What SillyTavern “world lorebooks” actually do
📌 How to read this: Mechanisms only—no “click this menu” walkthrough. UI depends on your SillyTavern build; field semantics belong to World Info (official).
Settings in plain English (start here)
Before the tables, think of each lore entry as a sticky note the app may paste into the big prompt the model reads:
| Label you’ll see | Plain meaning |
|---|---|
| Keywords | If these words show up in scanned chat (and rules pass), this note may attach. |
| Content | The actual text the model gains when the note attaches. |
| Scan depth | Don’t only read your latest line—also look N message pairs back for keywords. |
| Token budget / % | All lore notes together have a length cap; overflow entries may be skipped. |
| Insertion order | When several notes attach in the same slot, sort key; larger numbers usually sit lower in that bundle and often weigh more on the next reply. |
| Insertion position | Where in the prompt stack the note sits—near character text, near Author’s Note, at message depth @D, etc. |
| Constant (blue) | Try to include every turn without keywords (still respects budget). |
| Normal (green) | Needs keyword hits. |
| Recursion | Text from note A can mention a keyword for note B, chaining activations. |
Wording in your build may differ slightly; this is the idea layer.
📖 What a lorebook is
| Metaphor | Meaning |
|---|---|
| 📕 Dictionary / encyclopedia | A word appears in chat → a note can be merged into the big text the model reads |
| 🎯 Minimal pattern | Keywords match → Content may activate |
| ⚠️ Limit | Activation ≠ the model must obey verbatim; it only raises the chance the fact is used |
Logic example: A “used-book shop owner” card and the line “tonight I’m picking up a batch of rare catalogs on Lantern Row.” The model may not know what Lantern Row is; a lore entry keyed on that name explains the street, mood, and how it ties to the shop. When the name appears, the entry can enter context so the scene stays coherent.
🧠 Two scopes (conceptual)
| Type | Scope (idea) | Typical use |
|---|---|---|
| 🌐 Global | Shared across many characters | Shared setting tables, props, etc. |
| 👤 Bound to a character | Mainly for one card | Private backstory, character-only facts |
Whether exports bundle the primary bound book depends on format/version—no UI steps here.
🧩 Context and ordering: why placement matters
Treat one model request as a capacity-limited container:
| Idea | Intuition |
|---|---|
| 📦 Context | Total tokens the model can see; old material can fall out |
| 🔒 Stickier blocks | e.g. character description, some WI set to constant—often described in tutorials as harder to push out |
| 📜 More roll-off | Most chat history, some example chats—long runs push earlier lines away |
Community tutorials about SillyTavern’s “blue” prompt strip (chat-facing layer) often summarize:
⬇️ Within that strip, lower in the stack tends to weigh more on the current reply.
So “as low as possible” is not automatically best—you balance character text, Author’s Note relatives, and WI.
Author’s Note (concept): pins helper text above the N‑th chat message; WI entries placed “before/after Author’s Note” move with that anchor.
🎛️ Entry-level insertion position (relative strength table)
From the same tutorial lineage—conceptual strength, not a click map:
| Placement (label) | Relative strength (tutorial) | Note |
|---|---|---|
| ⬆️ Before character definition | Weakest | More “background” |
| ⬇️ After character definition | Slightly stronger, still weak | — |
| 📝 Before Author’s Note | Strong | Tracks AN depth |
| 📝 After Author’s Note | Very strong | Same; easy to over-pack |
| @D (depth slot) | Variable | Inserts above the N‑th message (exact rules per build) |
💡 Short chats (new threads): character text, WI around it, and the first message jointly bootstrap the run—tutorials warn against moving everything to “right around Author’s Note” on day one.
🔢 Insertion order when several entries fire
| Rule (aligned with official docs) | Meaning |
|---|---|
| ⬆️ Smaller number | Higher in the WI chunk for that same placement |
| ⬇️ Larger number | Lower in the chunk—usually stronger on the next token |
If placements differ, sort by placement tier first, then Order inside each tier.
Author-style tips:
| Tip | Meaning |
|---|---|
| 🎯 Importance → lower in stack | Within a tier, raise Order for what must dominate |
| 🌍 Big world bible | Can live higher (e.g. before character block) |
| 📟 Status HUD / rules | Often lower; if the model “forgets” state, try even lower |
Shorter intro: SillyTavern lorebook order for beginners.
🎲 Trigger probability
| Value | Behavior (concept) |
|---|---|
| 100 | Always attempt insert when otherwise eligible |
| 50 | Coin flip each eligible turn |
| 0 | Effectively off |
Tutorials note the UX can feel like “rare events happen often”—verify empirically.
🚦 Status modes
| Icon | Mode | Behavior (concept) |
|---|---|---|
| 🟢 | Normal | Keyword-driven |
| 🔵 | Constant | Tries every turn without keywords (tutorial analogy: “permanent prompt”) |
| ❌ | Disabled | Inactive |
🔵 Constant + depth: can mimic Author’s Note–like anchoring (still budget-limited).
🔌 Logic (AND / OR / NOT intuition)
| Logic | Intuition |
|---|---|
| AND | All required keys must appear |
| Comma lists | Often OR among primary keys (product-dependent) |
| NOT | Fire when A matches and B does not |
Secondary boxes are often empty in simple setups.
👥 Character filter / exclusion (concept)
| Mode | Story (tutorial) |
|---|---|
| ✅ Only these characters | Group chat: secret one actor knows |
| 🚫 Exclude characters | Group chat: everyone except one knows |
⚙️ Global activation settings
Merging global vs character lore
| Strategy (label) | Tutorial note |
|---|---|
| Even sort | Author finds behavior messy—don’t rely on it blindly |
| Character-first / global-first | Descriptions vs reality can drift |
Key claim (tutorial): Insertion position + Order beat these merge modes. Example spirit: Order 100 still ends below Order 99 in the same band—not flipped by “global vs character” alone.
Scan depth
| Idea | Meaning |
|---|---|
| Pair user + assistant messages | Depth 1 ≈ scan one recent pair; 2 ≈ two pairs… |
| Typical | 1–2; adventure cards may need more |
| Ceiling | Community cites ~ten pairs—verify your fork |
Recursive scanning
| Off | On |
|---|---|
| Text inside entry A cannot wake entry B by mentioning B’s keys | Chain: A’s injected text can trigger B |
Classic logic: Bessie → text mentions Rufus → Rufus entry loads.
Budget & eviction priority
| Control | Role |
|---|---|
| Context % / budget cap | Hard cap on WI tokens |
| Overflow alert | Warn when WI exceeds allowance |
When over budget (tutorial ordering):
| Priority (high → low) | Rule |
|---|---|
| 1 | 🔵 Constant beats 🟢 normal |
| 2 | Higher Order beats lower |
| 3 | Direct keyword hits in chat beat recursion-only activations |
🔤 Other toggles (logic)
| Toggle | Role |
|---|---|
| Case-sensitive keys | Rose vs rose |
| Whole-word | Block hotdog firing dog; bad for CJK without spaces—often leave off |
| Min activations | Alternate scan mode; tutorial: easy to misuse, rarely needed |
🧪 Regex (purpose only)
Good for staged mechanics (affection tiers, state machines)—flexible triggers; details live in dedicated regex guides.
🌟 Closing frame
| Lens | One line |
|---|---|
| 🔧 Engineering | Lorebooks are ordered, budgeted, chainable prompt inserts |
| 🎮 Play | Rulesets, mode instructions, not just “world bible” |
Some models also show extra attention to the first/last tiny slices of context (community stress tests)—treat as advanced, validate with logs.
Browse PNG + embedded JSON on Mac with Sillycard.
Topic: Blog · Latest posts