SillyTavern Lorebook Not Working? Check These Settings

If a SillyTavern lorebook imports but changes nothing, check whether its Content reaches the generated prompt. An absent entry needs an assignment or activation fix. An entry already in the request needs a different investigation: the model may not use supplied lore reliably.1

Keep the original book intact. A small disposable test is more useful than resetting settings or rewriting a whole fictional world.

Inspect the request before judging the reply

The short version: Look for the exact lore sentence in the assembled prompt, not just a convincing answer.

SillyTavern’s official prompting guide lists Prompt Itemization on an AI reply, the Prompt Inspector extension, and terminal or browser-console logs as ways to examine prompts.3

Start with Prompt Itemization for the reply you are investigating. If you cannot see the actual text there, use the documented Prompt Inspector extension. Its wand-menu toggle opens a prompt-data popup when you send a message. Inspect without editing.4

Search for a distinctive sentence from Content. Finding the keyword in your own message proves nothing about lore insertion. Finding the entry’s fact only in an older assistant answer also does not prove a fresh insertion.

Two limits matter: Prompt Inspector shows Chat Completion data before backend-specific post-processing, and its Cancel button discards edits but does not cancel the request.4 Redact private character text before sharing any prompt or logs.

Follow the symptom to the next check

The short version: Change the setting that explains your observation, then inspect another request.

These checks follow the documented World Info activation controls; the last row separates successful insertion from model behavior.1

Confirm assignment and literal matching

The short version: Opening a book in the editor is not proof that its entries are eligible for this conversation.

Check the global World Info selection or a context-specific assignment. Character lore follows that character across chats; chat lore belongs only to its assigned conversation. Persona lore follows the selected persona. Verify the active context rather than repeatedly importing the same file.1

Inside the entry, confirm its enable toggle and activation strategy. Constant entries do not need keywords; keyword entries do. The title or memo neither triggers the entry nor supplies the model with its facts. Write a standalone sentence in Content.1

An Optional Filter adds conditions to the primary-key match. AND ANY needs a primary match plus at least one secondary match; AND ALL needs every secondary key. NOT ANY requires none of those secondary keys, while NOT ALL blocks activation when all occur.1 Clear the filter in your test entry rather than guessing which word is missing.

Check case-sensitive and whole-word options, including per-entry overrides. With whole-word matching, a single-word key must occur as a whole word. Scan Depth counts recent messages, not exchanges or insertion position: one scans the last message. A keyword outside that range will not match through the normal chat scan.1

If the character itself never imported, use the separate character-card PNG diagnostic. That is an earlier failure than a working character with missing lore.

Run a controlled Amberbridge example

The short version: Compare an entry-disabled baseline with the same request after enabling one small entry.

This is synthetic TEST DATA and a suggested procedure, not a test executed for this article.

Record your global book selections and activation settings before changing them. Keep production books untouched. For the test, temporarily deselect other global books and use a disposable character and persona without other assigned lore.

Create a temporary book with one entry:

TEST DATA - Amberbridge clock
Primary key: Amberbridge
Content: The station clock in Amberbridge is painted mint green.
Strategy: keyword
Optional Filter: empty
Trigger %: 100
Insertion Position: After Char Defs

Leave inclusion groups and character filters empty. Leave generation-type restrictions unset. Set Sticky, Cooldown and Delay to zero, and turn off Delay until recursion. Use Scan Depth one, with no conflicting entry override, Min Activations zero, and Recursive Scan off. Use ordinary keyword matching without Vector Storage matching. Ensure the World Info budget can hold this short sentence. These choices remove documented alternative activation paths and filters.1

For Chat Completion, also check that the corresponding World Info (after) prompt is on in Prompt Manager.2

1. Disable the test entry. Start a fresh disposable chat, assign the temporary book to that chat, and send: What color is the station clock in Amberbridge?
2. Save the inspected prompt privately. Expected baseline: the Content sentence is absent from the World Info insertion. The answer itself is not the pass condition.
3. Enable the entry. Start another fresh chat with the same character and identical book assignment. Send the identical question and inspect again.
4. Expected enabled result: the exact Content sentence appears in the World Info insertion. If it does not, revisit assignment and activation before changing models.

The fresh chats avoid carrying the baseline answer into the second request. Between trials, change only the entry’s enabled state. After recording the result, detach and remove the temporary book, then restore the global selections and settings you recorded.

Check budget competition and insertion position

The short version: A matching key can still lose space or be sent somewhere you did not expect.

World Info has its own Context % or Budget allowance. When that allowance is exhausted, more entries can be excluded despite matching keys. Constant entries receive priority, followed by larger Insertion Order values; direct keyword matches have priority over recursively activated entries.1

Shorten a copied problem entry first. If the short version appears under otherwise identical conditions, inspect token pressure before enlarging the allowance. More lore also consumes context needed by the conversation.

Inclusion groups are another source of competition: only one simultaneously triggered member of a group is inserted. Prioritize Inclusion selects the highest Order among eligible members with that option enabled. An Order change does not repair a missing keyword match.1

Then check placement. An Outlet entry is not inserted automatically; it needs its named outlet mechanism. Author’s Note placements are ignored when that note is disabled. In Chat Completion, World Info before/after prompts can themselves be toggled off in Prompt Manager.1

Treat present-but-ignored lore separately

The short version: Once the exact content is present, investigate contradictory context rather than repeatedly repairing activation.

Check whether the character description or recent history states a different fact. Try one direct factual question in a clean test chat before testing whether the model uses the detail naturally in a scene. World Info’s documentation explicitly warns that inclusion does not guarantee the generated reply will use it.1

For ongoing story contradictions, keep a reviewed continuity ledger rather than treating every inconsistency as failed lore injection.

These are SillyTavern instructions. Local Waifu has different lorebook controls; consult its backstory guide rather than assuming every Tavern option or export transfers. The SillyTavern and Local Waifu comparison covers that broader choice without making a settings problem a reason to switch.

Sources

The short version: Official documentation reviewed on 2026-10-22; this is a documentation-based diagnostic, not a SillyTavern or Windows runtime test.

1 https://docs.sillytavern.app/usage/core-concepts/worldinfo - SillyTavern World Info

2 https://docs.sillytavern.app/usage/prompts/prompt-manager - SillyTavern Prompt Manager

3 https://docs.sillytavern.app/usage/prompts - SillyTavern Prompting and prompt inspection

4 https://github.com/SillyTavern/Extension-PromptInspector - SillyTavern Prompt Inspector README

Questions people ask

Back to the blog