Best AI RPG for a Long Campaign (2026): Which Platforms Actually Remember
A long AI campaign fails in two different ways, and most guides only measure one of them. The first is memory: the world forgets your companion died, resurrects an NPC, loses the sword you threw in a river. The second is longevity: nothing is forgotten, and you stop caring anyway, because the platform has run out of things to do to you.
We score those separately — Memory & Continuity and Longevity — as two of the seven axes behind every rating on this site. Across the 26 platforms and games ranked below, the results are less flattering than the marketing:
Only four have scored above 3 on both axes — and two of those four charge the same price for it, which is not money.
And nobody has ever scored 5 on either one. Of our seven axes, these two are the only pair with no product at the ceiling.
One scoping note before the tables. Everything ranked here is a platform you can play today. Several products currently in closed alpha or closed beta are deliberately absent: we don’t score pre-release software, because it isn’t fair to grade an unfinished build or to let an early verdict outlive the version it described. If you want to know what’s coming rather than what’s playable, our first looks cover those separately.
The Memory Ranking
| Platform | Memory & Continuity |
|---|---|
| Voyage | 4.25 |
| Hidden Door, WyrdTale, Tabled, Craft | 4 |
| Friends & Fables, Master of Dungeon | 3.5 |
| Old Greg’s Tavern, Deep Realms, MacerAI, Infinity DM, AI Dungeon, FableAI, AI Game Master, ArcQuill, Tidefall, Tales RPG, NOPOTIONS, aiga_ | 3 |
| Chub AI, Janitor AI, Dunia, Perchance, Vantiel | 2 |
| AI Realm, Character.AI | 1 |
Nothing has scored a 5, and that is not us being stingy. A 5 would mean a campaign that holds together indefinitely without you doing maintenance. Every platform here is managing a hard limitation rather than removing it — a model has a finite context window, and everything else is scaffolding built around that fact. The good implementations delay the failure. None prevent it. If you understand nothing else before subscribing, understand that, and read why your campaign falls apart at turn 50.
The Longevity Ranking
| Platform | Longevity |
|---|---|
| Voyage, Hidden Door, WyrdTale | 4 |
| Craft, NOPOTIONS | 3.5 |
| Janitor AI, Old Greg’s Tavern, Deep Realms, Chub AI, Vantiel, Tabled, Tidefall, Tales RPG, aiga_, Master of Dungeon | 3 |
| ArcQuill | 2.5 |
| Friends & Fables, MacerAI, Infinity DM, AI Dungeon, FableAI, Character.AI, Dunia | 2 |
| AI Game Master, AI Realm, Perchance | 1 |
Memory and Longevity are two of seven problems an AI RPG has to solve, and on our measurements Longevity is the second-hardest of the seven.
Why the Two Lists Are Different
Look at what moves between them, because this is the part that decides your six months.
Friends & Fables is 3.5 on Memory and 2 on Longevity. It has one of the better memory architectures we have tested — a Working Context the platform assembles before each response, filtering in only what the current scene needs. It genuinely works. And the platform still runs out of road, because its own developer has since shipped a successor and because Franz’s flaws surface precisely in the very long campaigns its structure is built for. It remembers you and stops being worth remembering.
Janitor AI is the reverse: 2 on Memory, 3 on Longevity. It forgets constantly. It also has joint-best NPCs in our directory — one of only three 5s we have given on that axis, alongside Tidefall and Voyage — and a community writing cards good enough that people keep coming back regardless. It forgets you and you return anyway.
AI Realm scores 1 and 1, and it is the most instructive failure here. Its ruleset is excellent — Mechanical Depth 4, behind only Friends & Fables and Craft. None of that survives a memory that cannot hold the campaign the rules are running. Depth without recall does not compound; it resets.
A platform can remember everything and still bore you. A platform can forget everything and still be the one you open on Friday. If you only check one number before subscribing, you will pick the wrong platform about half the time.
The Two Axes Nobody Has Ever Topped
Step back from these 26 and look at all 37 products we have scored, across all seven axes at once. Something shows up that neither ranking above can show on its own.
| Axis | Field average | Highest ever awarded |
|---|---|---|
| NPC Fidelity | 2.97 | 5 |
| Signature Design | 2.88 | 5 |
| Memory & Continuity | 2.71 | 4.25 |
| Determinism & Fairness | 2.61 | 5 |
| Player Agency | 2.51 | 5 |
| Longevity | 2.42 | 4 |
| Mechanical Depth | 2.27 | 5 |
Five of the seven have somebody sitting at the top of them. Somebody has scored a perfect 5 on mechanics, on fairness, on agency, on characters, on originality — every one of those has been solved at least once, by somebody, on hardware anyone can rent.
Nobody has ever scored 5 on Memory or on Longevity. In two years those are the only two axes with no ceiling holder — and they are the two this page is about. Voyage holds Memory & Continuity — 4.25 and that is as high as the field has reached. Longevity stops at 4.
The Longevity distribution is worse than its average makes it look:
| Longevity score | Products |
|---|---|
| 1 | 5 |
| 2 | 15 |
| 2.5 | 1 |
| 3 | 11 |
| 3.5 | 2 |
| 4 | 3 |
| 5 | 0 |
Twenty of the 37 sit at 2 or below. That is not a tail of bad products — it is most of the category, and it is the honest context for every “persistent memory” claim on a pricing page.
What actually predicts survival
The three cases above are anecdotes until you check them against the whole board, so we did. Here is Longevity compared against each of the other six axes across all 37 products.
| Longevity tracks… | How closely |
|---|---|
| Memory & Continuity | 0.64 |
| NPC Fidelity | 0.41 |
| Signature Design | 0.39 |
| Player Agency | 0.27 |
| Determinism & Fairness | 0.27 |
| Mechanical Depth | 0.17 |
A number near 1 would mean the two scores rise and fall together; a number near 0 means knowing one tells you almost nothing about the other.
Memory is the only strong signal, which is the statistical version of what this whole page argues — and if you only have time to read one of the two tables above, read the memory one.
Mechanical Depth is the weakest predictor on the board. How much machinery a platform has — dice, hit points, inventories, combat grids — barely moves whether it survives your novelty window. Friends & Fables and AI Realm are not exceptions to a rule here; they are the rule. If you are picking a platform for a six-month campaign, the ruleset is close to the least informative number on its scorecard, and why a language model struggles to hold one is a separate argument from whether it will keep you.
The Winners, and What They Cost
Voyage has the strongest pair on this page — Memory 4.25 and Longevity 4 — and it arrived on 31 August 2026, after everything below was written. Behind it, two platforms top both axes at 4 and 4: Hidden Door and WyrdTale. They have almost nothing else in common, and they arrive at the same pair of numbers by opposite routes. One more product is above 3 on both — Craft, at 4 and 3.5.
What makes Voyage the interesting one for this page is the price it does not charge. It scores Player Agency — 3.5, where the two 4-and-4 platforms below it both score 1. It still puppets — your character speaks in a cleaned-up version of what you typed and occasionally moves without being told — but it does not railroad, and it does not write your side of a conversation from scratch. That is the smallest bill anyone has paid for this much coherence.
Hidden Door gets there by not being a simulation. Game state lives in cards; stories assemble from a large library of human-authored tropes adapted to your choices. Bound the space of things that can happen, and consistency stops being a losing battle against an expanding context. Stories do not drift. Characters do not forget who they are. The world at turn fifty behaves like the world at turn five.
WyrdTale gets there by moving the game out of the model altogether. It is an MCP connector: the engine runs as a server your own Claude or ChatGPT talks to, holding entities, saves, factions, weather and combat state somewhere the context window cannot lose them. Nothing about continuity depends on the conversation staying short, because continuity was never stored in the conversation.
Now the part that matters, and it is the same sentence twice. Both score Player Agency 1 out of 5.
Hidden Door charges it by restriction: you cannot create at will, you are a co-author inside someone else’s deck of cards, and there is no system underneath to adjudicate anything you invent. WyrdTale charges it by puppeting: it restricts almost nothing and simply keeps writing your character — dialogue you never typed, actions resolved before you chose them. One says no. The other says yes on your behalf.
For most of this category that is the trade, stated openly for once: coherence is bought with freedom. What made it convincing is that these two paid the same price with entirely different architectures.
Then Tabled scored Memory 4 and Player Agency 4, and the rule stopped being a rule — Voyage has since gone further still, taking the top memory score with a Player Agency of 3.5. It keeps campaign state on a server exactly like WyrdTale does, and it does not puppet you — we walked our character away from the party mid-campaign and it simply let us. So the price is not structural. Two platforms chose to charge it; one did not have to.
Craft finishes the job. It scores Memory & Continuity — 4, Longevity — 3.5 and Player Agency — 3.5, which is the closest anything in this directory has come to holding all three at once. Its errors do not compound the way Franz’s do, and a worldbuilder can correct them mid-campaign rather than restarting. It still charges you — in waiting. Turns take long enough to process that we spent about as much of a session watching one resolve as playing it, which is exactly the kind of cost that ends a long campaign quietly rather than dramatically. There is no free coherence anywhere on this page; there are only different currencies.
What Tabled charges instead is worth knowing, because nothing here is free: it scores Determinism & Fairness — 2. The plot it intends to tell proceeds whether you are in it or not, thinner and in parallel, still paying you experience for fights you were not at. So the honest version of the exchange rate is broader and less comfortable than the original: coherence costs something, and the interesting question is which axis a platform decides to charge it to. Know that before you commit six months.
If You Can’t Accept That Trade
Five platforms are worth naming here, and none of them charges you your character to get there.
Voyage has the best numbers of the five — Memory — 4.25, Longevity — 4, Player Agency — 3.5 — and on this page’s question it is now the straightforward answer. What it charges is the narrator tidying your dialogue and occasionally moving your character in small ways you did not ask for, which is the same bill Craft sends and a much smaller one than Hidden Door or WyrdTale. One thing to know before you commit six months: inventory and equipment still misbehave.
Craft is next — Memory — 4, Longevity — 3.5, Player Agency — 3.5 — and it is the pick if you want a campaign that can be repaired rather than restarted, since the worldbuilder-authored memory layer is something you can go in and fix. Budget for the turn times before you commit to a long one.
Tabled is worth a look if memory is your priority and you want a group: it scores Memory — 4, second only to Voyage, and Longevity — 3, and it is 5e-compatible with server-audited dice and real character death. The caveat is not a flaw so much as a design choice — its campaigns are built to reach an ending, so it is the pick for a long finished story rather than an indefinite one, and its own longevity depends on you being willing to follow the story it brought.
Two more sit at 3 on both axes — the balanced middle:
Old Greg’s Tavern is the pick if tone matters. It is the only AI DM we have tested whose voice we would recognise blind, its persistence genuinely works across devices, and invited friends play free on the host’s credits, which makes it the cheapest way to run a group.
Deep Realms matches it on the numbers but gets there differently — and the difference is worth knowing. Its early quality comes from strong models rather than an engine, and that is exactly what caps it. When no system is holding the world and the model alone carries story and memory, Longevity cannot exceed 3 however good the first few hours feel. It is a rule we now apply generally, and Deep Realms is its clearest case.
And if you want the deepest mechanics available and will take the maintenance on yourself, Friends & Fables is still the most complete AI tabletop you can play — just start the summarisation habit in session one rather than the session it breaks.
The Habit That Beats All of Them
The highest-value thing you can do for a long campaign is not choosing the right platform. It is compressing your own history, on whatever platform you chose.
Summarise every twenty or thirty turns, feed the summary back, and you convert a long history into a short one the model can actually hold. This works everywhere, and it helps most exactly where memory is thinnest. Our campaign summarisation guide covers the method, and the free Campaign Memory Tool builds the compression prompt for you — no account, nothing stored, runs entirely in your browser.
A player who summarises on a 2-memory platform will outlast a player who doesn’t on a 4.
How We Measured This
Every number here comes from first-hand play against a fixed probe set, scored on seven published axes and reported to one decimal. Memory & Continuity asks whether the world remembers what happened and for how long. Longevity asks whether it survives past turn fifty and the novelty window. They are deliberately separate, because platforms fail them independently — which is the entire point of this page.
The comparison in “What actually predicts survival” is a straightforward correlation of Longevity against each other axis across all 37 scored products. It describes 37 things we have played rather than predicting the category, and it rests on a field bunched near the floor — a distribution that can move when a few good products arrive. The ceiling claim is the durable one: two axes, no ceiling holder, two years in.
You can sort the full comparison matrix by either axis to see where any platform lands, and the best AI roleplay platforms guide ranks the field by player type rather than by axis.
We do not score our own games, and no platform has paid for a listing or a rating.
Frequently Asked Questions
Which AI RPG has the best memory? Voyage, at 4.25 out of 5 on Memory & Continuity — the only score above 4 we have awarded on that axis across all 37 scored products, platforms and prompt-games together. It held a fact we planted early across an extended campaign and slipped only once or twice over the whole run. Behind it, Hidden Door, WyrdTale, Tabled and Craft all score 4. Hidden Door gets there by assembling stories from authored cards and tropes rather than open-ended generation, which bounds what the system has to keep track of. WyrdTale and Tabled both get there by keeping campaign state in a server outside the model entirely. Craft gets there a fourth way, by letting each world author write its memory structures itself. What separates them is the bill. Hidden Door and WyrdTale each score 1 out of 5 on Player Agency, for opposite reasons: Hidden Door will not let you create at will, and WyrdTale keeps writing your character for you. Tabled scores 4 and will let you walk away from the plot entirely — it just carries on telling that plot without you. Craft scores 3.5 and charges you in waiting instead: its turns take longer to process than anything else we have played.
Is memory the same thing as a campaign lasting? No, and conflating them is why people pick badly. Memory is whether the world recalls what happened. Longevity is whether the platform is still worth playing at turn two hundred. Friends & Fables scores 3.5 on Memory and 2 on Longevity — it remembers well and still runs out of road. Janitor AI is the reverse at 2 and 3. A platform can remember everything and still become boring, or forget constantly and stay fun.
Why does no AI RPG score 5 for memory? Because none of them have earned it. Every platform we have tested is managing a fundamental limitation rather than solving it — a language model has a finite context window, and everything else is scaffolding around that fact. The best implementations delay the failure; none prevent it. A 5 on this axis would mean a campaign that holds together indefinitely without the player doing maintenance, and nothing we have played does that.
Can I fix bad memory myself? Partly, and it is worth doing. Summarising your campaign at intervals and feeding the summary back is the single highest-value habit, because it converts a long history into a short one the model can actually hold. Our campaign summarisation guide covers the method and our free Campaign Memory Tool generates the compression prompt for you. This helps on any platform, but it helps most on the ones whose memory is thin to begin with.
Does paying more get better memory? Not reliably. Craft and Friends & Fables are the most expensive platforms in this comparison, at $199.99 and $39.95 a month at the top, and they score 4 and 3.5 on Memory — good, but not the best here. The best is Voyage at 4.25, whose free tier is a limited single-player demo plus unlimited play on public servers, and Hidden Door matches Craft on a genuinely usable free tier of its own. Higher subscription tiers usually buy a larger context window rather than a better memory architecture, and a larger window delays the problem rather than changing its shape.
What should I pick for a six-month campaign? Voyage — four platforms now score above 3 on both the axes that matter here, and it is the only one of them that does not take your character off you. The other three are Hidden Door and WyrdTale at 4 and 4, and Craft at 4 and 3.5. Hidden Door and WyrdTale both charge the same thing, which is authorship of your own character — Hidden Door if you can accept playing inside authored structure, WyrdTale if you would rather run it on your own Claude or ChatGPT subscription. Craft charges you in waiting instead. If none of those trades appeal, Old Greg’s Tavern or Deep Realms at 3 and 3 are the balanced picks. If you want maximum mechanical depth and will do the maintenance yourself, Friends & Fables — but plan for the summarisation habit from session one, not from the session it breaks. Tabled scores Memory 4, second only to Voyage, and costs you no agency, but it is deliberately not built for an open-ended six months: its campaigns are designed to finish.
Do better mechanics make an AI RPG last longer? Much less than you would expect, and this is the most useful thing in our data for anyone about to subscribe. When we compared Longevity against the other six axes across all 37 scored products, Mechanical Depth came out as the weakest predictor of the lot, well behind Memory and Continuity. Friends and Fables is the clearest case: it holds a perfect score on Mechanical Depth and still scores 2 out of 5 on Longevity, while AI Realm scores 4 on Mechanical Depth and 1 on Longevity. Rules make the hours you get better. They do not by themselves make there be more of them.