World Books

A World Book is a reusable collection of lore entries about a setting: its locations, factions, important NPCs, items, history, customs, and rules. You attach a book to a character (or a story or novel), and from then on the AI can draw on that lore during chat — without you having to re-explain the world every time.

Unlike memory, which is about you and this relationship, a World Book is about the world. The same book can power many characters at once.

How entries reach the character

Every entry is one of two kinds, and that single choice — Core canon vs. When relevant — is the only knob you set.

Core canon

Canon entries ground the scene on every turn. Use them for the handful of foundational facts that should always be true — the premise, the era, the core rules of the world. Keep this set small: canon takes up room in every single reply, so a few tight entries beat a pile of long ones.

The editor shows a canon token meter as you mark entries. If you load it up, you'll see a warning — on smaller models, canon near the bottom of the list can get squeezed out and won't be injected. Lean canon, everything else "When relevant."

When relevant

These are the bulk of most books. Reverie retrieves them by semantic similarity — it reads the recent conversation and pulls the entries whose meaning fits the moment, not entries that happen to contain a magic keyword. This is cross-lingual: lore you wrote in one language surfaces for a conversation in another, because matching is on meaning.

A few practical consequences:

• Small books inject in full. If a book's when-relevant entries are small enough to fit the lore budget, all of them ride along every turn — no retrieval needed.
• Large books surface the best matches. When a book is too big to fit, Reverie injects the most relevant entries in full, lists other established topics by name as a lightweight "world index," and lets the model fetch any of them on demand if the conversation drifts toward them.
• Naming an entry outright pulls it in. If you (or the character) mention an entry's title or one of its aliases verbatim, that entry surfaces even if it's short and wouldn't have ranked highly on its own — a safety net for proper nouns, item names, and codes.

Attaching a World Book

World Books attach to the thing that owns the world, not to an individual chat:

• To a character — when creating or editing a character, open Advanced settings → World books and attach one or more books. Every chat with that character inherits them.
• To a story or novel — attach books directly in the story/novel editor. At generation time these merge with the books inherited from the cast.

You can attach your own books or any public book from the library (browse from the same panel). Attaching the same book twice — once to the story and once to a character in it — still injects it only once.

In a group chat, each character's lore is attributed to that character, so one character's private canon isn't mistaken for another's.

Creating and editing entries

You create a book, then fill it with entries. Each entry has a title (the topic's name — a short noun phrase, not a sentence), the content (the lore itself), an optional list of aliases (other names the topic goes by, which improve recall), and the Core canon / When relevant toggle.

Ways to build a book faster:

• AI generate — describe a world in a sentence or two and get a complete draft book (metadata plus a set of entries, with a few marked canon) to review and edit before saving.
• AI split — paste a wall of setting notes and have it broken into self-contained, deduplicated entries that match your existing canon's style.
• Import — bring in an existing lorebook (e.g. a SillyTavern export); the client maps it to entries and imports the batch in one shot.
• Preview — type a sample message and see exactly which entries would surface for it, including a match score for the relevant ones. The preview runs the same selection the live chat does.

Keep entries focused on one topic each. A single sprawling entry retrieves less precisely than several tight ones — the editor nudges you to split very long entries for this reason.

Only the owner can edit a book's entries. Others can attach a public book and read it, but not change it.

Sharing

A book is Private by default. Set it to Public to list it in the library for other creators to attach, or Unlisted to share by link without listing it. Public books are translated for display so other-language users can browse them, but chat retrieval always works off your original text — the cross-lingual matching makes translation unnecessary for the AI.

Living World: let a book evolve with the story

Normally a World Book only changes when you edit it. Turn on Living world (book editor → Settings, private books only) and the AI can add to it too: while you chat with a character carrying that book, the model can record a lasting change it establishes in the roleplay — a city falls, a new faction appears, an NPC's role changes — as a real entry, the same way you'd write one by hand.

• New lore becomes a new entry. Something durable that has no entry yet — a place, faction, NPC, artifact, law, or major event — is recorded as a fresh "When relevant" entry. It never gets marked canon on its own; that stays your call.
• Developments append to existing entries. If something that already has an entry changes, the update is added after what's there — history is never rewritten or deleted.
• You review everything. Right below the toggle, the AI updates list shows every AI-written change with a one-click Undo. Regenerating a reply that wrote new lore rolls that lore back automatically, so a rejected reply never leaves its world changes behind.

Living world is only available on Private books you own, and turning a book Public or Unlisted turns it off — a book that other people can attach never accepts unreviewed AI writes.

World Books vs. Memory vs. Scenarios

These three layers are easy to confuse. The quick rule: World Book = the world, Memory = you, Scenario = this chat's starting situation.

Don't put plot or relationship facts in a World Book — those belong in Memory. Don't put a chat's starting setup in a World Book either — that's a Scenario. The World Book is for things that are true about the world regardless of who's chatting or what's happened between you.

Tips

• Mark only 1–3 entries as canon. The premise and the world's defining rules belong in context every turn; everything else should be "When relevant" and retrieved on demand. A book with no canon adds detail but won't set the scene — make a key setting entry canon if you want the world to ground every reply.
• Write entries the way someone would ask about them. Retrieval is by meaning, so an entry that reads like a clear answer to "who rules the north?" surfaces better than a bare list of names.
• Add aliases for short, name-like topics. Proper nouns, item names, and codes are exactly where semantic matching is weakest and a verbatim alias hit covers the gap.
• Use Preview before you rely on a book. Type the kind of message you'd actually send and confirm the right entries light up.
• Keep canon lean. Lore is a support layer, not the conversation — if canon starts crowding the reply, move entries to "When relevant."