Use this when regenerating or reprocessing ranges that were previously hidden.
The indicator uses the highest processed message to show where processed history ends and unprocessed chat begins.
Modes:
- Off;
- Memory boundary divider;
- draggable jump button;
- divider plus jump button.
The jump button scrolls toward the first unprocessed message and remembers its dragged position.
A practical initial setup is:
- show the boundary divider and jump button;
- leave two messages unhidden;
- enable temporary unhide for generation;
- use no auto-hide until the user has confirmed that a Memory was saved correctly;
- then switch to hiding all processed messages for the main token-saving benefit.
Normal Memories are commonly keyword-triggered. Good keywords are concrete and distinctive:
- character names and aliases;
- named locations or organizations;
- important objects;
- event names;
- identifiers;
- specific discoveries or actions.
Weak keywords such as important event, conversation, or secret are too broad.
The memory content determines what the model learns. Keywords help determine when SillyTavern retrieves it.
- Normal: keyword/rule-driven activation.
- Constant: always active, subject to applicable budget and entry controls.
- Vectorized: uses vector-related retrieval when the user’s setup supports it.
Vectors are optional. STMB works through keywords without the Vectors extension.
Common starting recommendations:
- Match Whole Words: off;
- Scan Depth: relatively high, such as 8;
- Max Recursion Steps: approximately 2;
- Context percentage: sized to the user’s total context and competing prompt material.
These are recommendations, not hard requirements.
If the Memory Book is the only active lorebook/World Info source, leave Delay Until Recursion disabled. Otherwise no entry may start the first recursion cycle and the Memory may never activate.
When an AI “does not remember”:
- Verify that the entry exists.
- Verify that the correct Memory Book is active for the chat.
- Verify that the entry is enabled.
- Verify that keywords or activation mode match the current conversation.
- Verify that the lorebook budget is sufficient.
- Verify recursion settings.
- Use a World Info inspection tool or request log to confirm whether the entry was actually sent.
- If it was sent but ignored, the remaining problem is model behavior or competing context, not STMB storage.
Group Chat Mode applies to a real SillyTavern group containing two or more separate character cards.
SillyTavern Group
├── Alice character card
├── Bob character card
└── Clara character card
SillyTavern records which card authored each message, so STMB can preserve speaker attribution and detect participating group members.
No separate Group Chat Mode switch is required. Open a group chat and use STMB normally.
A detected participant is normally a character card that authored at least one message inside the selected scene.
STMB does not infer every person physically present from prose. Therefore:
- a silent observer may not be detected;
- a merely mentioned character is not a participant;
- an absent character discussed by the group is not selected;
- the user is not treated as a separate group-character Memory Book target;
- duplicate or unusual speaker identities may need correction.
The participant prompt means: Which group characters should this Memory be associated with? It does not prove who knew every fact or who was physically present.
This is the recommended starting layout.
Use Automatic Mode, Auto-Create, or a main Manual Mode book. Each scene produces one canonical entry in the group Memory Book. When participant names are available, the entry can receive an inclusive SillyTavern character filter.
An inclusive filter for Alice and Bob means the entry can activate when Alice or Bob is active. It does not create a synthetic “Alice and Bob” character or a separate subset book.
One group book is best when:
- the cast mostly shares one story;
- one omniscient/group-oriented summary is sufficient;
- minimal setup and fewer duplicate entries are preferred;
- STLO is not needed.
A single group Memory can still preserve asymmetric knowledge:
Alice found the transmitter and hid it. Bob believed the room was empty.
The advanced real-group layout uses:
- one canonical group Memory Book;
- one assigned character Memory Book for each group member.
Requirements:
- Manual Lorebook Mode;
- SillyTavern-LorebookOrdering (STLO) installed and enabled;
- a valid assignment for every required group member.
The canonical group book cannot also be a character book. More than one character may share the same character book; STMB writes one copy to that shared book rather than duplicates.
When a Memory is saved:
- the canonical version is written to the group book;
- participant selection is confirmed unless automatic acceptance is enabled;
- linked copies are written to selected participant books;
- STMB rolls back partial writes when possible if one required save fails.
Selecting no participants in the real-group participant confirmation applies the Memory to every current group member.
By default, the same group-oriented Memory is copied to participant books.
A profile can enable Use separate group and character prompts in group chats. Then:
- the Group Summary Prompt writes the canonical group version;
- the Character Summary Prompt writes an individualized version for each single-character target book.
Character-focused versions can preserve:
- private knowledge;
- mistaken beliefs;
- personal emotional reactions;
- relationship-specific priorities;
- what mattered to one participant.
This requires additional AI requests. A shared character book receives one shared copy rather than one duplicate per assigned character.
Memory Books decides:
- scene range;
- participants;
- summary content;
- which books receive copies;
- whether individualized prompts are used.
STLO decides:
- when a lorebook is active;
- which character can activate it;
- priority, position, budget, and ordering.
When STMB assigns a character book, it adds the character’s avatar basename to stlo.characterOverrides and enables stlo.onlyWhenSpeaking, while preserving existing STLO priorities, budgets, and overrides.
STMB uses merge-only behavior. Clearing or changing an assignment does not automatically remove the old STLO character override. Remove obsolete overrides in STLO manually.
Separate books and filters improve relevance. They do not guarantee that:
- one character can never receive another character’s information;
- the model never sees the canonical group version;
- previous-memory context is perfectly knowledge-partitioned;
- a character book represents only conscious knowledge.
Use them as context-routing tools, not as security boundaries.
Linked entries share metadata that lets STMB recognize the same original event, but later edits are independent.
Editing, deleting, compacting, or regenerating one copy does not automatically change the others. Regenerate or edit each version separately when all copies need the same change.
Adding a character:
- assign a valid book before the next distributed Memory;
- old Memories are not copied retroactively;
- old filters are not rewritten;
- manually provide historical context if needed.
Removing a character:
- existing entries remain;
- old filters and STLO overrides remain;
- linked copies are not deleted automatically.
Changing a character’s book:
- changes future routing;
- does not necessarily remove that character from the old book’s STLO overrides.
The canonical group book uses the automatic group-chat consolidation analysis prompt, which aims for an omniscient chronology while distinguishing objective events from individual knowledge.
Character books use the consolidation preset selected in the popup. Books may have different numbers of eligible sources. A book without enough material can be skipped with a warning while ready books continue.
A missing scene in a character book is a chronology gap. It does not prove absence, ignorance, or unconsciousness. A shared character book receives one consolidated entry.
Narrator Mode is for a normal one-on-one SillyTavern chat where one Narrator character card writes multiple fictional characters.
Normal SillyTavern Chat
└── Narrator card
├── writes Alice
├── writes Bob
└── writes Clara
Without Narrator Mode, SillyTavern sees all AI responses as authored by the Narrator card. Narrator Mode supplies a manual cast model so STMB can associate scenes and Memory Books with fictional characters inside the Narrator’s prose.
Narrator Mode is not available inside a real SillyTavern group chat.
Narrator Mode requires:
- Manual Lorebook Mode;
- one selected omniscient/canonical Memory Book;
- one unique Memory Book for every declared cast member.
Rules:
- a cast member cannot use the omniscient book;
- two cast members cannot share the same book;
- every declared member must have an available book;
- retired members retain their identity and reserved book assignment until restored or otherwise removed by the implementation;
- Auto-Create is incompatible because Narrator Mode depends on Manual Lorebook Mode.
Unlike the advanced real-group layout, Narrator Mode does not require STLO for active-character retrieval. STMB injects the selected cast members’ books into the active lorebook context during generation.
- Open the Narrator card’s normal chat.
- Enable Manual Lorebook Mode.
- select the main manual book; this is the omniscient Memory Book.
- Enable Narrator Mode.
- Open Manage Narrator Cast.
- Add each fictional character by name and assign a unique Memory Book.
- Use the floating Active Cast drawer to select characters present in the next exchange.
Narrator Mode must be disabled before Manual Lorebook Mode can be disabled.
The floating Active Cast drawer can be expanded, collapsed, moved, and used to select current cast members.
At generation time, STMB snapshots the active cast and stores it in message metadata:
- the user message receives the active-cast snapshot;
- the Narrator response receives the generation snapshot;
- a continuation merges its cast with existing cast metadata;
- swipe metadata is stored separately for each swipe;
- selecting a swipe can restore the active cast from that timeline point;
- deleting recent messages can restore cast state from the latest remaining tagged Narrator message.
The cast marker records association, not a semantic analysis of prose.
When a Narrator generation starts, STMB loads the Memory Books of the active cast and merges their entries into the character-lore collection used for that request, avoiding duplicate world/UID pairs.
Consequences:
- only active-cast books are added by this Narrator workflow;
- the omniscient book still follows its normal Manual Mode activation/configuration;
- per-character STLO filters are not required for Narrator Mode;
- cast selection must be correct before generation if the correct character books are expected in context.
For a selected scene, tagged Narrator responses are authoritative. STMB combines the cast IDs stamped on Narrator-authored messages.
If the scene contains untagged legacy Narrator messages, STMB falls back to continuity information from all messages and asks the user to confirm the scene cast. Current active cast members are preselected. An empty selection means no individual cast members were present.
This confirmation is specifically for legacy or incomplete cast metadata; fully tagged scenes do not need it.
A Narrator scene Memory is written as:
- one canonical omniscient entry in the main Memory Book;
- one linked copy in each selected participant’s unique Memory Book.
Narrator copies do not use native SillyTavern character filters. Instead, STMB stores Narrator participant and owner IDs in entry metadata.
If separate multi-character prompts are disabled, participant books receive copies of the omniscient summary. If enabled, each single-character book can receive a character-focused generation.
Narrator ownership and participant metadata are carried through consolidation sources. This lets higher-tier entries retain which character book owns a copy and which cast members participated in the underlying material.
Regeneration uses this metadata to determine whether the replacement prompt target is omniscient/group-oriented or character-focused.
As with real-group copies, linked Narrator entries are not live-synchronized after creation.
The cast manager can mark a member retired and later restore them. Retired members:
- are removed from active-cast choices;
- are removed from the active-cast ID set;
- keep stable identity/history metadata;
- retain their book reservation, preventing accidental reuse that would merge identities.
Use retirement for a character who leaves the active cast but whose historical Memory identity must remain intact.
SillyTavern native branches can become different continuities. If a branch and its parent write into the same unlocked Memory Books, contradictory timelines can mix.
Copy Memory Books when branching is enabled by default.
When STMB recognizes a newly created native branch:
- Automatic Mode copies the active chat-bound Memory Book;
- Manual Mode copies the main manual Memory Book;
- a Manual Mode real group copies each unique unlocked character Memory Book;
- Narrator Mode copies the omniscient book and each declared character book;
- persistent real-character locks are preserved instead of copied because a lock means “continue using this same book.”
All books copied in one branch operation use the same available lineage number:
Group Memories Branch 1
Alice Memories Branch 1
Bob Memories Branch 1
Branching from an existing branch retains the original lineage root rather than producing names like Branch 1 Branch 1.
Inside the copies, STMB:
- rewrites matching parent chat IDs to the new branch chat ID;
- redirects canonical group/character links when both linked books were copied;
- updates the new branch’s bindings to point to the copies.
It clones existing contents; it does not regenerate Memories.
Do not switch chats while branch copying is in progress.
If copying fails, STMB clears the new branch’s inherited writable bindings and records the failure so the branch cannot silently write into the parent’s originals.
Disable the setting only when the branch is intentionally meant to share the same Memory Books and continuing history as the parent.
A Clip saves selected chat text directly into a [STMB Clip] lorebook entry. It does not call an AI model.
- a preference;
- a promise or secret;
- a name or alias;
- an item or pet;
- a short relationship fact;
- a line that should be preserved exactly or nearly exactly;
- a quick “note to self” that does not justify a scene Memory.
- Highlight text inside a chat message.
- Click the floating scissors button.
- Choose an existing Clip entry or create a new one.
- Choose always-active or keyword-triggered behavior for a new entry.
- Review the current entry and updated preview.
- Rename if needed.
- Save.
The floating scissors button appears only after chat text is selected and can be disabled in the main panel.
Title:
Seraphina Healed Me [STMB Clip]
Content:
=== Seraphina Healed Me ===
- Seraphina healed the user’s wounds with magic.
=== END Seraphina Healed Me ===One Clip entry has one section. Focused titles support focused activation keywords.
An existing entry can be treated as a Clip entry by adding [STMB Clip] to the end of its title. Long Clip entries can be edited manually or compacted.
Clips save only the chosen text. They do not add source attribution automatically.
A Topical Clip reads confirmed STMB Memory entries from one selected Memory Book and asks an AI to produce a focused “about this topic” entry. Eligible sources can include scene Memories and consolidated summaries; Clip and Side Prompt entries are excluded.
Information about one subject is spread across several Memories, for example:
- a recurring NPC;
- a relationship history;
- a location or faction;
- an investigation or mystery;
- powers, injuries, promises, preferences, or secrets;
- an important object;
- an unresolved plot thread.
Topical Clip is organized by subject, not by the chronology of every source Memory.
Topical Clip uses:
- confirmed STMB Memory entries from the selected source book, including eligible consolidated summaries.
It does not use:
- raw chat messages;
- ordinary Clip entries;
- Side Prompt entries;
- unrelated ordinary lorebook entries.
- Open Memory Books.
- Click Topical Clip.
- Choose the source Memory Book.
- Enter the topic.
- Enter activation keywords, or leave them blank to use the topic.
- Choose a new entry or an existing
[STMB Clip]update target. - Optionally select only specific source Memories.
- Choose the generation profile.
- Generate the draft.
- review and edit it.
- Save only when correct.
The generated draft is never saved automatically.
After a successful run, STMB records which source Memories were used. A later update normally sends only new or changed source Memories together with the existing Clip content.
Use Rebuild from all source memories when:
- the current entry is incomplete or disorganized;
- the prompt changed;
- older Memories were substantially edited;
- the whole topic should be reconsidered.
Use Use only selected memories when the book is large, the topic is limited to one story period, names overlap, or strict evidence control is needed.
STMB estimates request size and warns when the configured token threshold is exceeded. Reduce sources, raise the threshold deliberately, or run once anyway.
Check that the draft:
- stays on topic;
- preserves names and relationships;
- includes major relevant facts;
- identifies contradictions rather than silently choosing one version;
- does not invent explanations unsupported by source Memories;
- merges updates without unnecessary duplication.
A custom Topical Clip prompt must include:
{{SOURCE_MEMORIES}}
Supported placeholders include:
{{MODE}}
{{TOPIC}}
{{KEYWORDS}}
{{EXISTING_CLIP}}
{{EXISTING_ENTRY_CONTENT}}
{{SOURCE_MEMORIES}}
Reset to Default if a custom prompt stops producing useful output.
A Side Prompt is a named STMB prompt that runs separately from the normal character reply. It usually creates or updates one continuing support entry rather than another sequential scene Memory.
In the Trackers & Side Prompts list, the power icon immediately changes the prompt-wide Enabled flag: green means enabled and dim means disabled. This control does not add, remove, or otherwise change the prompt's configured triggers.
- plot and unresolved-thread trackers;
- relationship state;
- NPC or faction status;
- inventory and resources;
- injuries, statistics, or reputation;
- timelines, dates, deadlines, and travel;
- mystery clues, suspects, and contradictions;
- inventions, research, and projects;
- continuity-risk reports;
- world-state summaries.
Avoid vague “track everything” prompts, duplicate scene summaries, or tasks that must appear inside the next roleplay response.
Side Prompts normally expect final plain text or Markdown ready to save. They do not require Memory JSON. JSON is allowed only when the user intentionally wants JSON stored as tracker text.
A typical run assembles:
- Side Prompt instructions;
- prior saved tracker entry, if any;
- optional previous Memories;
- optional Additional Context;
- selected or since-last scene text;
- optional Response Format instructions.
The prior entry is existing state to revise, not proof that every old statement should remain. Prompts should explicitly remove stale, resolved, contradicted, or duplicate information.
/sideprompt "Prompt Name"
/sideprompt "Prompt Name" 10-20
/sideprompt "Relationship Tracker" {{npc name}}="Alice" 10-20
Names with spaces should be quoted. A supplied range is inclusive.
Manual runs are best for targeted analysis and prompts requiring runtime macro values.
A Side Prompt can enable Run automatically after memory.
The chat then uses one of two automatic selection modes:
- individually enabled Side Prompts; or
- one selected Side Prompt Set.
A selected set replaces individually enabled automatic prompts for that chat. It does not add to them.
A Side Prompt can enable Run on visible message interval and specify a number of visible messages since its checkpoint.
Hidden and system messages do not count.
When a set is active, only rows in that set whose referenced prompt has the appropriate interval trigger are candidates.
A Side Prompt Set is an ordered run list, not merely a folder. The same template may appear more than once with different macro values.
Example:
- Relationship Tracker — Alice
- Relationship Tracker — Bob
- Plot Tracker
- Cleanup Report
Rows can store:
- a prompt reference;
- an optional label;
- runtime macro values;
- order;
- duplicate or delete actions.
Rows run top to bottom.
Manual set commands:
/sideprompt-set "Set Name"
/sideprompt-set "Set Name" 10-20
/sideprompt-macroset "Relationship Pass" {{npc_1}}="Alice" {{npc_2}}="Bob" 10-20
General Settings can define:
- a default set for solo chats;
- a default set for group chats.
Each chat may:
- inherit the applicable default;
- explicitly use individually enabled prompts;
- choose a named set.
An empty global default means individual mode.
If a selected set is deleted, STMB warns rather than silently substituting another workflow. A missing row prompt or unresolved macro causes that row to be skipped with a warning.
The set selects candidate rows. Each referenced Side Prompt still needs the relevant automatic trigger for after-Memory or interval execution. Manual set commands do not require those trigger checkboxes.
Side Prompts can use normal SillyTavern macros such as:
{{user}}
{{char}}
Non-standard {{...}} placeholders are runtime macros. They must be supplied manually or stored in a set row.
Examples:
{{npc name}}
{{faction}}
{{project_name}}
A prompt with unresolved runtime macros cannot run automatically. Automatic runs cannot pause to ask for values.
STMB registers integer macros for the effective main Memory Book:
| Macro | Count |
|---|---|
{{memtier0}} |
scene Memories |
{{memtier1}} |
Arcs |
{{memtier2}} |
Chapters |
{{memtier3}} |
Books |
{{memtier4}} |
Legends |
{{memtier5}} |
Series |
{{memtier6}} |
Epics |
{{memclips}} |
Clip entries |
{{memside}} |
Side Prompt entries |
The effective main book is the chat-bound book in Automatic Mode or the resolved main manual book in Manual Mode. In a multi-book group or Narrator setup, the counts do not add all character books together.
A count macro provides only a number, not the contents of those entries.
An explicit range uses exactly that inclusive range. Without a range, STMB uses the Side Prompt’s since-last checkpoint/cap behavior.
Use explicit ranges for debugging, targeted cleanup, or rerunning a known section.
A Side Prompt may include up to seven previous scene Memories.
Its Additional Context source can be:
- none;
- Follow chat, using the chat’s selected Context Setting;
- one fixed named Context Setting.
These are reference materials. The prompt should not blindly copy them into the tracker.
A Side Prompt normally saves to the effective Memory Book. It can instead use:
- a per-chat target override;
- a template-level target;
- the effective Memory Book as fallback.
A valid per-chat override wins.
Use alternate targets for a deliberate shared campaign book or dedicated tracker book. Do not scatter trackers without a retrieval plan.
A template can configure:
- title override;
- keywords;
- Normal, Constant, or Vectorized activation;
- insertion position and Outlet name;
- order mode/value;
- Prevent Recursion;
- Delay Until Recursion;
- Ignore Budget.
Title and keyword fields can expand applicable macros. Ignore Budget should be used sparingly because multiple always-included trackers can consume large amounts of context.
A Side Prompt can inherit normal Memory Books connection resolution or bind a specific STMB profile. An override is useful for a cheaper model or one better at structured maintenance. Excessive profile combinations make troubleshooting harder.
Compatible saves store a compact snapshot containing:
- Side Prompt template key;
- prior entry content;
- source chat and inclusive range;
- runtime macro values.
To regenerate, open the lorebook editor and click Regenerate side prompt. The replacement uses the saved snapshot with the current template and current profile/context settings.
Regeneration cannot complete when the template was deleted, the source chat/range is unavailable, or the target/source changed during generation. Only the content is replaced; existing title, keywords, and entry settings remain.
A good Side Prompt defines:
- the exact maintenance job;
- what source material to review;
- whether to revise, replace, merge, or append;
- stale information to remove;
- stable output headings and ordering;
- a strict length limit;
- final-output-only behavior.
Example:
Update the relationship tracker from the supplied scene. Preserve current facts, merge new developments into the existing sections, and remove resolved, contradicted, stale, or duplicate details. Keep each relationship to 1–3 concise bullets. Output only the updated tracker.
Useful guards:
Do not append a new section unless there is genuinely new information.
Remove resolved threads and obsolete speculation.
Output only the updated report; no preface or explanation.
Keep the entire output under 300 words.
Stable headings reduce drift across repeated updates.
If a prompt did not run:
- confirm the Memory or interval event actually occurred;
- inspect the chat’s individual/set selection;
- verify the referenced prompt still exists;
- verify the relevant automatic trigger is enabled;
- verify all runtime macros have values;
- check whether
/stmb-stopor a failed job canceled it.
If it ran twice:
- check manual plus automatic invocation;
- duplicate set rows;
- duplicate prompt copies;
- multiple tabs or chats triggering work.
If the wrong book received it, inspect both per-chat and template-level target scopes.
If output grows indefinitely, add explicit replacement, pruning, item-count, and word-count rules.
Consolidation combines lower-tier STMB Memories or summaries into higher-tier chronological recaps.
Scene Memory → Arc → Chapter → Book → Legend → Series → Epic
Consolidation works from existing STMB entries, not directly from raw chat.
Use it when:
- scene Memories are accumulating;
- older material no longer needs full scene detail;
- a major relationship, plot, or campaign phase has completed;
- token use should be reduced while preserving continuity;
- a cleaner higher-level chronology is needed.
Consolidated entries should emphasize lasting changes, turning points, goals, consequences, relationship shifts, unresolved threads, and stable state.
- Open Consolidate Memories.
- Choose the target tier.
- Select eligible source entries.
- Choose the consolidation prompt/profile settings.
- Decide whether source entries should be disabled after successful consolidation.
- Run and review the candidates.
- Approve the desired summaries.
Prompt for consolidation when a tier is ready watches selected target tiers. When the saved minimum eligible count is reached, STMB presents a yes/later prompt. Choosing Yes opens the consolidation interface. It does not silently consolidate.
Ordinary consolidation expects strict JSON:
{
"summaries": [
{
"title": "Short higher-tier title",
"summary": "Consolidated chronological recap",
"keywords": ["keyword1", "keyword2"],
"member_ids": ["001", "002"]
}
],
"unassigned_items": [
{
"id": "003",
"reason": "Why this source did not fit"
}
]
}The model may return one or several summaries. member_ids assigns each source to a returned summary. Outliers belong in unassigned_items rather than being forced into an unrelated recap.
A previous summary in the target tier can be supplied as canon context. It is not source material to rewrite. Consolidation prompts must distinguish it from the lower-tier entries being processed.
Consolidation previews can allow editing, accepting, regenerating one candidate from the same sources, or regenerating a pending batch.
Malformed or failed AI responses can be inspected and, where supported, corrected manually before commit.
When enabled, STMB disables source entries after successful consolidation so the higher-tier summary can take over retrieval. This is reversible through lorebook editing.
They should define:
- the compression target;
- whether to create one recap or the smallest coherent number;
- chronology and grouping logic;
- details that must survive;
- explicit handling of outliers;
- exact JSON structure.
They should preserve major beats, consequences, promises, relationship changes, identifiers, unresolved threads, and retrieval-friendly keywords while removing repeated scene-level detail.
Compaction asks an AI to shorten one existing STMB-managed entry and presents the original and draft before replacement.
[STMB Clip]entries;- Side Prompt entries;
- STMB Memory entries.
Ordinary non-STMB lorebook entries are not listed.
- Open Compaction.
- Choose a Memory Book.
- Choose a Compaction Profile.
- Optionally edit the Compaction Prompt.
- choose one entry.
- Compare original and compacted token estimates/content.
- Edit the draft if needed.
- Replace, copy the draft, or cancel.
The original is not changed until Replace with Compacted Version is selected.
- long Clip collections;
- repeated or stale tracker content;
- wordy scene Memories;
- always-active entries consuming too much context.
Compaction is not for adding facts, summarizing raw chat, creating a new Memory, or processing ordinary lorebook entries.
{{ENTRY_CONTENT}} required current content
{{ENTRY_KIND}} Clip, SidePrompt, or Memory
{{ENTRY_TITLE}} entry title
The prompt should preserve facts, names, pronouns, macros, wrapper headings, and end markers while removing redundancy and low-value wording.
Regeneration creates a reviewable replacement for an existing entry. It does not create a second numbered entry and never overwrites without approval.
- open the source chat;
- open the Memory Book in the lorebook editor;
- click Regenerate memory;
- choose the current profile, prompt, previous-memory count, and Additional Context;
- review title, content, and keywords.
The original scene range and sequence number are retained. If all source messages are hidden, reveal them or enable unhide-before-generation.
A higher-tier summary is regenerated from its exact linked lower-tier sources using the dedicated Regenerate Consolidation preset.
The full source set must still exist at the correct tier. A lower-tier source cannot be regenerated while an active parent summary depends on it; delete the parent first when intentionally rebuilding the lower tier.
See the Side Prompt snapshot rules in Section 16.16.
Immediately before replacement, STMB verifies that:
- the target entry is unchanged;
- the source chat range is unchanged;
- required consolidation sources are unchanged and available;
- the entry remains eligible.
If any check fails, nothing is overwritten.
Linked group, character, and Narrator copies remain independent.
Several context sources can appear in an STMB request. They are not interchangeable.
The message range being processed now. It is the target material for an ordinary scene Memory.
Earlier scene Memories from the effective Memory Book, included as read-only continuity context. The user can normally include 0–7.
They should not be summarized again merely because they appear before the current scene.
Selected lorebook entries supplied as stable reference material, such as:
- character or setting rules;
- canonical names and terminology;
- campaign constraints;
- an authoritative timeline;
- location references;
- facts assumed but not repeated in the scene.
Additional Context appears before previous Memories and the scene transcript. It is reference material, not another scene.
A Context Setting is a reusable ordered collection of Additional Context entries.
Workflow:
- open Context Settings;
- create a named setting;
- select lorebook entries;
- order them;
- choose the setting for the current chat or explicitly choose No Context.
The selection is stored per chat and works with Current SillyTavern Settings as well as saved profiles.
If a referenced book or entry disappears, STMB warns, skips the stale reference, and continues. If the entire Context Setting is deleted, chats referring to it continue without Additional Context until another selection is made.
Context Settings can be duplicated, imported, and exported as stmb-context-settings.json.
The current tracker text to revise. It is state, not evidence that all old statements remain valid.
Lower-tier entries that are the actual material being grouped and compressed.
Canon carried forward during consolidation. It is not a source to rewrite.
Ordinary Memory:
Memory prompt
Additional Context
Previous Memories
Current scene transcript
Side Prompt:
Side Prompt instructions
Prior entry
Previous Memories
Additional Context
Scene text
Response Format
Consolidation:
Consolidation prompt
Previous higher-tier summary
Selected lower-tier source entries
Prompts should label target material and reference-only material clearly.
STMB has three main structured generation systems plus several focused auxiliary workflows.
STMB expects one JSON object:
{
"title": "Short scene title",
"content": "The memory text",
"keywords": ["keyword1", "keyword2"]
}Rules:
- return only the JSON object;
- use the exact keys
title,content, andkeywords; keywordsmust be a JSON array of strings;- keep the title short and readable;
- use concrete retrieval terms;
- place any desired Markdown inside the
contentstring; - escape quotation marks correctly.
STMB can repair some fences, trailing commas, think tags, wrappers, or minor malformed output, but prompts should never depend on recovery.
A strong Memory prompt states:
- the desired memory style and level of compression;
- continuity-relevant information to preserve;
- filler, OOC, or unsupported material to omit;
- the exact JSON schema.
Weak prompts specify style but not structure, ask for analysis instead of a final object, blur previous context with the current scene, or use abstract keywords.
The built-in presets include styles such as:
- Summary — detailed beat-by-beat summary;
- Summarize — structured Markdown-oriented timeline/beats/interactions/outcome;
- Synopsis — comprehensive structured synopsis;
- Sum Up — concise beat summary with timeline;
- Minimal — approximately one or two sentences;
- Northgate — literary summary style for creative writing;
- Aelemar — plot points and character-memory focus;
- Comprehensive — broader synopsis and keyword extraction.
The exact built-in text may be recreated for the current SillyTavern locale. Recreating built-ins removes local edits to those built-ins but should not delete unrelated custom presets.
When separate group/character prompts are enabled, STMB marks the request target as:
groupfor a canonical real-group or omniscient Narrator Memory;characterfor one individual character-book version.
The prompt should explicitly use the target perspective without inventing knowledge not supported by the scene and supplied context.
Side Prompts normally return plain text or Markdown. Write them like maintenance instructions, not Memory prompts.
A strong Side Prompt:
- defines one narrow job;
- explains how to use the previous tracker;
- removes stale state;
- imposes stable headings and length limits;
- returns only the final tracker.
Ordinary consolidation requires the schema in Section 17.5. A strong prompt:
- preserves chronology;
- creates the smallest coherent number of summaries;
- assigns every used source through
member_ids; - identifies leftovers through
unassigned_items; - preserves major changes and unresolved continuity;
- uses concrete keywords.
The dedicated Regenerate Consolidation preset is for one replacement summary and is not selectable as the normal consolidation default.
The prompt must include {{SOURCE_MEMORIES}}, remain focused on the requested topic, distinguish source evidence from inference, merge new material into existing Clip content, and surface contradictions.
The prompt must include {{ENTRY_CONTENT}} and should shorten without adding unsupported facts. It should preserve structural wrappers and macros needed by the entry.
Before finalizing any STMB prompt, answer:
- What material is the actual analysis target?
- What material is reference-only?
- Does this path expect strict JSON or final plain text?
- What information must survive for later retrieval?
- What should be omitted, merged, carried forward, or left unassigned?
Return-format correctness comes before style.
Can create, edit, duplicate, delete, import, and export ordinary Memory prompt presets. Assign a preset through a Memory Books profile.
All ordinary Memory presets must preserve the required Memory JSON schema.
Controls prompts used to group lower-tier entries into higher-tier summaries and selects the normal default consolidation prompt.
The regeneration-only consolidation preset cannot be used for ordinary consolidation.
Built-in prompts can be recreated in the current app locale. Back up locally modified built-ins before recreating them.
STMB integrates with SillyTavern’s Regex extension at two stages:
- Outgoing/User Input: transform the assembled prompt before it is sent.
- Incoming/AI Output: clean or standardize the raw response before parsing/saving.
Enable Use regex (advanced), then open Configure regex and select one or more scripts for each direction.
Important: STMB’s own selection controls execution. A script selected by STMB can run even when that script is disabled in the Regex extension’s normal interface.
Use Regex only when the transformation is understood. A bad outgoing rule can corrupt required schema instructions; a bad incoming rule can corrupt otherwise valid JSON.
Profile title formats can use:
{{title}}— AI-generated title;{{scene}}— source range;{{char}}— character/group name;{{user}}— user name;{{messages}}— scene message count;{{profile}}— profile name;- supported date and time placeholders.
Supported numbering tokens include forms such as:
[0] [00] (0) {0} #0
#[000] ([000]) {[000]}
STMB assigns sequential, zero-padded numbers according to the chosen format.
All printable Unicode characters are allowed in titles, including emoji, accented text, CJK, and symbols. Unicode control characters in U+0000–U+001F and U+007F–U+009F are removed.
Lorebook filenames used by Auto-Create are separately sanitized for filesystem-reserved characters and length.
The optional queue requires Chat Top Bar / Chat Top Info Bar.
The Memory Books Jobs drawer can show:
- queued;
- active;
- completed;
- failed;
- canceled;
- blocked;
- Needs Review.
It can cancel active work, reopen review jobs, inspect failures, retry work, and dismiss terminal history rows.
Retry scopes:
- Retry: rerun one non-Memory job, such as a Side Prompt or consolidation job.
- Retry All: rerun/resume the Memory and associated after-Memory Side Prompt work. If the Memory was already saved, STMB can resume from that result rather than duplicate it.
- Retry Memory: rerun/resume only the Memory and intentionally skip after-Memory Side Prompts.
Use Retry All to restore the combined workflow; use Retry Memory when tracker work should not run.
Without Chat Top Bar, STMB still performs its normal workflows but lacks the queue UI.
STMB provides visual states for scene controls, including inactive, selected, valid range, in-scene, and processing states. Exact colors depend on the SillyTavern theme.
Accessibility support includes:
- keyboard navigation;
- focus indicators;
- ARIA attributes;
- reduced-motion behavior;
- mobile-friendly controls.
When teaching from a screenshot, describe the visible icon and label rather than relying on a specific color.
The precise layout can change between releases, but the following settings define current behavior.
- Always Use Default Profile: skip normal confirmation windows unless another warning/review is required.
- Show memory previews: review/edit Memories and applicable Side Prompt outputs before save.
- Show consolidation previews: review consolidation candidates.
- Show notifications: enable toast messages.
- Show floating Clip button: show scissors after chat text selection.
- Refresh Editor: refresh the lorebook editor after writes.
- Memory boundary indicator: off, divider, jump button, or both.
- Allow Scene Overlap: allow a selected range to overlap an existing Memory range.
- Max Response Tokens: output-length override for STMB generation.
- Token Warning Threshold: request-size threshold that triggers a warning.
- Default Previous Memories: normally include 0–7 previous Memories.
- Manual Lorebook Mode: use per-chat manual Memory Book selection.
- Auto-create lorebook if none exists: create/bind a book in Automatic Mode.
- Lorebook Name Template: name Auto-Created books with macros.
- Copy Memory Books when branching: clone active unlocked books for native branches.
- Narrator Mode: use one omniscient manual book plus unique declared-character books in a normal Narrator-card chat.
- Character Memory Book locks: persist a character-card book assignment across compatible Manual Mode chats.
- Auto-create memory summaries: enable automatic scene Memories.
- Auto-Summary Interval: messages per automatic Memory.
- Auto-Summary Buffer: newest messages excluded from the current automatic range.
- Unhide hidden messages before memory generation: temporarily reveal source ranges.
- Auto-hide messages after adding memory: none, all processed, or last range.
- Messages to leave unhidden: recent overlap preserved near the boundary.
- Prompt for consolidation when a tier is ready: show a yes/later readiness prompt.
- Auto-Consolidation Tiers: tiers monitored for readiness; this does not perform silent consolidation.
- Default After-Memory Side Prompt Set for solo chats.
- Default After-Memory Side Prompt Set for group chats.
- Per-chat inheritance, individual mode, or named set selection is configured in the Side Prompts interface.
- Use regex (advanced): enable STMB-specific outgoing/incoming script selection.
- title format;
- activation mode;
- position and Outlet;
- order mode/value;
- Prevent Recursion;
- Delay Until Recursion.
/creatememory
Create a Memory from the currently marked scene.
/scenememory X-Y
Set the inclusive range and create a Memory, for example /scenememory 10-15.
/nextmemory
Create a Memory from the message after the highest processed boundary through the current eligible end.
/stmb-catchup interval=x start=y end=z
Process an existing long chat in consecutive chunks.
/sideprompt "Name" {{macro}}="value" [X-Y]
/sideprompt-set "Set Name" [X-Y]
/sideprompt-macroset "Set Name" {{macro}}="value" [X-Y]
/sideprompt-on "Name" | all
/sideprompt-off "Name" | all
/stmb-highest
/stmb-set-highest <N|none>
/stmb-stop
Stops all in-flight STMB generation everywhere, including Side Prompts. Work already committed remains saved.
Symptoms:
- Memory Books missing from the magic-wand menu;
- chevrons missing;
- no floating Clip button after selection.
Checks:
- extension installed and enabled;
- page reloaded;
- character/group chat open;
- wait up to ten seconds;
- expand message actions;
- inspect console only after those checks.
Both ► and ◄ are required for a marked scene. Verify Current Scene in the panel.
If the range overlaps an existing Memory, choose another range or enable Allow Scene Overlap.
Automatic Mode:
- bind a lorebook to the chat; or
- enable Auto-Create.
Manual Mode:
- select a main manual book;
- repair a deleted selection;
- unlock a broken character lock before changing it.
Real multi-book group:
- STLO must be available;
- every required member needs a valid assignment;
- the group book cannot be reused as a character book.
Narrator Mode:
- Manual Mode must be enabled;
- an omniscient book must be selected;
- every declared member needs a unique non-omniscient book.
Check in this order:
- provider/model/profile are valid;
- response was not truncated;
- maximum response tokens are sufficient;
- selected prompt still requires exact JSON;
- schema was not corrupted by Regex;
- provider supports the selected structured-output mode;
- try Skip Structured Output only if the provider rejects schemas;
- try a more instruction-following model before rewriting the prompt;
- use the raw response/manual JSON correction interface when available.
Common causes include code fences, commentary, a missing key, keywords not being an array, refusal text, or cut-off output.
They were probably auto-hidden. Change Token Saving settings. Hidden messages are not deleted.
Check:
- Auto-create memory summaries enabled;
- enough messages beyond the highest processed boundary;
- interval plus buffer requirement met;
- no postpone checkpoint still active;
- valid Memory Book available;
- no other Memory job currently blocking the trigger;
- current chat not switched during work;
- group generation finished before the trigger is expected.
A first manual Memory is recommended but not technically required in the current version.
Check:
- correct book active;
- entry enabled;
- relevant keywords;
- activation mode;
- budget;
- recursion and Delay Until Recursion;
- STLO routing if used;
- World Info inspection/logs.
Do not regenerate the Memory until retrieval has been tested.
This is model-use behavior. Possible responses:
- make the Memory shorter and more explicit;
- improve insertion position/priority;
- reduce competing context;
- use an OOC reminder;
- choose a model that follows supplied context more reliably.
See Section 16.18. In particular, a selected set suppresses individually enabled prompts outside that set.
Verify:
- readiness prompt enabled;
- target tier selected for monitoring;
- enough eligible source entries exist;
- sources are not already disabled/ineligible;
- the saved minimum count for that tier is met.
Hover or inspect the stated reason. Common causes:
- entry predates required snapshot metadata;
- source chat/range unavailable;
- source entries missing or wrong tier;
- active parent consolidation blocks a lower source;
- original sequence number cannot be determined;
- Side Prompt template deleted.
Check:
- Copy Memory Books when branching was enabled before branch creation;
- it was a native SillyTavern branch;
- source books existed and could be loaded;
- chat was not switched during copying;
- the branch was not previously marked completed/failed;
- locked books were intentionally preserved rather than copied.
Check:
- Active Cast selection before generation;
- whether the message was a continuation that merged cast metadata;
- whether a swipe restored older cast state;
- whether the scene contains legacy untagged messages requiring confirmation;
- whether the declared character was retired;
- whether each character book still exists.
No. Keyword activation is sufficient and is generated automatically. Vectors are optional.
Usually yes for organization, budgeting, reuse, and diagnosis, but it is not mandatory.
No. It can hide processed messages from active context.
Yes. Mark scenes and create Memories only when desired.
Yes in current STMB. With no processed baseline, it begins at message 0 once the interval plus buffer is met. A manual first run is still recommended to verify setup and choose the desired starting boundary.
No. STMB can prompt when a tier is ready, but the user confirms and reviews the operation.
Yes. It is the recommended starting setup and does not require STLO.
When individual continuity, knowledge, speaker-specific retrieval, or character-focused summaries justify the extra setup and AI requests.
No. Group Chat Mode reads separate SillyTavern character-card authors. Narrator Mode manually declares fictional characters written by one Narrator card.
No for its active-cast retrieval path. It does require Manual Lorebook Mode, one omniscient book, and unique per-character books.
No. They are linked for origin/consolidation metadata, not continuous mirroring.
If no other lorebook entry starts recursion, a delayed Memory entry may never activate.
Verify entry retrieval, then enable automatic Memories, choose interval/buffer, enable token hiding, and add Clips or a narrowly defined Side Prompt only when needed. Use Topical Clip and Consolidation after enough Memories exist.
This section preserves only history that affects current use.
- Current documented release: v8.5.0, August 1, 2026.
- SillyTavern requirement: 1.14.0 or later.
- Narrator Mode was added in v8.5.0.
- Branch book copying, Side Prompt regeneration, and character Memory Book locks were added in v8.4.0.
- Multi-character real-group Memory distribution arrived in v8.0.0.
- Additional Context moved from profiles to reusable per-chat Context Settings in v7.0.0; older profile context is migrated.
- Topical Clip was added in v6.10.0.
- Compaction and Clips were added in v6.6.0.
- Side Prompt Sets and per-prompt targets were added in the v6.4–v6.5 period.
- Consolidation became a multi-tier Arc-through-Epic system in v6.0.0; older Arc metadata is migrated.
- Job Queue integration was added in v6.8.0 and remains optional.
- Current profile defaults use Delay Until Recursion disabled unless a user/profile explicitly changes it.
Only entries with the stmemorybooks flag and required metadata are recognized as STMB Memories. Use the supplied lorebook converter for older entries that predate current metadata.
The old bookmark feature was removed from Memory Books in v4.0.0 and split away from the core extension. Do not teach Memory Books bookmark controls as current behavior.
Built-in prompts can be regenerated according to the active SillyTavern language. Back up customized built-ins before recreation.
Side Prompt import is additive. Existing prompts are preserved; imported key conflicts are renamed rather than overwriting the existing prompt.
Memory Books uses Bun for bundling/minification.
bun run buildInstall the repository’s pre-commit build hook with:
bun run install-hooksThe hook builds before commit, stages build artifacts, and aborts if the build fails.
Memory Books is Copyright © 2024–2026 Aiko Hanasaki and licensed under the GNU Affero General Public License v3.0. Modified versions must preserve applicable notices, identify modifications, and comply with AGPL source-availability requirements.
User says “Memory Books is not working.”
│
├─ Is the menu/control visible?
│ ├─ No → installation/loading/UI checks.
│ └─ Yes
│
├─ Can a scene be selected?
│ ├─ No → expand message actions; set both chevrons; inspect overlap.
│ └─ Yes
│
├─ Is there a valid effective Memory Book?
│ ├─ No → bind, auto-create, select manual, or repair multi-book bindings.
│ └─ Yes
│
├─ Does generation return valid complete output?
│ ├─ No → profile, provider, output tokens, JSON schema, Regex, model.
│ └─ Yes
│
├─ Does the entry exist in the intended book?
│ ├─ No → save/rollback/permission/job failure.
│ └─ Yes
│
├─ Does SillyTavern activate and send it later?
│ ├─ No → keywords, activation mode, book binding, budget, recursion, STLO.
│ └─ Yes
│
└─ Does the model use the supplied entry?
├─ No → model compliance, placement, competing context, entry clarity.
└─ Yes → workflow is functioning.
For a new user, teach only this sequence first:
- Open the magic-wand menu and find Memory Books.
- Use Automatic Mode with a bound book or enable Auto-Create.
- Select Current SillyTavern Settings.
- Expand message actions and mark a short complete scene with ► and ◄.
- Create and preview one Memory.
- Open the Memory Book and verify the saved entry.
- Verify that the entry can activate later.
- Enable automatic Memories and choose interval/buffer.
- Enable auto-hide only after explaining that hidden messages are not deleted.
- Introduce Clips, then Side Prompts, then Topical Clip/Consolidation only when the user has a concrete need.
Do not begin with custom prompts, Full Manual endpoints, multiple character books, Regex, or consolidation unless the user’s actual problem requires them.
Memory Books is an external continuity pipeline built on SillyTavern lorebooks:
Select or schedule chat material
→ generate a structured representation
→ save it with retrieval metadata
→ optionally hide processed transcript
→ let SillyTavern retrieve relevant entries later
The system works best when:
- scenes are coherent;
- prompts clearly distinguish target from reference context;
- JSON workflows return exact schemas;
- keywords are concrete;
- Memory Books are deliberately assigned and activated;
- long-running trackers prune stale state;
- consolidation reduces old detail without erasing continuity;
- users verify retrieval rather than assuming saved means sent;
- advanced multi-book routing is used only when its precision is worth the complexity.