How World Info Actually Fires

This tutorial covers how World Info works in current SillyTavern builds, how to structure entries, and how to test them before they poison a live roleplay.

Every entry has keys (trigger words), content (the injected text), and a set of insertion rules. On each generation, SillyTavern scans recent messages for key matches and assembles the matching entries into a context block.

Key mechanics to understand:

• Scan depth — how many recent messages are checked for keys. Higher depth catches older references but costs more tokens.
• Token budget — the maximum percentage of context World Info may consume. Entries beyond budget are dropped, usually by priority order.
• Insertion position — before character definitions, after them, or at a specific depth in the chat. Position affects how strongly the model weights the lore.
• Recursion — entries can trigger other entries. Powerful, but a common source of runaway context bloat.

If you have ever seen a character “forget” a location after ten turns, the cause is usually a key that stopped appearing in the scan window, not a model failure.

Structuring Entries That Survive Long Chats

Treat each entry as a self-contained fact, not a paragraph of prose.

One fact per entry

Split “Lyra Vane is the Archivist of the Ninth Library, exiled from the Concordat in 1187, and she distrusts the Order of Ash” into three entries. Each gets its own keys and fires independently. Combined entries either fire too often or miss when only part of the fact is relevant.

Write keys that match natural phrasing

• Use lowercase keys; matching is case-insensitive in most configurations.
• Include aliases and nicknames — lyra, lyra vane, the archivist, archivist vane.
• Avoid single common words like the or city; they will fire constantly.
• Add a secondary key requirement for ambiguous names. If two characters share a first name, require both tokens.

Keep content declarative

Write facts, not instructions. “The Ninth Library forbids open flame” works better than “Remember that the library has a fire rule.” The model treats injected lore as world state, and imperative phrasing in lore entries tends to leak into narration style.

Use selective and constant flags deliberately

Flag	When to use
Constant (blue)	Core setting facts that must always be present, e.g. genre, era, magic rules
Selective (green)	Most lore — fires only on key match
Vectorized	Semantic matching when exact keywords are unreliable; requires an embedding source

Reserve constant entries for a handful of foundational facts. Every constant entry taxes every single generation.

A Practical Build Order

1. Define the spine first. Era, technology level, magic or power rules, and the central conflict. These become constant entries.
2. Add factions and locations. One entry each, keyed on the faction or place name plus common shorthand.
3. Add named characters. Separate entries for identity, relationships, and secrets. Secrets get narrow keys so they only surface when probed.
4. Add objects and terminology. Invented words need entries, or the model will improvise definitions mid-scene.
5. Link entries with recursion sparingly. A location entry can reference a faction entry, but keep chains to one level deep.

For character cards, decide early what belongs in the card versus the World Book. Personality, speech style, and appearance belong in the card. Backstory, relationships, and setting details belong in the book — that way the same lore can be shared across multiple cards in a series.

Testing Before You Commit

Use SillyTavern’s World Info panel to inspect which entries fired on the last generation. Then:

• Run a short probe scene that mentions each major entity once and verify the correct entries activate.
• Check the token count of the assembled block. If lore dominates the context, trim constant entries first.
• Test a scene with no proper nouns. Only constant entries should fire.
• Temporarily set scan depth low to confirm entries still trigger on immediate references.

If an entry never fires, the keys are wrong. If it fires constantly, the keys are too generic.

Sharing and Reusing Lore

World Books export as JSON and can be attached to any character card or chat. This makes them the natural unit for sharing settings — a single book can serve an entire campaign, a franchise of cards, or a collaborative universe.

MiniTavern supports this workflow directly: the Character Card Market hosts cards and lore packs you can import, and the Chrome extension plus the Web Tavern let you run SillyTavern-compatible chats without a desktop install. The iOS and Android apps carry the same World Info engine on mobile.

Conclusion

World Info is a retrieval system, not a prompt dump. Write one fact per entry, key it on the words your roleplay actually uses, keep constant entries minimal, and verify activation before a long session. Done well, the lore disappears into the background — which is exactly the point. Build your first book, test it against a probe scene, and export it so the next card can reuse it.