A lorebook is a collection of short background entries (places, side characters, items, rules of the world) that an AI roleplay app adds to the model’s prompt only when a matching keyword comes up in the chat. It lets a character “know” a large world without stuffing every detail into every reply. SillyTavern calls the same feature World Info; Janitor AI, NovelAI and Character.AI call it a lorebook.
Cherrypop.ai publishes this, and we make a competing AI companion app. Based on each vendor’s pricing and help pages and published user reports, as of September 2026.
What is a lorebook, in plain terms?
Every language model reads a limited amount of text per reply, called its context window. The character definition, your persona, the recent chat and any instructions all compete for that space. Put a whole fantasy kingdom into the definition and it costs tokens on every message, even when the chat is about breakfast.
A lorebook solves this by keeping details on standby. Each entry has two working parts:
- Keys (triggers): words or phrases such as a name, a place or a question like “how old are you”.
- Content: the text the model actually reads when a key shows up.
When you (or the character) mention “the Lantern Market”, the entry for the Lantern Market slips into the prompt for that turn. When nobody mentions it, it costs nothing. Character.AI’s help center sums up the division of labor well: always-relevant details go in the character definition, sometimes-relevant details go in the lorebook.
How a lorebook works: triggers, depth and token budgets
The more advanced systems, SillyTavern and NovelAI especially, expose four controls. Simpler ones like Character.AI’s handle most of them for you.
1. Keywords and matching
Keys are usually case-insensitive (NovelAI’s regex keys are the exception). Advanced tools add extras: SillyTavern and NovelAI both accept regular expressions as keys, SillyTavern has optional secondary filters (AND ANY, AND ALL, NOT ANY, NOT ALL), and NovelAI lets you join keys with “&” so an entry fires only when both words are present. SillyTavern also matches single-word keys as whole words by default, so “art” will not fire on “party”.
2. Scan depth (how far back it looks)
The app only searches the most recent part of the chat for keys. In SillyTavern this is Scan Depth, counted in messages: 1 scans only the last message, 2 the last two, and so on. NovelAI uses Search Range, counted in characters of story text, up to 10,000.
3. Insertion position and order
Where the entry lands in the prompt changes how strongly it steers the reply. SillyTavern’s docs note that entries placed closer to the end of the prompt have more impact on the reply. SillyTavern offers positions before or after the character definitions, around example messages, in the Author’s Note, or “@ D”, a specific depth in the chat where depth 0 is the bottom of the prompt. NovelAI uses numbers: 0 is the top of the context and -1 the very bottom. Both also have an Insertion Order that decides priority when several entries fire at once.
4. Token budget
Lore cannot be allowed to push the chat itself out of context, so there is a cap. SillyTavern sets it as a percentage of the context (Context %) or a fixed token Budget; once it is used up, no more entries activate even if their keys appear. NovelAI sets a Token Budget and Reserved Tokens per entry. This is why short entries matter: a bloated one can crowd out three useful ones.
How lorebooks work on SillyTavern, Janitor, NovelAI and Character.AI
| Platform | Name | Where it lives | Notable controls |
|---|---|---|---|
| SillyTavern | World Info / lorebook | Global, or bound to a character, persona or single chat | Scan depth, budget, insertion position and order, recursion, probability, sticky/cooldown/delay |
| Janitor AI | Lorebook (a script type) | Scripts, under your profile, assigned to your characters | Keys, content, probability, message depth, min messages, order, inclusion groups |
| NovelAI | Lorebook | Inside each story | Search range, token budget, reserved tokens, insertion position, phrase bias, categories |
| Character.AI | Lorebook (beta) | Standalone, linked to one or many of your characters | Up to 8 keywords per entry, entry templates, public or private, entry visibility |
SillyTavern (World Info)
SillyTavern has the deepest system. Entries can be constant (always inserted) or keyword-triggered, and can trigger each other through recursion, so mentioning a city can pull in the entry for the guild that runs it. You can bind one primary World Info file to a character, and it is embedded in the character card when you export it. Timed effects add pacing: Sticky keeps an entry active for N messages, Cooldown blocks it for N messages, Delay waits until the chat has N messages.
Janitor AI lorebook
On Janitor AI, a lorebook is not a field in the character editor. It is one type of “script”, created from the Scripts page in your profile menu and then assigned to your characters. Janitor’s help guides describe entries with a category, an entry name, keys (“wake-up words”, comma-separated), content, a probability and optional group settings; its newer overview adds message depth, a minimum message count and order. Janitor’s own beginner guide advises keeping entries to one to three sentences and starting with one entry, simple keys and 100% probability before building more. If you are tuning a Janitor bot, pair the lorebook with a lean definition; our Janitor AI prompts guide covers what goes in the permanent fields.
NovelAI lorebook
NovelAI’s lorebook is built for long-form story writing. Beyond keys and an Always On switch, its Placement tab exposes search range, token budget, reserved tokens, insertion order and position, key-relative insertion and cascading activation (entries triggered by keys inside other lore). Phrase Bias lets an active entry nudge the odds of specific words. NovelAI labels the Placement tab as intended for experienced users.
Character.AI lorebook
Character.AI, a general-audience app, launched Lorebook in beta on July 24, 2026. Per its help center, early access to creating and editing them requires a c.ai+ subscription. A lorebook is its own object: you add entries (title, content, up to 8 keywords), publish it, and link it to one or several characters you own so they share one canon. Public is the default; private lorebooks are for c.ai+ subscribers. The title is for humans only, the character never sees it. Characters with an active lorebook carry a +Lore tag. For more on the platform, see our Character.AI review.
How to write good lorebook entries
- One subject per entry. A location, a person, a rule. Mixed entries fire at the wrong time.
- Name the subject inside the content. SillyTavern and Character.AI both note that the model does not see the entry title, so an entry that never says “Mara” leaves the model guessing who it describes.
- Keys people actually type. Name, nickname, title, role, plus natural questions (“what do you look like”).
- Short. A few sentences covers most entries and protects the budget.
- Tell the model what to do, not just what exists. Character.AI’s guide puts it well: descriptions say what exists, constraints say what to generate.
- Fix recurring mistakes first. If the model keeps changing a character’s eye color or forgetting how magic works, that correction is your most valuable entry.
Three examples you can adapt, for an original fantasy setting:
| Type | Keys | Content |
|---|---|---|
| Location | Lantern Market, the market, night market | The Lantern Market opens only after dark on the docks of Veyle. Stalls sell charms, stolen maps and spiced wine. Guards look away for a coin. When a scene is set here, describe noise, smoke and haggling, and let {{char}} be recognized by at least one vendor. |
| Side character | Mara, Captain Voss, the captain | Captain Mara Voss, 38, runs the harbor watch. Dry, tired, fair. She owes {{char}} a favor and hates that she does. She never raises her voice; she lowers it. She speaks in short sentences and ends arguments by walking away. |
| World rule | glassmagic, spell, casting, magic | Glassmagic in Veyle needs a shard of enchanted glass held in the hand. Each spell cracks the shard; a fully cracked shard shatters and cuts the caster. {{char}} never casts without a shard and always counts the cracks out loud. |
Common lorebook mistakes
- Keys that are too common. “You”, “the”, “love” or “home” fire constantly and spend your budget on irrelevant lore.
- Keys nobody types. A key of “Veylean Harbor Authority” never fires if everyone says “the harbor”. Add the short forms.
- Wiki-length entries. A 600-word history of the kingdom crowds out the chat and other entries. Split it or cut it.
- Core personality in the lorebook. How the main character talks and behaves belongs in the definition, which is always present, not in an entry that may not fire.
- Too many always-on entries. Constant entries are just a second definition with extra steps. Reserve them for rules that truly apply every turn.
- Contradicting entries. If two entries give Mara different ages, the model picks one, or blends them.
- Expecting a guarantee. An active entry is information, not a command the model must obey. If a detail keeps slipping, make the entry more specific and more instructive.
Getting lorebook-style results on a platform without lorebooks
Many companion apps, Cherrypop.ai included, have no lorebook feature. You can still get most of the benefit, because most characters do not need 60 entries. They need five or six facts that stay consistent. On Cherrypop, the character creator at cherrypop.ai/create gives you three places to put them:
- Scenario: the world and the situation. Setting, the one or two side characters who matter, the key world rule, and how the character knows you.
- Additional Personality Details: the character’s identity: voice, habits, backstory, and corrections for anything the model tends to get wrong.
- First Message: the opening scene. Use it to show the setting and a side character in action, not just describe them.
The trick is compression. Rewrite each would-be entry as one line and keep only what affects scenes. The three entries above become a Scenario like this:
“Veyle is a harbor city. The Lantern Market on the docks opens after dark: charms, stolen maps, bribable guards. Captain Mara Voss (38, dry, owes {{char}} a favor, never raises her voice) runs the harbor watch. Magic needs a shard of enchanted glass; every spell cracks it, a shattered shard cuts the caster. {{user}} has just hired {{char}} to find a missing map.”
Then act as your own trigger. When you want a detail back in play, mention it in your message (“I head for the Lantern Market to find Mara”), which puts the name right in the recent chat, where the model is most likely to pick it up. The AI Assisted option, on by default, can enhance the text you write in these fields. For more on steering a scene, see our guide on how to AI roleplay.
Cherrypop.ai is free to start, with limits, and also has a catalog of ready-made characters if you would rather not build one.
FAQ
What is a lorebook?
Keyword-triggered background entries that an AI roleplay app adds to the prompt when a matching word appears in chat. SillyTavern calls it World Info.
What is a lorebook in Character AI?
A beta feature (launched July 24, 2026) for storing side characters, locations, items and rules as entries with up to 8 keywords each. You link a lorebook to one or more of your characters. Per the help center, early access to creating lorebooks requires c.ai+.
What is a lorebook in Janitor AI?
A script type you create under Scripts in your profile menu, fill with entries (keys plus content), and assign to your characters. Janitor suggests one to three sentences per entry.
Does Cherrypop.ai have lorebooks?
No. Put your world’s essentials in the Scenario, Additional Personality Details and First Message fields, keep them short, and mention a detail in chat when you want it back in the scene.