JanitorAI Lorebook Guide (2026): Create, Import and Fix World Memory

A lorebook is conditional world knowledge, not unlimited long-term memory. Good entries inject one relevant fact when a precise key appears; bad entries fire constantly or never fire.

Start with five short entries, give each two or three distinctive keys, attach the book or script to one test character, trigger each key explicitly, and verify the current JanitorAI editor before assuming another platform’s JSON will import.

What a JanitorAI lorebook is — and is not

A lorebook is a set of conditional world facts: locations, factions, relationships, rules, objects and past events. A key in the conversation activates a matching entry so the model sees that fact when it matters.

It is not a perfect diary of everything that happened. Chat history, character definition, persona, scenario, summaries and lore entries occupy different layers. A lorebook is best for authored canon, not automatically learned memories.

Choose lorebook facts that deserve conditional context

Use entries for details that matter sometimes but not in every reply: a city law, the history of a scar, a family relationship or the effect of an artifact. Put permanent behavioral rules in the character definition and the immediate starting situation in Scenario.

If a fact must shape every response, hiding it behind a keyword is risky. If a fact only matters once in fifty turns, keeping it permanently in the character card wastes context.

A clean entry template

Give each entry a descriptive title, two or three precise keys, and a compact body written as facts. Example keys might be “Veyra Gate, western gate”; the body can state who controls it, what is forbidden and what the character personally knows.

Avoid dialogue instructions inside pure world entries. Separate facts from performance so a later change to tone does not require rewriting the world. Keep one subject per entry to make failures easy to diagnose.

Design keywords that activate reliably

Prefer proper nouns, aliases and distinctive phrases. Add the spelling a user will naturally type, but avoid giant synonym lists. Broad keys such as “city”, “friend”, “the” or a character’s pronoun can inject lore almost every turn and crowd out the current scene.

Test exact case, word-boundary and scan settings in the editor you actually have. JanitorAI’s scripting and lore interfaces have changed over time, so a community screenshot may not match the current labels.

Create and attach the smallest useful book

Build five entries rather than fifty: one place, one relationship, one rule, one object and one past event. Attach the lorebook or script to a single private test character, save, then reopen the editor to confirm the association persisted.

Use the official JanitorAI help center and current editor as the authority for today’s buttons. Publicly accessible official help does not provide a stable cross-version JSON contract for every lorebook workflow, so interface verification is part of the process.

Import without assuming JSON formats are portable

“JSON” describes the container, not the schema. SillyTavern World Info, Character Cards V2 extension data and JanitorAI scripts can use different field names, activation rules and nesting. Back up both sides, inspect the structure and import one entry before a full book.

If the current JanitorAI editor offers import, use its own export or documented example as the template. Never rename a file or paste untrusted script merely to make an import button accept it.

Test whether an entry really fires

Start a clean chat and type one exact key in a natural sentence. Ask a factual follow-up whose answer exists only in that entry. Repeat with a near miss that should not activate. Then test two entries together to see whether the prompt becomes confused.

Do not judge by one eloquent answer; a model can guess. Use a unique invented fact, such as a private gate password, so a correct response is strong evidence that the entry entered context.

Fix a lorebook that does not work

Check in this order: the book is attached to the intended character; the entry is enabled; the key matches; activation probability and conditions allow it; scan depth can see the key; token budget has room; the model receives the resulting prompt. Change one setting and retest.

If lore fires constantly, narrow the key and remove recursive chains. If it fires but the model ignores it, shorten the entry, remove contradictions with the card and put the critical fact first. If several entries conflict, merge or rank the canonical version.

Move world memory to LumiChat carefully

LumiChat’s creator accepts a separate World Book JSON, but that does not promise direct compatibility with every JanitorAI script. Use a clean export, compare entry fields, remove executable or platform-specific code, then test a private draft.

For a portable library, keep a neutral source document with title, keys, content and notes, then generate platform-specific JSON copies. That makes migrations auditable and prevents one editor’s hidden defaults from becoming your only copy.

Five-minute quality checklist

• Back up the original file.
• Verify the current account and plan screen.
• Test one narrow example before scaling.
• Check privacy before uploading personal material.
• Record the result so later changes are easy to spot.

Official sources and verification notes

These links are primary documentation or the official product help center. Limits and editor labels can change, so verify the current screen before paying or migrating data.

• JanitorAI official website
• JanitorAI official help center
• JanitorAI official app FAQ
• SillyTavern official World Info reference for portable concepts

Continue with a practical test

Use a small, non-sensitive test character first. Compare the imported fields and one controlled conversation before moving a full library.

Open LumiChat Character Creator · Browse LumiChat characters · Read the privacy policy

Frequently asked questions

Is a lorebook the same as long-term memory?

No. It is authored conditional context triggered by keys; it does not automatically remember every chat event.

How many entries should I start with?

Five focused entries are enough to validate the workflow before scaling.

Why does my lorebook never trigger?

Check attachment, enabled state, exact key, conditions, probability, scan depth and token budget in that order.

Why does lore appear in every reply?

The keyword is probably too broad or recursive activation is pulling related entries.

Can I import SillyTavern JSON directly into JanitorAI?

Do not assume so. JSON schemas and activation rules differ; use the current JanitorAI importer and test one entry.

Can LumiChat import a JanitorAI lorebook?

LumiChat accepts World Book JSON, but JanitorAI scripts may need conversion. Remove platform-specific code and test privately.