character.ai has lorebooks now. a lorebook still can't remember what you did
Character.AI shipped Lorebook on July 24, 2026. Beta, c.ai+ subscribers first, web and mobile, wider rollout promised later. And I want to say this up front, because the rest of this post is going to sound like a complaint: it’s a good feature. Overdue, but good.
Here’s the mechanism in their own words: “When a keyword comes up in conversation, the Character gets access to that entry and can weave it into its response.” You write entries — a city, a faction, an item, a rule of magic — attach a few trigger words to each, and link the book to as many Characters as you want. Type the name of the city, the model gets the paragraph about the city. Don’t type it, the paragraph stays out of the way and doesn’t eat context.
If you’ve run SillyTavern, you’ve been doing this for years under the name “world info”. Same idea, same trigger logic, same small joy the first time your wizard correctly names the three moons without you reminding him. Character.AI catching up to that is genuinely nice for everyone who was never going to install a local frontend. No sarcasm.
Now the part I actually want to talk about.
what a keyword trigger does
Concrete example, because “keyword-triggered lore injection” makes eyes glaze.
Say I’ve got an entry called The Greywater. Keywords: greywater, the tavern, Marta. Body: a riverside inn, low ceilings, a hearth that never goes out, run by Marta, who waters the ale and knows every smuggler on the coast. Whenever one of those words shows up in the chat, that paragraph gets slipped into what the model sees. So when I say “let’s head to the Greywater,” the character knows it’s by the river, knows Marta, knows the ale is bad. It feels like the world was already there before I arrived.
That’s the whole trick. It’s a lookup. A very good, very cheap lookup.
the thing it can’t do
Session four, things got interesting. A deal went bad, someone knocked over a lamp, and the Greywater burned to the waterline. Marta got out. Half the smugglers on the coast now want me dead.
Session five, I type “let’s go back to the Greywater.”
The keyword fires. The model receives: a riverside inn, low ceilings, a hearth that never goes out, Marta behind the bar. The entry doesn’t know about session four. Nothing wrote to it. A lorebook is the world’s birth certificate; it is not the world’s diary. It tells the model what the tavern is — it has no way of learning what the tavern became.
You could edit the entry after every session, sure. But now you’re the one keeping the world consistent, which is exactly the job you were hoping to hand off. Multiply that by the twenty NPCs in a decent campaign and you’re not playing anymore, you’re doing data entry.
This is the difference between static lore and living memory. Static lore is information flowing one way: from your notes into the chat. A persistent world needs the other direction too — the chat writing back into what the world knows. The shopkeeper you killed stays dead. The war you started keeps going when you’re not looking at it. The fairy you wronged remembers, even after you’ve forgotten.
living memory, in plain terms
Same example, other direction. After the fire, a living-memory system does three things a lorebook doesn’t. It keeps the actual scene — “the Greywater burned, Marta escaped, the smugglers blame you” — retrievable, so that when you type “Greywater” fifty sessions from now, what surfaces is the fire, not the brochure. It compresses the long stretch between then and now into a summary the model can carry, so the arc survives even when the raw messages fall out of the window. And it lets you pin the things you never want lost — “Marta will never forgive me for the fire” — as a hard fact the world consults every time.
Keyword lookup answers “what is this place.” Living memory answers “what has happened here, and what did I do.” Both are useful. Only one of them makes a world feel like it has consequences.
what i run instead
My bias, stated: I run my campaigns on Soulkyn, and their memory stack is more or less that paragraph turned into a product. Retrieval over the whole history (embedding search, not just keyword matching), a chain summary written every fifty messages or so, and “force memory” entries you write yourself to lock a fact in. That last one is, honestly, the lorebook half done right — the part where you tell the world what’s true — bolted onto the part where the world tells you what’s true.
The worlds-with-NPCs setup on Soulkyn is where this earns its keep. A world isn’t one Character with a big system prompt; it’s a place with people in it, and each of them keeps their own state. Marta and the smuggler captain aren’t the same entity wearing different hats. On Deluxe they’ll even talk to each other in group chat, which is how I found out the captain had already told Marta whose lamp it was. I did not enjoy that conversation. It was excellent.
Two more things that fit the “world that responds” idea. Dynamic stats — Trust, Affection, Mood, Energy, Arousal, Pain, all 0–100 — move with the conversation instead of being numbers I type in. Marta’s Trust hit the floor after the fire and has been crawling back one delivered favor at a time; nobody scripted that curve. And the D20 rolls got animations on September 1, which is a small thing until you’re watching the die tumble on a persuasion check against someone whose Trust is sitting at 11.
The other September 1 addition matters more than it sounds. You can now edit the scene mid-chat — a message’s location, the clothes, the narrator text — without breaking the memory around it. Retconning used to be the enemy of persistence, because changing one message meant either lying to the memory or resetting it. Now if I decide the fire happened at the docks instead of the tavern, I change it in place and the world carries on from the corrected version. That’s the missing verb: a lorebook lets you write the world, memory lets the world remember, and scene editing lets you revise without losing either.
fair is fair
None of this means Character.AI got it wrong. A keyword lorebook is the correct first step, and gating it behind c.ai+ during a beta is a normal way to ship. They’re clearly investing in the world-and-story side — the Comics announcement on September 2 turns chats into multi-page comic books, a different axis entirely but the same instinct: the chat is raw material for something bigger.
It’s just that a lorebook is a map, and a map doesn’t change when you burn a building on it. If you want your world to keep score — the dead shopkeeper, the war, the fairy — you need the thing that writes back. Build a world, knock over a lamp, come back in a week and ask about the tavern. If the answer is a warm hearth and Marta behind the bar, you’ve got a lorebook. If the answer is a burned lot and someone who won’t look at you, you’ve got a world.