SillyTavern World Info: set up a lorebook and troubleshoot entries

When a SillyTavern lorebook seems to do nothing, check whether its entry reached the request sent to your model. Rewriting the character or adding more lore will not identify where that process failed.

This guide uses one original entry, two contrasting inputs, and a budget check to separate activation problems from generation problems. EchoMelon publishes this guide and operates a character-chat product. We checked SillyTavern's official documentation on November 3, 2026; we did not install it, connect a backend, or run these tests. The procedures below are for you to try, not reported results.

You need a working SillyTavern setup and a backend you can use. Check its charges before generating. Keep your existing stories and settings intact; use a disposable practice chat and record any settings you temporarily change.

The fields have different jobs

The World Info documentation distinguishes the text supplied to the model from the controls deciding when to supply it:

Field	Job
Title / Memo	Your organizational label
Key	Text used for keyword matching
Content	The actual material to insert
Strategy	Keyword, constant, or vector-based activation
Insertion Position	Where the material belongs in the prompt
Order	Ordering and priority among eligible entries

A descriptive title does not replace Content. Likewise, finding a keyword is not proof that the entry survived the remaining conditions and budget.

Make a small, isolated practice entry

Open the World Info editor and create a practice lorebook with one entry. Bind it to the practice conversation through Chat Lore in Character Management. Check that other global, character, persona, or chat lore will not interfere with this exercise; restore any shared selections afterward.

Use these deliberate test settings, rather than treating them as an optimized preset:

Setting	Practice value
Entry enabled	Yes
Strategy	Green, keyword-based
Probability	100%
Optional Filter / Inclusion Group	Empty
Insertion Position / Order	After Char Defs / 100
Scan Depth / Min Activations	1 / 0
Recursive Scan / World Info vector matching	Off
Sticky / Cooldown / Delay	0 / 0 / 0

Check entry overrides too. Scan Depth 1 examines the last message; other activation paths would make the negative check harder to interpret. A 100% probability removes that random filter, not other restrictions.

Start with enough World Info budget for this entry's measured tokens. There is no universal token count for the text below across all tokenizers and configurations.

Copy the seed-library example

The scene involves two fictional adults: Nessa Venn, a 42-year-old collection keeper, and Tavi Reed, a 29-year-old visitor controlled by the player. Keep this setup separate from the World Info entry:

Nessa manages a community collection. Tavi is visiting.
Portray Nessa and observable surroundings.
Leave Tavi's words, thoughts and decisions to the player.

Use a neutral opening without the lorebook's key or rules:

Nessa sets a closed paper packet on the table.
"What would you like to check before we continue?"

Set the entry's Title / Memo to Seed loan procedure and its sole Key to Cinderwell Seed Library. Paste this into Content:

Cinderwell Seed Library stores rare seeds in numbered paper packets.
Unnumbered packets stay on the red inspection tray, outside the lending drawers.
A loan requires both a lot number and a matching catalog record.
Nessa Venn can check the catalog but cannot invent a missing lot number.
The visitor may request a catalog check or leave the packet for inspection.

These are ordinary text fields, not a JSON or PNG import package. Do not also paste the entry into the character definition, greeting, or user input: then finding its rule in the request would no longer establish that World Info supplied it.

The example gives you a recognizable phrase, a concrete rule, and a choice that remains with the player. It does not require a particular generated sentence.

Run the positive check and inspect the request

From the practice opening, send:

At Cinderwell Seed Library, I find an unnumbered packet.
What must happen before it can be loaned?

Keep the first reply. SillyTavern's prompt-viewing guide identifies Prompt Itemization on an AI reply as an inspection route; local terminal logs and the browser console are alternatives.

Look for the complete entry, particularly Unnumbered packets stay on the red inspection tray. Then inspect the reply separately. Does it preserve the lot-number and catalog requirements? Does it leave Tavi's decision open?

Record two observations: entry present or absent, and reply consistent, contradictory, or inconclusive. A sensible answer about a library could be a guess. An entry present in the request could still be ignored. Neither observation substitutes for the other.

For Chat Completion setups, also check the relevant World Info prompt in Prompt Manager. Its default World Info prompts can be toggled off. Text Completion uses Advanced Formatting instead; do not assume the same controls apply.

Try an input without the key

Start another fresh practice conversation with the same neutral opening and settings. Chat Lore belongs to an individual chat: explicitly bind the same practice lorebook in this new chat and verify the entry is enabled before sending the negative input. Otherwise, a missing binding could look like a successful keyword check. Send:

At the seed library, I find an unnumbered packet.
What must happen before it can be loaned?

Under the isolated keyword setup, this is your negative check: the exact key is absent. Inspect the request again. If the entry still appears, investigate other active sources, entry overrides, or activation paths before attributing it to semantic matching.

Do not run this immediately after the positive exchange in the same history. You would be introducing another possible source for the rules you are trying to trace.

Distinguish a budget problem from a trigger problem

In a disposable copy of the positive check, first verify that Chat Lore still binds the same practice lorebook and the entry is enabled. Then lower the available World Info budget below the entry's measured size, using a valid positive value your interface accepts. Record the original value and restore it afterward. Check whether the entry is omitted; do not invent a result if your configuration behaves differently.

The official rules give constant entries precedence, then larger Order values; direct matches outrank recursive ones. Capacity still limits inclusion. Raising Order does not create more space, and changing insertion position does not prove better compliance.

When diagnosing a real story, use this sequence:

Observation	Next check
Entry absent before any meaningful comparison	Confirm the correct book, enabled entry, and Content
Input contains the key but entry is absent	Check scan scope, filters, probability, groups, timed effects, and budget
Entry fits but is absent from the final request	Inspect insertion position and your prompt-building preset
Entry is present but output contradicts it	Look for conflicting instructions or history; change one factor and retain both outputs

For the last case, repeatedly adding copies of the same lore makes diagnosis harder. First identify the specific contradiction, such as lending the unnumbered packet without checking the catalog.

Keep a useful record

Save the date, SillyTavern version, backend/model, context settings, entry configuration, test input, whether Content appeared, the first reply, and any cost. Keep this local if the request includes private material. The User Settings documentation explains version display and prompt logging.

This exercise checks a small insertion workflow, not long-term memory or an overall success rate. For events that change during play, our recap and continuity guide addresses a different task. If you are still deciding whether to manage a frontend and backend, start with our SillyTavern review.