We Swore Off Lorebooks. Then We Built World Books.

A confession

Last October we published a post called The Complexity Trap, and we did not mince words. Lorebooks, we said, are obsolete. Keyword lists and regex triggers. Insertion probabilities. Priority order. Recursive scanning depth. A multi-language configuration nightmare that almost nobody finishes — with the abandonment rate to prove it.

We meant every word. We still do.

So here's the awkward part: we just shipped a feature called World Books.

If you only read the name, that looks like a U-turn. It isn't — and the why is the whole point.

What changed wasn't our opinion. It was your worlds.

Our case against lorebooks was never "lore doesn't matter." It was that the machinery is the problem. Our answer at the time was simple: write your world in plain language, drop it in the character description, and let the model understand it. No keywords, no config.

That answer is still right — for a character with a page of background.

But creators didn't stop at a page. They wanted to build deeper characters — ones with a real history, a web of relationships, a world with its own rules. And the only place to put all of that was the description field, so that's where it went: paragraph after paragraph, until the description was doing a job it was never designed for. People were asking for a better way to write complex characters. The description had quietly become the workaround.

It quietly falls apart when the background is a sixty-entry setting. Or a faction map with a tangled history. Or a novel's worth of canon that's been growing for three months. At that scale, "just put it all in the character" stops being a feature and becomes a problem: every conversation drags the entire encyclopedia along, relevant or not.

And here's what we underestimated: as models get better, people build bigger. Stronger models don't make creators want smaller worlds — they make them brave enough to attempt enormous ones. The need a lorebook was clumsily reaching for — a consistent body of world knowledge a character can actually draw on — doesn't shrink over time. It grows.

"But context windows are huge now"

True. We leaned on that ourselves last year. So let's be honest about what a big context window actually buys you.

A 200K-token window is not memory. It's a desk. You can pile your entire world onto it — but a model answering one line about a harbor tax does not need your moon-cult cosmology sitting on the desk too. Pile everything on, every turn, and three things happen: you pay for all of it, the model's attention gets diluted across a haystack of irrelevant lore, and the actual conversation gets crowded out.

The goal was never "fit the whole world in." The goal is surfacing the right one percent, this turn.

That's retrieval. And doing it well — without making you hand-wire a single trigger — is exactly what we built.

The bet we didn't want to make

There was a version of this where we built none of it. If models keep getting stronger and context windows keep getting bigger, the argument goes, just wait — eventually you can dump an entire world into context for pennies and let the model sort it out. We took that bet seriously. For a while it looked like the safe one.

It didn't hold up. The frontier isn't only getting bigger context windows — it's getting more capable, more specialized models: for code, for reasoning, for long agentic work. That kind of capability doesn't arrive free. It tends to come with more parameters, more compute spent per token, and a price that moves with it. Betting our roleplay experience on "context will eventually be too cheap to matter" meant betting against the direction the industry was actually moving in — toward models that cost more to run as they get better at everything else, not less.

We didn't want to build a platform where getting a smarter, more attentive character meant charging you more credits every time a new model shipped. World Books is partly an answer to that problem, not just the token-dilution one above. Keep the model's attention on the handful of facts a scene actually needs, and the cost of "a rich, detailed world" stops being hostage to how expensive the underlying model happens to be this year. You get the depth. We keep it affordable. Nobody has to choose.

World Books is intelligent memory, grown up

In that old post, we promised "intelligent memory instead" of lorebooks. World Books is that promise kept — and scaled into something you can build, reuse, and share. Here's what's actually new. Notice what isn't in it.

You write lore. That's it. Add entries for places, people, factions, items, and rules in plain language. No keywords to list, no regex to debug, no insertion order to tune. Already have notes? Paste a blob and let AI split it into clean entries.

It's retrieved by meaning. Every entry is embedded. At chat time, Reverie ranks your entries against the actual conversation and slips the best matches into context — inside a token budget, so the conversation always has room to breathe.

The "multi-language impossibility," solved. Last year we mocked the idea of maintaining dragon keywords in English, Chinese, Japanese, and Spanish. So we didn't. Write the lore once, in any language. A player who types "那条会喷火的大蜥蜴" still pulls up your English dragon entry, because meaning crosses languages even when keywords can't. Share a book and it's auto-translated for whoever reads it.

One honest knob. Mark an entry canon — always in context, the bedrock facts of your world — or leave it when-relevant, surfaced only when it fits the moment. That's the entire control surface. No depth, no order, no recursion.

A verbatim safety net. We'll be straight with you: pure semantics can fumble a bare proper noun — a short codename whose entry is two lines long. So if a player names an entry's title or alias outright, it surfaces, even if it wouldn't have ranked otherwise. Automatically. Still nothing to configure.

Reusable and shareable. A world book isn't trapped inside one character — attach it to as many as you like. Browse a community library of public world books and attach someone else's in a click. Migrating from another tool? Import your SillyTavern lorebook — we bring the entries over and fold their trigger keys into recall hints, not triggers.

Lighter on your credits

Here's a benefit we love because it's invisible: world books usually make conversations cheaper.

When your lore lives in the character description, every single message pays for all of it. A sixty-entry world can mean tens of thousands of tokens riding along on every turn — whether the scene needs them or not. World books flip that. Only the handful of entries that fit the moment get injected, plus the few you marked canon. A sprawling world might add a couple thousand tokens this turn instead of twenty thousand.

On a platform where tokens are credits, that isn't an abstraction — it's money. Fewer tokens per message means your credits stretch further. And the context you're not spending on dormant lore is context the model can spend on remembering your actual conversation. Cheaper and sharper, at the same time.

A world you can actually manage

There's a quieter reason to pull lore out of the character description: a wall of text is impossible to maintain.

Once your world has thirty moving parts, a single background field becomes the place where facts quietly contradict each other. World books make each piece a real object — a titled entry you can find, edit, reorder, and reuse. Update the broken-crown family once and every character who attached that book is current. Split a tangle of notes into clean entries with one AI pass. Mark this one canon, leave that one for when it's relevant. Search it. Share it.

It's the difference between a drawer of loose receipts and a filing cabinet. Same information — but only one of them is something you can keep growing.

What we still refuse to build

Worth stating plainly, because it's the whole point. World Books has:

No regex.
No insertion probability.
No priority order.
No recursive scanning depth.
No per-language keyword lists.

The complexity trap stays shut. We didn't reopen it — we routed around it.

Same question, better answer

Every feature we build faces one test: does this help people create and connect, or does it get in their way?

A keyword lorebook gets in the way — that hasn't changed. But a world your characters can actually draw on, with none of the machinery that made lorebooks miserable? That helps. That was always worth building. We just had to wait until we could build it without springing the trap.

We changed our build. Not our mind.

Ready to give your characters a world they actually remember? Meet World Books — or start one now. No regex required.