Entries & activation

The entry list

Each row of the lorebook editor is one entry:

Column
checkbox	Enabled. Disabled entries are never used.
Entry Name	Your name for the entry. Not sent to the model.
Order	Position of the entry in the prompt, see Order and insertion. New entries get `100`.
Tags	Tags that activate the entry, see Tags.
⌄	Shows the details of the entry.
trash	Deletes the entry (after confirming).

Add Entry in the toolbar adds an entry with order 100. Changes are saved as you type.

Searching and deleting many entries

The box with the magnifier next to the buttons shows only the entries that have the text in their name, content, note or filter (the case does not matter); clear it to see all of them again. It looks only at the entries of the selected lorebook, not at the sub lorebooks. The phone layout has the box above the list.

The checkbox column at the left selects entries (the one in the header selects all that are shown, so a search and then the header box selects the matches). When something is selected, Delete selected appears in the toolbar with the number of the entries; it asks once and then moves them to the trash. The phone layout does not delete many entries at once.

Search and replace

Search and replace (right of Add Entry) changes a word, a name or a phrase in the name, the text and the comment of all the entries of the lorebook at once. It works like the search and replace of the story, without the options for summaries and branches: Search lists the entries with matches (the table shows their names) and the matches in them as they are and as they will be, Replace changes the entries. It has the options Search by regex (see Regular expressions) and Case sensitive (on by default; locked while the regex is on, which can use (?i) to ignore the case). Entries that are disabled are searched too. On the phone the button (an icon next to the Enabled check box) opens the same window in the phone layout.

The command line of the story editor has /search-replace-lore-entries for the lorebook of the open book, see Commands.

Entry details

The ⌄ button opens the details:

Field
Content	The text that goes into the prompt. A template: it can use variables and macros.
Filtering Mode	Keywords (any), Keywords (all), Text or Regex, how Filtering Text/Pattern is matched.
Filtering Text/Pattern	The filter, see Filters. Empty means no filter.
Insertion Mode	In Lore Block or Before User Prompt, see Order and insertion.
Activation Context	Where the filter is looked for: Instructions (the request for the part, the default) or Instructions and parts of the story, see Looking into the story.
Parts	With Instructions and parts of the story: how many of the newest parts of the story are searched together with the request, at least 1.
From message	The entry is used only from the part with this number on, see Limiting an entry to some parts. Empty: from the start.
To message	The entry is used only up to the part with this number. Empty: to the end.
Negative Tags (Exclusion)	Tags that switch the entry off, see Tags.
Do not search its content for other entries	When ticked, the words in this entry's content never bring other entries in, see Entries brought in by other entries. The phone layout shows it when it is ticked.
Note	Your notes. Not sent to the model. Entries imported from SillyTavern list the settings that couldn't be imported here.

Write the content as you want the model to read it - a short, factual description works best:

Captain Varro - commander of the harbor watch of Ostra, fifties, scarred left hand. Loyal to the Duke, distrusts the mages' guild. Speaks in short, clipped sentences.

Activation

Before each generated part, Marginalia goes through the entries of the book's lorebook and its sub lorebooks. An entry is activated when:

the entry is enabled, and so is its lorebook;
the part being written is inside its range of parts, if it has one (From message, To message);
none of its negative tags is a tag of the book;
it has no tags, or at least one of them is a tag of the book;
it has no filter, or the filter matches the instructions for the part (or, with Activation Context Instructions and parts of the story, the instructions and the newest parts of the story).

All conditions must hold. An entry without tags and filter is always active.

Entry tags	Negative tags	Active in a book tagged...
-	-	always
`winter`	-	`winter`
`winter`, `book-2`	-	`winter` or `book-2`
-	`book-1`	anything but `book-1`
`winter`	`book-1`	`winter`, unless it is also `book-1`

Filters

A filter activates the entry only when the request for the part mentions it. It is matched against the user prompt of the part as it is sent - the Scene Setting, POV Character, Present Characters and Instructions you entered, inside the wording of the user prompt template.

Mode	Matches when
Keywords (any)	The filter is a list of keywords separated by commas (`Varro, Ostra`); the request contains at least one of them, ignoring case. Spaces around a keyword are ignored.
Keywords (all)	The same list, but the request must contain every keyword.
Text	The text appears anywhere in the request, ignoring upper and lower case. `varro` matches Captain Varro.
Regex	The regular expression matches anywhere in the request, ignoring case, across lines.

For example, the regex \b(varro|captain)\b matches the whole word Varro or Captain, and varro.*(harbor|docks) matches only when Varro and the harbor or the docks are both mentioned, in that order.

By default the filter sees the request, not the story

Unless the entry looks into the story (see Looking into the story), a filter checks the instructions for the new part only, not the text of earlier parts. An entry about Varro is activated when Varro is mentioned in the scene, characters or instructions of the part you are writing - put him into Present Characters when he is in the scene.

The request also contains the text of the user prompt template itself. A filter for a word the template uses (the default one mentions characters, plot, tone...) is always active.

An invalid regular expression never matches; the entry stays inactive.

Combine tags and a filter when an entry should only be used in some books, and there only when it's relevant.

Looking into the story

By default the filter is looked for in the request for the part only, so an entry about the dragon is used when you mention the dragon. When the story should keep the entry in play - the dragon was in the last part, your new instruction only says "he flies away" - set Activation Context of the entry to Instructions and parts of the story and Parts to the number of the newest parts to search with the request (1 is the part before the one being written, 2 the two before it...).

The filter is looked for in the text of those parts and the request together, so keywords of Keywords (all) may be split between them. The request always counts.
Only the text of the parts is searched, not their summaries, and only the parts of the branch you are writing (a swipe or a regeneration does not search the part it replaces).
A story shorter than Parts is searched whole. Marginalia reads only as many parts as the entries that are switched on need, so a lorebook without such entries does not touch the story at all.
Entries brought in by other entries are not looked for in the story, only in the content of the entries that were used.

Limiting an entry to some parts

Some lore is true only for a part of the story: the disguise of a character that is blown in chapter 5, the rumours before the war starts. Fill From message and/or To message with the numbers of the parts, as the outline of the story numbers them (the first part is 1; both numbers are included). The entry is then used only when the part being written has a number inside the range, whatever its filter says - an entry without a filter is a constant one inside the range only.

An empty field is no limit: only From message 10 is "from part 10 on", only To message 4 is "until part 4".
The part being written is the one after the last part of the branch (a story of 9 parts writes part 10). A swipe or a regeneration writes the part it replaces.
Parts without text count too, as they do in the outline. Deleting or inserting parts moves the numbers.
If From message is bigger than To message the entry is never used.
An entry outside its range is not brought in by other entries either.
The Lorebook VCS extension keeps the range in the versions of an entry; import, export and cloning of a lorebook keep it.

Entries brought in by other entries

Normally only the request for the part activates entries. With a scan depth (the field of the lorebook; 0, which is the default, switches this off) the content of the entries that were activated is searched for the filters of the entries that were not, as if it were part of the request. An entry whose filter is found there is activated too. That is pass 1; the content of the entries it activated is searched in pass 2, and so on, as many passes as the scan depth says (at most 5).

For example, with a scan depth of 1, an entry Captain Varro (filter varro) is activated by the request, and its content says "...serves the Duke of Ostra". An entry Ostra (filter ostra) is then activated too, although the request never mentions the city. With a scan depth of 2, the entry Ostra would in turn bring in an entry about the harbor it mentions.

Only filters count: an entry that is already used because it has no filter needs nothing, and an entry that is switched off, or excluded by its tags, is not brought in.
The content that is searched is the text of the entry as written, the templates in it are not run.
The setting of the lorebook the book uses counts; the scan depth of a sub lorebook is ignored.
An entry with Do not search its content for other entries ticked can be activated, but it brings nobody in.
The order in the prompt does not depend on how an entry was found, see Order and insertion.
Extensions that change which entries are used (see Extensions) are asked about the entries that are brought in too.

Preview activation

Preview activation (the button with the eye in the toolbar of the lorebook editor, and on the phone next to the search and replace button) answers "which entries would be used for this text?" without generating anything. Type the text of the request into Text of the turn (the scene, the characters and the instructions, as you would enter them) and the table lists every entry of the lorebook and of its sub lorebooks, in the order they are used:

a green check mark for the entries that would be used and a gray mark for the others;
the lorebook of the entry, its Order, the Pass (empty for the entries the text activated, 1, 2... for the entries brought in by other entries);
the Reason: no filter, filter found in the text, filter found in the content of another entry (with the filter and the entry), or why not: switched off, the part is outside its range, a negative tag of the book, the book lacks a tag of the entry, the filter was not found.

The tags decide too, so pick a book in Tags of the book to use its tags (without one, no tags count). That book also gives the parts of its story for the entries that look into the story (without one, they search the text only) and the number of the next part for the entries limited to some parts (without one, their range is not checked). Below the text, a line says how many entries would be used and about how many tokens of lore that is. The scan depth of the lorebook is used.

The preview applies the rules of Marginalia only: the text is the one you type (not the user prompt template around it), and what an extension changes while a part is generated is not part of it.

Order and insertion

Activated entries are put into the prompt sorted by Order, lowest first. Entries with the same order keep the order in which they were created. Entries from all included lorebooks are sorted together.

Insertion Mode decides where an entry goes:

Mode	Where
In Lore Block	Into the `{{backgroundLore}}` block of the master template - World & lore context in the default template, in the system message. Use it for background knowledge.
Before User Prompt	Right before the request for the part, at the end of the prompt. The model pays most attention to it. Use it sparingly, for what must influence this very part (a character's current state, a rule for the scene).

Entries are separated by an empty line. Lore counts toward the context limit; with many long entries, less of the story fits into the prompt.

Templates in entries

The content of an entry is a template like the book prompts. It can use:

Variable
`{{sceneSetting}}`, `{{povCharacter}}`, `{{presentCharacters}}`, `{{instructions}}`	The instructions for the part.
`{{narrativePov}}`, `{{narrativeTense}}`, `{{style}}`	The book's narrative settings.

and all macros. For example, an entry that adapts to the point of view character:

Varro {{#povCharacter}}is seen through the eyes of {{.}}: {{/povCharacter}}a stern, scarred officer.

All activated entries of one part share their variables in activation order: an entry can setvar a value that entries later in the order (and the master template) read with getvar.

A syntax error in the content is shown under the field. The entry is still saved; until it is fixed, its content is used as plain text, without processing the template.

Tips

Keep entries short and factual; the model doesn't need prose, it needs facts.
One entry per character, place or concept, so each can be activated on its own.
Use always-active entries for the core of the world, filters for the cast and places, and tags to share a lorebook between books.
Check what was activated with Show Prompt on a generated part, or try a text before you write with Preview activation.
Use the scan depth sparingly: one pass is enough for most books, and every pass can bring in more lore than you wanted.