Which AI companion apps can actually run a campaign, why the world state falls apart after twenty turns, and the system prompt we use to keep it honest.
There is no game engine behind an AI RPG. Nothing is tracking your hit points, your inventory or the fact that you burned the bridge in turn nine. The model is writing plausible text, and plausible text is not the same as consistent state, which is why a campaign that felt brilliant for ten turns quietly contradicts itself by turn twenty five. Gold you spent comes back, an NPC you killed walks into the tavern, and the model starts taking your character's turns for you. That is the real problem with an AI roleplay RPG companion, and it is a chat problem rather than an image one.
Below: the apps whose memory and roleplay handling survive a long session, our verdict on what running a campaign this way is actually like, and the system prompt, turn format and state block we would start from.
Highest overall score of the eight and it locks a character once you are happy, which for a campaign means your recurring cast stays the same cast instead of being re-rolled every session.
Built around realistic, passionate relationships, so it plays a companion travelling with you better than it plays a neutral narrator. Good for a two-hander, weaker for a party.
Privacy-first with a free tier, which is the sensible place to stress-test a long scenario before you pay for one. Campaigns fail slowly, so you want cheap failures.
Strong memory at $12.99, the cheapest of the high scorers. On this page memory is not a nice extra, it is the entire mechanic, so this is the value pick.
Same score band, different model underneath, so it drifts in different places. Useful when a campaign has become unrecoverable rather than merely wobbly.
Sits in the same 8.6 band as Lovel and Xotic. Worth a second run when a scenario has collapsed on the platform you started with and you want a clean context.
The largest character library here, so you can cast recurring NPCs from characters that already exist instead of writing and maintaining every one yourself.
Romance-leaning and not image-first, at $12.99. Fine if your campaign is almost entirely conversation and you rarely ask for a picture of the scene.
These scores are our editorial judgement of the category, not of any one app and not measurements. They describe what tends to hold and what tends to break when you run a campaign with an AI companion anywhere.
| Criterion | Rating | What it means for you |
|---|---|---|
| Turn discipline | The most common failure. Left alone, the model narrates three turns ahead, resolves your action and decides how you felt about it. You have to state the turn boundary in the system prompt and re-state it when it slips. | |
| Stat and inventory tracking | The weakest point of the whole category. Numbers are text, not state. Hit points heal themselves, spent gold reappears, and an item you handed over is back in your pack two scenes later. If numbers matter to you, keep them yourself in a note. | |
| World consistency after twenty turns | Geography rearranges, a dead NPC returns, and a locked door becomes an open corridor. Distance from the event matters more than how important it was, so major plot points decay just as fast as small ones. | |
| Stakes and failure | The model wants to please you, so almost every attempt succeeds and every threat is survivable. Ask for a real failure rate up front, or you get a story with no tension in it. | |
| Improvisation and scene writing | The genuine strength, and the reason to do this at all. Walk off the map into an unplanned village and you get a village, with people in it, immediately. No human GM improvises that fast. | |
| Session length before a reset | You get a solid evening before the earliest turns fall out of context. Ending each session with a written recap and pasting it in next time is what buys you a campaign rather than a series of one-shots. |
We revise these as memory handling improves across the apps. If your campaign holds together better than this suggests, tell us how you set it up and we will update the page.
Almost all of the work here is in the first message. You are not describing a character, you are writing a small rulebook, and the parts you leave out are the parts that break first. Set this up once, save it, and paste it at the start of every session.
| Setting | Suggested value |
|---|---|
| Role assignment | Say explicitly that it narrates the world and every NPC, and that you control exactly one character. Without this line the model will write your dialogue for you. |
| Turn format | Ask it to end every response with the situation and then stop. Something as simple as "end each turn with what I can see and then wait" prevents most of the narrating-ahead problem. |
| State block | Ask for a short block at the end of each turn listing location, hit points, notable inventory and time of day. It does not create real state, but restating it keeps it in context and makes drift visible immediately. |
| Resolution rule | Decide how attempts are settled and say so. "Roll d20, tell me the number, 10 or under fails" works, and asking for the number makes it much harder for the model to quietly award you every success. |
| Agency rule | One line: never decide my character's actions, words or feelings. Repeat it the first time it is broken rather than letting it slide, because it escalates. |
| Scene scope | Keep each scene to one location and one objective. Sprawling multi-location turns are where the world state contradicts itself fastest. |
| Recap cadence | Every ten turns, ask for a five-line summary of established facts. Paste that summary at the start of the next session. This is the single habit that turns a chat into a campaign. |
Prompt to copy
You are the narrator and game master. You control the world and every NPC. I control one character and only I decide what she says, does or feels. Never write my actions for me. End every turn with what I can see and hear, then stop and wait for my input. Do not resolve more than one of my actions per turn. When I attempt something uncertain, tell me to roll a d20, ask for the number, and treat 10 or under as a failure with a consequence. After each turn, add a short block: LOCATION, HP, INVENTORY, TIME. All characters are adults. Setting and opening scene follow in my next message.
Four directions the same category can go in. Copy one, paste it into the character builder, and change the parts you disagree with.
One companion, one dungeon, tight scope. The easiest structure to keep coherent, because there is only one NPC to track.
Setting: a flooded dwarven mine, abandoned two centuries ago. I play a mercenary. You play Vessa, an adult scout who has been down here before and will not say why. Rules as previously given. Keep the map small: no more than six locations, and remember which ones we have already cleared. Vessa acts on her own initiative and can disagree with me. Track torches as a resource, four remaining. Open on the mine entrance at dusk.
Travel campaign built on supplies running out. Gives the model a countdown it cannot argue with, which is the cheapest way to create real stakes.
Setting: a long road across a plague-emptied province. I play a courier. You play Ilma, an adult physician travelling with me for safety. Rules as previously given. Track three resources every turn: rations, water, days remaining. Rations start at 8, water at 5. They only ever go down unless I explicitly find more, and you must never restore them out of kindness. Ilma has her own goal that conflicts with mine and I do not yet know what it is. Open on the second morning of the journey.
Planning phase then execution phase, with a visible timer. Structure is the point: it gives the model something to hold onto when memory slips.
Setting: a merchant bank in a canal city. I play a thief. You play Rennick, an adult forger and my only accomplice. Rules as previously given. Run two phases. Phase one is planning: I ask questions, you answer as the world, and we agree a plan of exactly five steps. Write the five steps down and restate them at the start of phase two. Phase two is execution: each turn is one step and one complication, and the guard rotation advances whether or not I am ready. Do not let me succeed at a step I did not plan.
For a world you intend to return to for weeks. Front-loads the shared facts so later sessions have something firm to be measured against.
Before we play, run a session zero. Ask me ten questions about the world, one at a time, covering geography, who holds power, what magic costs, what my character wants and who she has already wronged. When I have answered, write a CANON document of no more than fifteen numbered facts and treat it as fixed. At the start of every future session I will paste the CANON back to you. During play, if a new fact contradicts the CANON, stop and tell me rather than rewriting it silently. All characters are adults. Then begin play with my character arriving somewhere she is not welcome.
These are the words that decide whether your campaign holds together. Most people arrive from tabletop and assume the model knows the same vocabulary, and it does know the words, it just does not enforce anything behind them unless you ask.
| Term | What you actually get |
|---|---|
| Game master mode | The instruction that the model narrates the world and every NPC while you play one character. Without saying it out loud, most apps default to a single companion in conversation with you, not a narrator. |
| Turn | One exchange with a clear stop. If you do not define where a turn ends, the model will keep going and resolve several of your intentions at once. |
| Agency bleed | The model writing your character's words, actions or feelings. It starts with something small, usually a reaction, and grows until you are reading rather than playing. Correct it the first time. |
| State block | The short LOCATION / HP / INVENTORY line at the end of each turn. It is not real bookkeeping, it is a way of keeping the numbers inside the recent context so they drift more slowly and drift visibly. |
| Inventory drift | Items reappearing after being spent, sold or given away, and quantities changing without an event. The single most reliable symptom that you have left the model's working memory. |
| World state | Everything that is true because of what has already happened: doors opened, people killed, favours owed. It is the first thing to decay, and it decays by distance from the current turn rather than by importance. |
| Retcon | The model rewriting an established fact instead of admitting the contradiction. Ask it to flag conflicts rather than resolve them and you catch these while they are still fixable. |
| Resolution roll | Any dice or randomiser used to settle an uncertain action. Asking the model to state the number before the outcome is what stops success rates quietly rising to a hundred per cent. |
| Session zero | An opening conversation that fixes the world's rules before play begins. Fifteen numbered facts you can paste back in later are worth more than any amount of atmospheric preamble. |
| Recap | A short summary of established facts, generated at the end of a session and pasted in at the start of the next. The nearest thing to a save file this format has. |
Because nothing is being stored as state. The model only sees a window of recent conversation, and once your early turns fall out of that window they stop existing. Ending each session with a written recap and pasting it back in at the start of the next is the only reliable fix.
Weight memory and roleplay handling over image quality, which is the opposite of most categories on this site. The apps at the top of this page score highest on the combination, and the cheaper strong-memory options are worth a look given how much message volume a campaign uses.
State once, in your opening message, that you control exactly one character and that it must never write her actions, words or feelings. Then correct it the first time it slips, because agency bleed compounds: if you let a small reaction pass, you will be reading a novel within ten turns.
Not reliably. It can write numbers that look like tracking, but there is no engine checking them, so they heal and refill on their own. Ask for a state block after each turn so drift is at least visible, and keep anything you genuinely care about in a note of your own.