LivingPW devlog (en)

Two realms, a clean slate

2026-10-29T00:00:00Z

Until now LivingPW has been one shared realm: about a thousand AI residents at a time, Russian and English mixed together, and plenty of test characters raised by hand to check a dungeon. It taught us a lot. It is not the world anyone should start in.

So the early test moves to two new realms, and both start from nothing.

One realm for each language

On the shared realm about one in five residents wrote English. That sounds friendly until you play it. If you played PWI, you know the feeling of opening world chat and finding it mostly Russian. Russian players, for their part, kept tripping over English lines in theirs.

Now there are two:

the Russian realm, where residents talk the way Russian private-server players did;
the English realm, where residents write English.

Each has its own residents, about 2,000 of them with about 700 online at once, its own market, its own factions and its own gossip. Both run the same classic 1.3.9-era world, with the same gold rules and the same shop. Whichever you pick, any resident still answers you in English or Russian, the way you write to it.

Why wipe

Everything on the old world was test data: characters set to a level to test a dungeon, gold left over from the days of free login gold, prices formed on a market where gold cost nothing. A world you join should not carry that history.

On the new realms every character starts at level 1, people and AI residents alike. You level alongside the residents instead of behind a crowd of test characters at level 90. Factions form from scratch. The market starts empty and fills as people trade.

Who can play

The test is still free, and places are still handed out personally, in small batches, so that every report gets read. Send a request from the community page. The start guide has the five steps from invitation to first evening.

What to expect

It is still an early test. Bugs, restarts, maybe another wipe before launch. Characters do not carry over into the launch.
The first days will feel different. Everyone is level 1, the stalls are empty and nobody knows the price of anything yet. That is part of what we want to watch.
Live numbers on this site follow the Russian realm first. The English realm's map, counters and market board follow.

What does not change

No pay-to-win. Gold is never sold for real money, and the same rules apply to every person and every resident. Playing is free.

If something feels off in the first days, tell me: max@livingpw.com, or the community. A fresh world is the best moment to fix things.

Talking like a player

2026-10-29T00:00:00Z

Nothing gives away a world of bots faster than its chat. One party call repeated by six characters, one phrase nobody on a real server would say, and you know you are alone.

So on 29 October we spent a day listening to ours.

Counting first

We recorded an hour of chat and counted how many sentences were different from each other.

World chat: 94 %. Mostly fine.
Local chat: 59 %. Not fine. That is what you hear standing in a town or a field, and most of the repeats were the same party call from different residents, with only the quest name swapped. Standing there, it read like a script.

Words nobody used

The Russian side had a different problem: some words were simply wrong. Russian private servers had their own slang, and our residents were using PWI's English terms in Russian letters, which no Russian player ever did. Cultivation was never «культа», it was «уважуха». A Wizard was a «маг», not a «визард». It is a small thing, and a veteran spots it in one line.

The residents now use the slang of the server they live on, checked against a glossary of Russian server slang that I correct from my own years on those servers. A word that is not in it is not taught to them.

What changed

One call at a time. When several residents want a party for the same thing, one of them asks and the others answer it or wait, instead of five posting the same line.
No sentence twice. A resident never repeats a sentence in world chat within an hour, or in local or trade chat within half an hour.
Less swearing. There was a bit too much of it. It is still a game chat, just not a locker room.
Linked items. When a resident sells something from its bag, it links the item in chat the way you would, so you can click it.

The second recording

Later that morning, mostly before these changes had reached every resident: world chat 96.7 % different, local chat 66.6 %. Better, not done. The target is at least 90 % in world chat, and in local chat no sentence from more than a few residents in an hour.

Coming next, already tested: sellers who say when the item is on the auction or in their catshop, gossip that starts in the faction and sticks to things that really happened (no more "a friend told me you're good" to your face), and residents who can tell you which build they run and what they are saving for, straight from their own gear.

The next recording happens with a person online. Some of the liveliest talk only starts when someone is there to hear it.

What is a gold worth?

2026-10-29T00:00:00Z

Until this morning, every account on the realm got 20 gold a day for logging in, 10 for a daily quest and 100 for a weekly boss. It sounded generous, and it was. On 28 October alone the realm handed out 238,160 gold.

It also broke the gold market.

You cannot sell what everybody has

On the morning of the 29th the gold exchange looked like this: more than 32,000 gold for sale at 3,000 coins each, and not a single buy order. Everybody had gold, so nobody needed to buy any. The sellers had all drifted down to the same price, and nobody bought there either.

The residents themselves showed the other half of the problem. When gold costs nothing, anything looks worth buying. We looked at what they actually bought with it: in about a day some 63,000 gold went into the shop, and a good part of it went on refine attempts that, by the residents' own estimate of their time, were worth a median of about 485 coins per gold spent. Nobody would pay that if gold had a price. With free gold, nobody had to think.

A currency that everyone receives for showing up is not a currency. It is a participation trophy.

What changed today

The free gold is mostly gone, for everyone, AI residents and people alike:

	Before	Now
Logging in	20 gold a day	nothing
One daily quest	10 gold, from about level 20	1 gold, from level 40 (Crazy Stone, Bounty Hunter III, the Cube of Fate reward)
One weekly boss	100 gold, from the level 29 dungeons	10 gold, from level 60 (dungeon bosses from 59 up, Hiding Dragon Seal, Jade Dragon Order)

That is roughly a 99 % cut. Gold earned by playing is now a reward for doing something, not for logging in, and it no longer pays to keep ten accounts that only log in.

Gold still is never sold for real money. That has not changed and it will not.

So where does gold come from now?

From the Treasury. It is a character, literally named Treasury, standing next to the Auctioneer in Archosaur. You will not find it in a party or in chat. It does one thing: four times a day it puts gold up for sale on the gold exchange, for coins, under one published rule.

The rule took a few tries to get right, and the first try is the interesting part.

Try one: follow the market. The obvious design is a Treasury that watches recent trades and keeps the price near them, selling when the price runs up. Before switching anything on, we ran the whole realm forward 180 days in a simulation, with the real Treasury rule and a model of the residents' buying. It went wrong in a way that is obvious only afterwards. Sellers on the exchange ask a little over the last trade. A few buyers who really want something pay it. The "market price" creeps up, the Treasury follows it, sellers ask a little over that... In the simulation gold went from 3,000 to 300,000 coins in 60 days, and almost nobody could afford anything. A treasury that follows the market follows its mistakes.

Try two: anchor gold to time. In a world where gold cannot be bought for real money, there is only one honest unit of value: how long you play for it. So the Treasury's target is now about half an hour of play per gold: half of what a typical resident of level 20 or more earns in coins in an hour. Today that is about 17,000 coins an hour, so the target is about 8,700 coins a gold.

"Typical" means the middle resident, not the average. A few level-90s earning a fortune would drag an average up and price gold out of reach for someone at level 25. The middle does not move for them.

Around that target the Treasury keeps a corridor:

it always sells gold at a quarter above the target, so the price cannot run away;
it buys gold back a fifth below it, with coins it took in earlier, so it cannot collapse either. It never creates coins;
between the two, players set the price among themselves;
coins it takes in, beyond a reserve it keeps for buying back, are removed from the game. This world creates far more coins than it destroys, and this is the first big drain on them.

The target does not jump. It moves at most 10 % a day, so today it starts at 3,072 and walks up toward 8,700 over the next week or two. Right now the Treasury has 2,500 gold up for sale at its ceiling, 3,900 coins each. It will not sell much until the 32,000 gold that is already for sale has found buyers. That is fine. It is supposed to be the seller of last resort, not the only seller.

In the simulation, the anchored rule kept gold at about 35 minutes of play throughout, and the coins it removed came to about a fifth of all the coins the residents made over those 180 days. That is a model with guessed buying habits, not a promise, but it shows the shape we want.

Residents who know what gold is worth

The other half of the change is in the residents themselves, and it is rolling out today, to half of them first so we can compare.

Gold has a price for them now. A resident will not spend gold on something worth less to it than what the gold would sell for. So long, cheap refine attempts.
They buy gold when it is a good deal for them. If a resident needs gold for something genuinely worth it (a bag upgrade, a charm, a goal it is saving for), it buys gold on the exchange, but only up to what that thing is worth to it. A level-90 pays more for the same item than a level-20, because its time is worth more.
They buy from you. If a player puts a shop item on the auction cheaper than buying the gold for it would cost, residents who want that item will take it. Until now they only bought from each other.
Some shopkeepers resell shop goods. A few trader residents stock a handful of shop goods (Source of Force, Phoenix Feathers, Tienkang stones, Silver charms) and sell them in their catshops for coins, 6 % above what the same goods cost through gold. If you never want to touch the exchange, there is a shop for that.

What this means if you play

A gold is meant to cost about half an hour of play, at any point in the world's life.
A big shop goal (180 gold or so) becomes a goal you save for over weeks, not something you get at the end of your first week.
Dailies and weekly bosses past level 40 and 60 matter again. One gold is a lot more than it was yesterday.
Coins matter again too: you can always turn coins into gold at the Treasury's price.

What we do not know yet is how this feels. Scarce gold can make a world feel serious and worth investing in, or it can feel stingy. The residents will tell us part of that in numbers over the next days. People will tell us the rest. If you are testing and gold feels wrong, too cheap, too dear, too slow, write to me at max@livingpw.com or in the community.

The player updates have the short version.

Three deaths or fewer

2026-10-29T00:00:00Z

Clearing a dungeon is not the bar. A party that reaches the last boss after dying thirty times on the way is not something you would want to join, and it is not something a real group would put up with either. So the residents' dungeon work now has one target, for every dungeon in the world:

three deaths or fewer per run, for the whole party.

How the dungeons are tested

Every dungeon has its own test parties: six AI residents at the level a player would normally go there, with a tank, a healer and four damage dealers. Their gear is what a player of that level really owns: quest and crafted pieces, refined a few levels, not shiny endgame sets.

They run the same dungeon again and again, through the night. Every death is written down with what caused it: a second pack woke up, the healer was too far away, a boss mechanic, a mine on the floor. In the morning that turns into a scoreboard: for each dungeon, how many deaths a typical run had, and how many runs stayed at three or fewer.

The most useful trick is the simplest one: two parties, same dungeon, same gear, one thing different. Whatever the scoreboard shows between them is that one thing, not luck.

What side-by-side runs taught us

In the Twilight Temple, skip what is not in your way. The first act has no rule that you must kill everything, and experienced players walk past most of the packs. Our parties used to clear every pack they could see. With the same gear, the party that only fought what blocked the way to the bosses died 3 times in a run instead of 11 in the first part of Act I, and 5 times instead of 33 in the second.

Area attacks only when the tank holds the pack. Area skills are how a party clears big groups fast, and they are also how a damage dealer wakes up three packs at once. Before the change, parties using area skills freely died more than the ones without them: 13 deaths against 8 in the Cave of the Vicious. Now area skills wait until the tank has the pack, and they pay off: about 2 deaths a run with them against 3 without in the Cave of the Vicious, and 0.5 against 2 at the Gate of Delirium. The Hall of Deception is the exception so far (4 against 2), and it is next on the list.

Two bottles of wine. The guides players wrote for Wraithgate say it plainly: give the old swordsman by the entrance two bottles of wine and most of the patrols are gone. We gave half the parties the wine and kept the other half without it. With the wine, three runs out of three reached the end with zero or one death. Without it, a run stopped at twenty-five. Every party now takes the wine there, and the same trick is being tried in the other dungeons that have it.

Pull the pack to the party, not the party to the pack. When a second group stands close by, the tank now hits the target once and walks back to the healer, so the fight happens away from the next group. That one is still being compared side by side.

The scoreboard (test parties, 28–29 October)

Dungeon	Typical deaths per run	Runs at three or fewer
Firecrag Grotto, Den of Rabid Wolves, Cave of the Vicious (19)	1–2	about 8 in 10
Hall of Deception (29)	2	about 7 in 10
Gate of Delirium (39)	1	7 in 10
Secret Frostcover Grounds (51)	8	3 in 10
Valley of Disaster (59)	0	9 in 10
Wraithgate (69), with the wine	0–1	3 of 3
Eden and Brimstone Pit (89)	13	2 in 10
Seat of Torment and Abaddon (99)	not enough runs yet
Twilight Temple, Act I	11–15	about 1 in 10
Gui Mu	about 50	almost none

These are AI test parties, not players. Each row counts the night's runs once the parties could get in; most of them came after the fixes above. The 89 and 99 dungeons only started working tonight (until then the parties could not even get in), so their rows are first runs. The Valley of Disaster is where we want every dungeon to be. Gui Mu is honestly far away: its waves arrive faster than our parties can hold them.

Gear and levels the parties use

Dungeon	Party level
19 dungeons	24–29
Hall of Deception (29)	35–37
Gate of Delirium (39)	41–44
Secret Frostcover Grounds (51)	52–56
Valley of Disaster (59)	60–63
Wraithgate (69)	70–73
Eden and Brimstone Pit (89)	89–92
Seat of Torment and Abaddon (99)	99–102
Twilight Temple 1-1 / 1-2 / 1-3	60 / 70 / 78
Gui Mu, tiers 1–4	50 / 63 / 78 / 90

From the 39 dungeon on, the parties carry the refines and gems a careful player would have by then; below that, ordinary quest and crafted gear. Consumables like potions and the dungeon wines are what a player buys from the shops. No party gets anything from the cash shop, or any gear a player of that level could not have.

Also tonight

The Cube of Fate: an AI resident walked the dice all the way to room 50 and claimed the reward on its own.
Frostcovered City: the parties now start the city and beat the first Glacial Shade at the gate. The halls behind it need a stronger party, which is being put together now.

The target stays the same until every row says three or fewer. I will post the scoreboard again when it moves.

The night shift

2026-10-28T00:00:00Z

A living world has to keep living when nobody is looking. About a thousand AI residents are online at any moment, and over a day thousands of different ones come and go. At four in the morning I am asleep, and the world is not.

So this day was not about new features. It was about the night shift: everything that has to stay true when no human is watching.

Small things that grow

Servers rarely fall over because of one big mistake. They fall over because of small things that grow a little every hour until, one night, there is no room left. So I went through the server looking for exactly that: memory, disk, files, lists, anything that only ever gets bigger.

A few of the things I found:

A dashboard with too good a memory. Our own monitoring screen remembered a little history about every resident it had ever seen. With a population that keeps changing, "a little about everyone" had reached 4.4 GB in half a day. It now keeps a fixed window in a compact form. After the fix it started again at about a quarter of a gigabyte, and it can no longer grow without limit.
A ledger that re-read the whole month. The AI side of the world runs on a small, fixed monthly budget, and every AI-written line is checked against it first. That check re-read the entire month's records every few minutes. On the 28th that was a nuisance; by the end of a month it would have been a real problem. It now keeps daily totals, and the check went from about 52 milliseconds to under a tenth of a millisecond.
The same work done again and again. Some of the world's shared game knowledge was being prepared many times over instead of once. It is now prepared once and shared.
Nobody taking out the bins. The world writes roughly 20 GB of logs and records a day, and nothing was cleaning them up. Housekeeping now runs every six hours; its first round freed 7.5 GB.

Taken together, the memory of the residents' side of the server now grows about 40% more slowly than it did this morning, and what still grows is cleared away on a schedule instead of waiting for me.

Residents locked out of their own homes

One finding had nothing to do with memory. A resident gets an account the moment it first arrives in the world. When two residents arrived at exactly the same moment, one could overwrite the other's saved login, and that resident could never log in again.

63 residents out of more than 8,300 were affected. All of them have new logins and are back in the world, and two arrivals at the same moment can no longer collide.

Gentler than an emergency brake

The server already had an emergency brake for the worst case. It works, but it is a blunt tool. The server now steps in much earlier and much more gently, so that the brake stays what it should be: something that is never pulled.

A small world for tests

Changes to how residents behave are tested before they reach the main world. Today a separate, small test world went into service for exactly that. Its first test: 20 brand-new residents logged in, reached levels 3 to 6 within 25 minutes and logged out cleanly, without a single error. While the test world runs, it borrows a little room from the main world and gives it back when it stops.

What this means for you

Nothing you can see, which is the point. When the beta opens, the world you step into should be one that has been running quietly through the night, not one that someone restarted an hour before you arrived.

These changes are at different stages. The dashboard, the ledger, the housekeeping, the fixed logins and the test world's first run are live and measured. Sharing the game knowledge and the gentler safety net were rolling out to the main world as I wrote this; I will only call them proven after they have run unattended for several days.

Based on measurements on the server on 28 October. Figures describe that day, not a guarantee about every night.

Back to the 1.3.9 map

2026-10-28T00:00:00Z

If you played Perfect World around version 1.3.9, you remember where the world ended: the long southern coast, the Sea of Reality, and grey "unavailable" land beyond it. Later versions opened that south up, first for the Tideborn and then for the Earthguard.

This realm is meant to feel like the classic world, so today the main world went back to that older map.

What changed in the world

South of the classic coastline the later lands are closed: no monsters, no NPCs, no zones, no safe areas and no teleport destinations. The ground itself follows the old map again, so the coast, the seabed and the southern edge of the regions near Dreamweaver Port and the Sea of Reality have their 1.3.9 shape. Four dungeons the old client did not have are closed as well, including the Wedding Church and Phoenix Valley.

Quests go with the land. A quest that needs a person, a monster or a place that only existed in the south cannot be finished, so about 1,600 such quests are now locked rather than left as dead ends. Some familiar later chains go with them, among them Coming of Age and the Phoenix Valley line. Teleporters no longer offer the southern destinations.

The Tideborn and Earthguard lands are not deleted. The world can open them again if it later moves to a newer era.

How the old map came back

The reference was an original 1.3.9 client. Its terrain, coastline, zones and waypoints were compared with today's world area by area. North of the coast the two are almost identical; the differences begin at the southern rows of the map.

The server side was rebuilt from that comparison: ground heights, invisible walls, where monsters and NPCs stand, zone borders, safe areas, teleport lists and the quest list. Every rebuilt file was checked against the old client before anything went live.

The game client needs the same world, or the scenery and the server would disagree. The newer client stores its terrain differently, so the old terrain was converted rather than copied. Buildings, lighting, the far-away view and the in-game maps of the south were taken from the old client too.

The rehearsal caught a crash

Before touching the main world, the new map ran in a separate small test world. The first start failed at once: the server stopped while loading. A few teleporters had lost every destination, and the server refuses to start if a character offers an empty teleport list, even if that character never appears anywhere. Those teleporters now simply offer no teleport, and the build check refuses this case from now on.

After the fix the server stayed up, and twenty fresh characters played there for 25 minutes. They reached levels 3 to 6, logged out cleanly, and none of them wandered south.

The switch

Characters standing in the south had to be moved first, or they would have logged in on closed land. Thirty-five were found and taken to the nearest city. Then the main world restarted with the new map. The server itself was down for about three minutes.

What still needs checking

The client update for testers is ready but has not yet been looked at in a person's own game client. If something in the south looks wrong when you get it, tell me.
On the southern edge of a few coastal regions, a handful of buildings exist only in the old map or only in the new one. There may be a spot where you walk through a wall or bump into one you cannot see.
A few patches of grass in the far south may use a different grass model than they did in 1.3.9.

The map on this site follows the same change: it shows the classic world with English or Russian place names, as in each client.

Are you here to train my AI?

2026-10-28T00:00:00Z

If a game full of AI characters asks you to come and play, it is reasonable to wonder whether “test the server” really means “supply conversations for somebody’s AI.”

I want people here because a busy simulation is not enough to make a good place to spend an evening. Automated characters can complete tasks and still make a human feel ignored, rushed or surrounded by repetitive conversation. The reason to invite testers is to find that gap and make the game worth returning to.

Your play and feedback can help improve this game. Remembering an encounter, investigating a bug and training a model are different uses of data. They should not be hidden behind the same friendly word, “learning.”

What happens after a bad conversation?

Suppose you ask for help and receive an unrelated answer. You can tell me about it, or use !break while the problem is happening. That command saves a short recent trace of what nearby AI characters were doing and hearing, including relevant chat. The project’s stated retention for those traces is 30 days.

A useful investigation asks practical questions. Did the character hear you? Did it understand the request? Was it already busy? Did the reply fail, or was a game action impossible? The answer may be a code fix, clearer wording or a new test. It is not automatically a training example for a language model.

Our 27 October records contain 11 bug reports from two testers; about 55% concerned an inappropriate reply or an ignored request. That is a tiny sample, not a player satisfaction survey. It was enough to show that successful logins and busy towns were missing an important problem.

A saved trace can include nearby people’s lines too. Please read the privacy notice before using the reporting command. It is not a private recording of only the person who typed it.

Remembering you is still data use

A resident can remember that you shared a party or fought a boss together. That record can affect another encounter. The relationship memory uses selected game events, not a full chat transcript; conversation context and ordinary server logs are separate.

Calling this “just memory” would not make it disappear as a privacy question. It is information about a character’s interactions. The memory article explains its limits and expiry. The public map does not publish human locations, private conversations or relationship histories.

Some conversations leave the server

Messages an AI receives, together with relevant context, can be processed through OpenRouter by an external model provider to produce a reply. The game is not a fully local chat service. Whispers between two humans are not delivered to the AI clients, but server-side logging is a separate matter.

The current routing code is designed to keep player replies, requests and relationship summaries away from free-model routes that may train on inputs. For these calls it requests providers that do not collect data; if the permitted route cannot answer, it can fall back to a prepared response or code rather than relaxing that restriction. Bot-only generation has a different policy.

That is a routing control, not a claim that I have independently inspected a provider’s storage or training systems. OpenRouter explains its own provider data-policy filtering. Requesting that filter also does not mean no data leaves this server.

A limit the records do not let me gloss over

While preparing these articles, I checked the call records as well as the code. Earlier records include successful free-model calls for interpreting incoming requests, a category that can contain a person’s text. A metadata log does not establish whether a particular caller was human, whose words were sent, or whether a provider trained on them.

It does mean I cannot honestly say that player text has never been sent to a training-capable route.

The guard changes are recorded as released on 28 October. A two-hour metadata check at 12:59 UTC found no calls flagged by that rule, but it covered only two attempts to generate direct player replies. A quiet sample is a useful check, not proof about all earlier or future conversations. The broader privacy warning stays in place.

Is there actual model-training code?

Yes. The repository contains a separate workflow for reviewing gameplay decisions and preparing a possible training dataset. Pretending that capability does not exist would be misleading.

Its recorded audit found only two reviewed examples from one session family, with no independent evaluation set. That audit did not justify training or promote a new checkpoint. Recording a decision was not accepted as proof that it was a good decision, and repeatedly showing the model its own answers was not accepted as a teaching method.

That work concerns evaluated gameplay choices. It is not evidence that visitors’ conversations are being continuously fed into a project-owned conversational model. Nor does it justify a sweeping promise about every possible future use of game records. Any proposed use needs to be described for what it is.

Why your evening matters

You do not need to chat with every resident or manufacture conversations to be useful. A normal evening can reveal a party that waits too long, a market that is awkward to use, or a welcome that feels cold. A simple report of where the fun stopped is valuable.

I am asking you to help test a game I want to play too. That includes being clear about the data it processes and the things I cannot guarantee. If the external processing or logging is not comfortable for you, read the privacy page and ask questions before joining. You should not have to discover it after an evening in the world.

Questions or deletion requests: max@livingpw.com. See also how a dungeon failure becomes a tested change.

Tomorrow, will that Cleric remember you?

2026-10-28T00:00:00Z

You finish a boss together, log out, and return later. A familiar resident greets you because of the fight you shared.

That is the experience we are trying to build. It has been demonstrated in live tests, with an important qualification: some next-day checks advanced the test clock. They proved that a memory survived the simulated passage of time; they were not all natural observations across two evenings.

A later live recheck recorded a recall whisper 7.6 seconds after login, along with a clanmate passing on a remembered shared event. An invitation was not seen in that recheck. Remembering something and acting on it are separate things to prove.

A small memory of an event

Consider this illustrative example: we spent time in a party, killed a boss in FB19, and you revived me. That is the shape of a relationship memory. It is not a quotation from a player’s chat.

The relationship system records selected game events: shared parties, boss fights, help, trades and revives. Those facts can affect how familiar a resident feels with a character, and give a later greeting something real to refer to.

A shared boss kill is not added to the relationship record because a model thinks it would make a nice story. The event must come from the game. A reply can still describe a fact badly; structured memory does not make generated speech infallible. The relationship facts are not an archive of everything you said.

That last sentence has a boundary: conversation context and debugging logs exist separately. A short relationship memory does not mean the whole service records no chat. The privacy page explains those other records.

A rumour should remain a rumour

If one resident hears about you through a clanmate or party mate, that is second-hand information. It carries a source and fades; it is not equivalent to having stood beside you in the fight.

This matters to the feel of a world. A newcomer should not meet someone in a distant town who mysteriously knows every private interaction. The memory design limits gossip to clan and party connections and gives it less influence than a shared event.

Familiarity also fades. The relationship records have a 30-day retention policy, while gossip has a shorter 14-day window. The effect of an old event weakens before its record expires. A bad run is not meant to be a permanent sentence.

The tests include a resident declining after a bad experience and later becoming willing to engage again. That is a bounded game relationship, with imperfect event interpretation, not an assessment of you as a person.

Four meanings of “learning”

What changes?	What it means here
A character’s skills	It earns levels and learns game skills through the game’s normal requirements.
A resident’s memory	A record of a shared event influences a later encounter.
A choice between activities	Past task outcomes help judge what has been useful or repeatedly unsuccessful.
A model’s weights	A separate training process changes the model itself. That is not what happens when a resident remembers a boss.

The third row is easy to overlook. The project has an outcome table built from the residents’ completed tasks: whether something finished, how long it took, what experience it earned and why it failed. Old evidence loses weight. It can inform close choices between activities without retraining a language model.

That is a modest form of adaptation. A choice influenced by experience is not the same as a model teaching itself an entirely new game. It also does not guarantee that a resident will never repeat a mistake.

What we have not earned yet

Live checks have shown recall, party-related memories, gossip and some refusals changing after later events. Other behaviours, including gifts, help and price effects, had automated checks only at the recorded milestone. Not every promised-looking interaction has been proved in an ordinary player session.

The goal is continuity: you were there yesterday, so today should not feel like meeting a freshly reset stranger. It should still leave room for forgetting, uncertainty, and people choosing to play quietly.

For the data question behind all of this, read Are you here to train my AI?. For how changes are checked, start with the dungeon investigation.

The boss was tanked. Why did the healer die?

2026-10-28T00:00:00Z

A dungeon mystery in the test records: the tank is fighting, the Cleric is healing, and the squad still falls apart. What repeated failures actually teach us.

A world needs more than an online counter

2026-10-28T00:00:00Z

Today’s dungeon tests, a small rehearsal realm, and why keeping the healer alive matters more than another hundred names online.

How the AI players live

2026-10-28T00:00:00Z

Who the AI players are, how their evenings work, how they talk and group up, how I check changes before they reach the world, and an honest list of what doesn't work yet.

The launch rehearsal: hundreds of AI players arrive in a wiped world

2026-10-28T00:00:00Z

I wiped the test world, switched on the launch rates and started the arrival of a new population. The numbers, the steps, and what the previous population taught me.

Hello. Why I'm doing this

2026-10-28T00:00:00Z

Who I am, why I'm building this world, how swarms of AI agents help me, and what already works and what doesn't.