What Is an AI Lorebook? How It Works, What It Costs, and When a Persona Beats It
18+ Only
An AI lorebook is a file of short entries, each tagged with keywords, that gets pulled into the model's prompt when one of those keywords shows up in the chat. That is the whole mechanism. It is not a memory, it does not learn, and it has no idea what you said five minutes ago. It is closer to a dictionary that opens itself.
That confusion is worth clearing early, because "lorebook" now covers three different things. This page covers what the feature does, which platforms document one as of October 2026, the three ways they fail, and the case for skipping them.
What an AI lorebook actually is
SillyTavern's documentation is the clearest description anyone has published: World Info, also known as Lorebooks or Memory Books, is a dynamic dictionary that only inserts relevant information when keywords associated with an entry appear in the message text. Every platform here implements that same idea and gives it a different name.
A keyword-triggered dictionary, not a memory
A lorebook entry is written by you, in advance, in a fixed form. Nothing about the conversation is stored in it. If your character's home town is Bellhaven and you write an entry with that keyword, the next time Bellhaven is mentioned the entry drops into the prompt and the model can describe it. If the conversation moves elsewhere, the entry stays dormant and costs nothing.
Memory does the opposite: it records what happened and feeds it back later. On OurDream the character keeps an advanced memory of past chats, which is how a detail you set three sessions ago can survive. Different mechanism, different failure mode — memory can be wrong about you, a lorebook can only be incomplete about the world.
Why the trigger model exists
Context is finite and expensive. Cramming a setting bible into a character definition means paying for it on every message. Character.AI says so directly: details in the character definition take up space in every response, relevant or not. The lorebook moves that cost from "always" to "only when needed".
What an entry contains
Three things: a name, keywords, and a content block. Only the content block reaches the prompt — SillyTavern warns that keywords and titles are not inserted, so each entry must read standalone. That explains most unsatisfying lorebooks: people write fragments that only make sense next to a title the model never sees.
| What people expect | What a lorebook does |
|---|---|
| The character remembers our chats | No. It injects facts you wrote by hand when a keyword appears. |
| I can dump a whole wiki in and it stays consistent | Partly. Only entries whose keywords fire reach the prompt, and each one consumes context. |
| It learns new lore as the story develops | No. You write it, or it does not exist. AI-assisted generation still needs your button press. |
| One lorebook works for every character | Yes, and that is its real strength. Character.AI documents linking one to several characters. |
Which platforms actually ship one (checked October 2026)
Feature lists lie by omission, so this section states only what each platform documents.
Character.AI: Lorebooks, in beta, paid tier
The help centre describes Lorebooks as a reusable collection of world entries — side characters, locations, rules — that activate when their keywords come up. The feature is in beta, and early access to creating and editing requires a c.ai+ subscription. Entries autosave, and one file can be linked to several characters.
NovelAI: the original, with import and export
NovelAI treats the Lorebook as a repository for supplemental information added to context as each entry comes up, covering characters, locations, objects and factions. It supports import and export of the whole file — which is why lorebook files circulate between platforms at all.
SillyTavern: World Info, the configurable one
This is where the mechanic is exposed in full. Entries can be bound to a character, a persona or a single chat, so one file manages both "lore this character always knows" and "lore that applies only in this conversation". It also documents recursive activation, scan depth, maximum depth and a context budget percentage. If a lorebook behaves differently in two apps, this is why: the trigger engine is not standardised.
SpicyChat, Janitor AI, Chub: lorebooks as a library
SpicyChat documents a full system with private and public visibility and a ceiling of 5,000 entries per lorebook. Janitor AI and Chub host large public libraries, which is why "ai lorebooks download" is a query at all. The trade-off is obvious: a public lorebook written for someone else's canon is rarely the one your character needs.
| Platform | What it is called | Where it lives | Trigger control |
|---|---|---|---|
| Character.AI | Lorebooks | Profile tab, linkable to several characters | Keywords only |
| NovelAI | Lorebook | Per story, import and export | Keywords, generation assistance |
| SillyTavern | World Info | Global, character, persona or single chat | Scan depth, recursion, budget |
| SpicyChat | Lorebook | Attached in the character editor | Keywords, public or private |
| OurDream.ai | No separate file | Persona, backstory, scenario, memory | Automatic memory, no manual index |
What a lorebook does well, and the three ways it fails
The strength: one world across many characters
This is the use case that justifies the feature, and single-character roleplay rarely needs it. A dozen recurring characters in one city: a lorebook is the only sane way to keep that city consistent — same streets, same rules, whichever character is on screen.
Failure 1: the keyword you did not write
The entry about the Crimson Order is tagged "Crimson Order". In the first scene the faction is named in full, the entry fires, everyone learns the lore. From then on your character says the order, and the entry never fires again. The lore sits in the file and not in the conversation. The fix is unglamorous: tag the variants a human would type, singular and plural, with and without the article, plus the character's nicknames. Then check your last twenty messages for those words.
Failure 2: silent token burn
Every activated entry occupies prompt space. Write a 400-word entry, tag it with the character's own name, and you added 400 words of overhead to every reply forever for nothing. Hence SillyTavern's context budget percentage: lore competes with the conversation.
Failure 3: maintenance debt
Lorebooks do not update themselves. A character alive for months accumulates contradictions — the sister's name changes, a detail invented in session nine contradicts an entry from session two. Nothing warns you: the model receives both versions, picks one, and now you have a continuity error you cannot trace. Serious keepers build a pruning pass; most people do not, and the file quietly rots.
How automatic memory differs from a lorebook
The two are complementary, but they solve different problems, and picking the wrong one is the main source of disappointment here.
You write the index, or the model does
With a lorebook you are the indexer: you decide what the world contains, which words summon it, and what it says. Total control, total responsibility. With automatic memory the model decides what mattered. You lose control and gain the ability to be surprised by your character remembering something you forgot you said.
| Task | Lorebook | Automatic memory |
|---|---|---|
| Keeping a fictional city consistent across characters | Yes, the only reliable way | No, memory is per character |
| Remembering a detail you mentioned once, casually | No, unless you wrote it | Yes, this is the whole point |
| Controlling exactly what the character believes | Yes, verbatim | No, summarised and lossy |
| Working in five minutes without setup | No, expect a draft and two revisions | Yes, it accumulates by itself |
The hybrid most experienced users end up with
Run both and keep the split strict: lorebook for the world, memory for the relationship. Facts that must never drift go in the world file. Anything about what the two of you have been through belongs to memory. Mixing them produces characters that recite their own backstory unprompted.
What OurDream does instead, and where that loses
OurDream offers no attachable world file, and it is not worth pretending otherwise. Its character creator documents four fields that carry the same job for one character: persona, backstory, scenario and an advanced memory of past chats, plus appearance and voice.
Why persona and memory rather than entries
The builder's documented advice is that a specific persona and a clear scenario do most of the work, because a vague character produces vague replies. A lorebook keeper reaches the same insight after twenty entries: the character's voice filters the world more reliably than a list of facts.
The honest gap
If your project is a shared universe with six recurring characters and a map, OurDream is the wrong tool and a lorebook platform is the right one. Where the persona route wins is the single-companion case: no index to maintain, no keyword that fails to fire, no pruning pass after three months. The full trade-off is in the OurDream vs Character AI comparison.
How to write a lorebook that fires
1. List entities before you write a single entry
Names only: people, places, objects, factions, rules. Writing entries first means discovering halfway through that three of them overlap and disagree.
2. Write the always-on core first
The setting, the rules governing it, the central tension. Four bullet points, no more. It can live in the scenario field or in one constantly-triggered entry — either way it must not be long.
3. One entity per entry, three to five bullets
A 500-word entry about one city dominates every message it activates on, and you cannot tell which line changed the reply. If an entry wants to be long, it is probably two entities.
4. Tag the words a human would type
Not the concept name. The words: short form, plural, nickname, the misnomer your character would use. Highest-leverage step, most often skipped.
5. Link entries by name, and know what your app does with that
Referencing one entry inside another builds depth: the Bellhaven entry mentions the Crimson Order and, if recursion is on, the order's entry can follow. SillyTavern caps that recursion with a maximum depth — your guard against a chain reaction. On platforms without recursion the link is decorative.
6. Prune after ten sessions
Read back which entries actually fired. One that never fired is either unnecessary or badly tagged. This habit separates a lorebook that stays useful for a year from one that becomes a liability by month three.
Who should bother, and who should skip it
Bother if
You run a multi-character story in one world, you want verbatim control over what a character believes, or you already have written lore from outside the app that you want respected.
Skip if
You want one companion who remembers your conversations, you pay per token, or you know you will not prune. To start talking in five minutes, a persona and a scenario get you further than a half-built lorebook ever will.
The cost check before you commit
Lorebook features sit behind paid tiers — Character.AI requires c.ai+ for early access. On OurDream, generation is metered in DreamCoins, documented in OurDream pricing and the cost breakdown.
Try OurDream.ai — persona, scenario and automatic memory, no lorebook to maintain →
Frequently asked questions
What is an AI lorebook in plain terms?
A file of short entries, each tagged with keywords. When one of those keywords appears in your chat, the matching entry is inserted into the model's prompt. It is closer to a dictionary that opens itself than to a memory: nothing is stored about what you said, only what you told it in advance.
Is a lorebook the same as long-term memory?
No, and confusing the two is why people are disappointed. A lorebook holds facts you wrote by hand and knows nothing about the conversation. Memory holds things that happened in the conversation and is generated automatically. A lorebook can tell a character that the city of Ashfall burned forty years ago. It cannot tell the character that you said you were tired at 2am last Tuesday.
Which AI apps actually have a lorebook feature?
As of October 2026, four document one clearly. Character.AI has Lorebooks in beta for c.ai+ subscribers. NovelAI has a native Lorebook with import, export and AI-assisted entry generation. SillyTavern calls it World Info and gives the most control over activation. SpicyChat offers Lorebooks with public and private visibility and a documented ceiling of 5,000 entries. Janitor AI and Chub host community lorebooks around the same format.
Do lorebook entries cost tokens even when they do not fire?
An entry that never activates costs nothing, because it is never sent to the model. The cost appears when entries do fire: each one occupies prompt space, and SillyTavern exposes a context budget percentage precisely to cap how much of the window lore can consume. A lorebook that fires six entries per message is paying rent on every reply.
Why does my lorebook never trigger?
Almost always the keywords. People tag the entry with a concept name instead of the words that will appear in the chat. If the entry is about the Crimson Order and the keyword is only "Crimson Order", it fires the first time the faction is named and never again, because after that everyone says "the order". Tag the variants a human would type, plus the nicknames your character uses.
Does OurDream.ai support lorebooks?
Not as a separate attachable world file. OurDream's character builder documents four fields that carry the same job for a single character: persona, backstory, scenario and an advanced memory of past chats, plus appearance and voice. If you need one world shared across a dozen characters, a lorebook platform is the right tool. If you need one character that stays itself across a long relationship, the persona and memory route covers most of the same ground with less maintenance.
Related guides on this site
- AI roleplay worldbuilding — building the world a lorebook is supposed to index.
- AI fantasy character creator — the character side, before any world file exists.
- AI character chat with memory — the mechanism a lorebook is confused with.
- Best AI companions with memory — how automatic memory compares across apps.
- Custom personality — persona design, which does most of the work a lorebook promises.
- Chat with multiple characters — the case where a shared world file earns its keep.
- AI roleplay for fantasy storytelling — where continuity problems show up first.
- Best AI roleplay platforms 2026 — which of them expose the trigger engine.
- OurDream vs Character AI — the two documented positions on world files, side by side.
- OurDream vs Nomi and OurDream vs Kindroid — how the memory-first alternatives stack up.
- OurDream pricing — what generation and voice actually cost in DreamCoins.
Sources checked October 2026: Character.AI Help Center — Lorebooks; SillyTavern documentation — World Info; NovelAI documentation — Lorebook; SpicyChat documentation — Lorebook; help.ourdream.ai character creation guides.