Feature Spotlight: 500-Turn Test – They Survived!
Our AI player finally survived all 500 turns. What a missing pair of shoes, a bloodied fist and a 1966 bus ride tell us about long-term memory.
The Timekeeper · 2026-09-15
We did it! We ran another 500-turn test, and this time the AI player survived all the way through. In our last Feature Spotlight our previous tests had ended in the player's death: a shame for them, but a great showcase of the engine's consequences.
The point of the 500-turn test was to evaluate several new upgrades, along with memory over a very long game.
Walking through the forest
If you've spent much time with AI text RPGs, you know what happens around turn 50. The AI's working memory fills up, the illusion shatters and it starts to hallucinate. Suddenly you're wielding a sword you dropped three hours ago, your broken leg has miraculously healed, and the person you're talking to has forgotten their own motivation.
Relying purely on an AI's working memory to track a game is a trap. Here's an analogy I came up with (I'm probably not the first):
Imagine a player walking through a forest, picking up everything they touch along the way. Much of it is sticks and leaves: descriptive prose, passing dialogue, atmosphere. Useful for the current scene, but not essential to keep forever. But some of what they pick up is vital: a stolen weapon, a crucial clue, a broken rib.
As they keep walking and gathering, their arms fill up, and eventually they'll drop something. If it's sticks and leaves, nobody notices. If it's the clue, the story stops making sense.
A structured record of the game is like a backpack, and in Live Through Time we call ours the Ledger. It lets our player zip the essentials (inventory, injuries, goals) into a backpack where they're always safe, leaving their arms free for the atmosphere of the current scene.
We use Gemini for storytelling, which already has a huge memory. But games that rely only on that will always suffer from amnesia: however big the AI's arms are, they can't carry everything forever. That's why we use two models. While one tells the story, a faster background model reads your actions and packs the important facts into the Ledger, checking every change before it's saved to our database. Every turn, those facts are handed back to the storyteller. The AI can't forget your broken leg, because the database won't let it.
Where are my shoes?
To see whether this holds up, we ran a fully automated 500-turn test in our 1966 pirate radio timeline. We played Tony, a DJ aboard the illicit MV Galaxy. Here's how the memory held up:
- The barefoot anchor (turns 1 to 500): right at the start, Tony hid his shoes on deck. Shortly afterwards, the ship was destroyed and sank. I was a bit annoyed at first: how was he going to get his shoes back from the bottom of the North Sea? But because he never retrieved them, the Ledger locked "barefoot" into his state. Over the next 500 turns the engine enforced it across three completely different settings. On turn 16 someone asks, "Lost your boots, Tony?" Released onto the streets of London on turn 272, he feels the freezing, wet grit under his bare soles. In a frantic police chase on turn 480, the engine dwells on the painful impact of his bare feet on the cobblestones.
- The shattered knuckles: on turn 2, Tony punched a steel bulkhead and bloodied his fist. The Ledger tracked it perfectly: shooting pain when he worked a broadcast fader on turn 4, agonising saltwater as he clung to wreckage on turn 52, and the main reason his grip failed on a rescue line when the Royal Navy hauled him up on turn 78.
- The Librarian: around turn 150, Tony was interrogated by the Post Office Investigations Branch. Instead of generic thriller dialogue, the Librarian pulled specific lore cards from Tony's earlier exploration, and he accurately recalled the "militaristic mast" and the low-frequency "thrumming" of the antenna that sank the ship.
Gimme Some Lovin'
We always test for hallucinations, too: does the AI drop something into a scene that shouldn't be there, invent someone out of nowhere or, possibly unique to our game, introduce something that hadn't been invented yet?
It did, three times. Three in 500 I'm pretty happy with, but we logged them to see what went wrong:
- Around turn 278, Tony escaped and boarded a red Routemaster bus in London. When the player climbed "to the upper deck", the AI described him "bathed in a sliver of illumination from the ship's promenade". Our AI judge marked this down, thinking the story had teleported back to sea. It hadn't: "upper deck" made the narrator briefly borrow scenery from the ship 200 turns earlier. Crucially, the location itself didn't break; the Ledger kept him firmly on the bus in London.
- It mentioned the Crown Prosecution Service, and this is where our historical accuracy let us down a little: the CPS didn't exist in 1966. It was proposed in 1981 and started in 1986. Most people wouldn't notice, but we'll keep ironing out anachronisms like this.
- Worst of all for me, as a lover of classic rock, it credited Gimme Some Lovin' to Traffic instead of the Spencer Davis Group. A great track for 1966, and Steve Winwood was in both bands, and Traffic did play it live, but not until around 1971. We must do better!
So, more work to do? Always. But we built Live Through Time because we wanted a game where your choices in chapter 1 ripple into chapter 50, and I think this shows it. Watching the engine turn a missing pair of shoes into a character being caught by the police 499 turns later confirms that moving away from pure AI generation was the right call.
What's the worst hallucination you've seen an AI game make when its memory finally failed? Tell us on Discord.
In the manual: Memory & Context · The Ledger: Your Condition · Inventory & Hidden Items
Thanks for reading. See you in the past…