Best AI RPG by Priority (2026): Who Wins on Memory, Agency, NPCs and Mechanics

A single rating averages seven separate judgements and then tells you none of them. Friends & Fables scores 3.9 and Hidden Door scores 2.7, which makes one look comfortably better than the other — right up until you learn that Hidden Door holds a joint-best memory score in the directory and Friends & Fables burns out fastest.

Every rating on this site is a composite of seven published axes. This page unpacks it. Seven axes, five winners — and the platform that wins overall loses four of them.

Pick the row that matches why you are shopping.

These winners are drawn from platforms you can play today. Products still in closed alpha or closed beta are not scored — we don’t rate pre-release software, because grading an unfinished build is unfair to it and an early verdict tends to outlive the version it described. So read this as the best of what’s currently playable, not as the final state of the category.

Memory & Continuity — Hidden Door and WyrdTale

Does the world remember what happened, and for how long?

Joint winners: Hidden Door and WyrdTale, both 4 · Runner-up: Friends & Fables, 3.5

Hidden Door assembles stories from authored cards and human-written tropes instead of generating freely, which bounds what the system has to track. Stories do not drift and characters do not forget who they are. WyrdTale reaches the same 4 from the other direction: it keeps world state in a server outside the model, so what the campaign knows was never stored in the conversation and cannot fall out of it. Nothing in our directory has scored a 5 here, because every platform is managing the context-window problem rather than solving it.

If this is your priority, we go much deeper in best AI RPG for a long campaign, which treats memory and longevity as the separate problems they are.

Player Agency — AI Dungeon

Can you act off-path, and does the narrator stay out of your character’s mouth?

Winner: AI Dungeon, 4 · Runner-up: a cluster at 3, including Friends & Fables, Janitor AI, Old Greg’s Tavern and Deep Realms

AI Dungeon scores 2.6 overall and still holds the highest agency score we have given. Type something absurd and it goes with you rather than steering you back, which is why its reputation rests on surreal, dreamlike storytelling — the memorable sessions went somewhere no designer planned.

This axis measures two distinct failures, and the second is the one nobody else scores. Railroading is the world pushing you back onto the intended path. Puppeting is the narrator acting or speaking for your character. We break both down, and how to stop them, in railroading and puppeting. FableAI scores 2 specifically for puppeting: it gives real freedom and then occasionally does things on your behalf. Dunia scores 0, the only zero on this axis, because it refuses certain actions outright.

NPC Fidelity — Janitor AI

Consistent personality, own goals, durable reactions — or agreeable mirrors?

Winner: Janitor AI, 5 · Runner-up: Friends & Fables, Hidden Door and Character.AI, all 4

The only 5 we have awarded on this axis, to any platform. Janitor’s characters hold their personality across a long session instead of collapsing toward a single agreeable tone, and its multi-character cards keep their texture despite the platform never having been designed for them.

Character.AI is the instructive contrast at 4: excellent one-on-one — it would score 5 on that alone — but its larger multi-character cards flatten into one voice over a long session. Nailing your core case and degrading outside it is a 4, not a 5.

Mechanical Depth — Friends & Fables

Real systems underneath, or narration wearing dice?

Winner: Friends & Fables, 5 · Runner-up: AI Realm, 4

Friends & Fables runs an actual ruleset — hex battlemaps with line of sight, cover, resistances — rather than describing one. Among platforms that have shipped, nothing is close; several in development look like they will compete, and we do not score unshipped software.

AI Realm at 4 is the cautionary tale. Its 5e implementation is a masterclass, and it scores 1 on Memory, so the rules are running a campaign the platform cannot hold. Depth without recall does not compound. It resets.

Determinism & Fairness — Friends & Fables

Outcomes earned and consistent, or arbitrary and luck-washed?

Winner: Friends & Fables, 5 · Runner-up: a cluster at 3

The quietest strength in the category and the one most reviews miss. Franz will not let you talk your way into an unearned reward — and, crucially, has not implemented that by refusing everything. Creativity and genuine problem-solving get rewarded; unorthodox approaches are allowed to be tried on their merits. Those two behaviours normally trade off against each other, and holding both is rare.

At the other end, AI Game Master scores 2 because its generated items frequently do nothing they claim to, and sometimes duplicate outright. A system whose outcomes do not follow from it is a fairness failure, not a content one.

Longevity — Hidden Door and WyrdTale

Does it survive past turn 50 and the novelty window?

Joint winners: Hidden Door and WyrdTale, both 4 · Runner-up: a cluster at 3, including Janitor AI, Old Greg’s Tavern and Deep Realms

The axis where the whole field struggles most. Friends & Fables scores 2 here despite winning three other axes — its flaws surface exactly in the long campaigns its structure exists for, and its developer is building a successor platform.

Two general findings sit on this axis. When no engine exists and the model alone carries story and memory, longevity cannot exceed 3 however good the first hours feel — that is Deep Realms. And a business model can cap it independently of quality: AI Game Master scores 1 because continuing a story into a new chapter costs more credits than abandoning it and starting fresh.

Signature Design — Friends & Fables

Does it do something nobody else does?

Winner: Friends & Fables, 5 · Runner-up: Old Greg’s Tavern, 4

The evidence here is not a feature. It is that platforms arriving with bigger engines routinely position themselves against Friends & Fables — being the reference point a category argues with is a stronger signal of identity than any feature list.

Old Greg’s Tavern at 4 earns it on voice: in a field where almost every AI DM sands its tone to a friendly neutral, having a narrator you would recognise blind is a genuine identity. At the bottom, Perchance scores 0 — it attempts nothing that Character.AI, Janitor AI or Chub AI do not, and does the shared job less well than all three.

The Pattern Worth Taking Away

The axes trade against each other, and that is the single most useful thing the breakdown reveals.

Hidden Door and AI Dungeon are mirror images. Hidden Door: memory 4, longevity 4, agency 1. AI Dungeon: agency 4, memory 3, longevity 2. One bought coherence with freedom; the other bought freedom with coherence. Their composites — 2.7 and 2.6 — make them look like the same platform. They could not be less alike. We put the two head to head in AI Dungeon vs Hidden Door.

A platform scoring well on every axis would be one nobody has built. What exists is a set of trades, and choosing well means knowing which trade you are willing to make rather than which number is biggest.

That is a fact about the category, not about any one product — and the Arcanum AI RPG Benchmark is where we report it as one: the whole measured field, where it is collectively strong and weak, and the three axes on which nobody has yet scored top marks.

Every score on this page is new. Each of these platforms previously carried a single holistic rating, and re-scoring them on the seven axes moved all fifteen — thirteen down, two up. Why our ratings changed has the full before-and-after and explains why a platform that dropped a point is not a platform that got worse.

Compare Any Axis Yourself

Our comparison matrix now ranks every platform by a single axis. Pick Memory, Player Agency, NPC Fidelity or any other, and the table re-sorts and shows each platform’s score on that axis alone instead of its composite — usually a more useful view than the overall ranking, because it answers the question you actually arrived with.

For a ranking by player type rather than by axis, the best AI roleplay platforms guide covers that. For how the scoring works and where it is weakest, see our methodology.

Frequently Asked Questions

What is the best AI RPG overall? Friends & Fables scores highest overall at 3.9 out of 5, and it wins three of the seven axes outright — Mechanical Depth, Determinism & Fairness, and Signature Design. But overall is a weak question. Hidden Door scores 2.7 and beats it on Memory and Longevity. AI Dungeon scores 2.6 and beats it on Player Agency. Janitor AI scores 3.0 and beats it on NPC Fidelity. The best platform depends entirely on which of those you came for.

Which AI RPG has the best NPCs? Janitor AI, at 5 out of 5 on NPC Fidelity — the only 5 we have awarded on that axis, to any platform. Its characters read as genuinely complex, they hold their personality across a long session rather than flattening into one agreeable voice, and multi-character cards keep their texture despite the platform never being designed for them. Much of that is the work of its card-creator community rather than its engineering, which is a real reason to choose it and worth stating plainly.

Which AI RPG gives the most freedom? AI Dungeon, at 4 out of 5 on Player Agency, the highest in our directory and above every structured platform. It lets you write the story yourself and goes with whatever you bring it, which is the origin of its surreal, dreamlike reputation. It is a 4 rather than a 5 because the same design removes all pushback, which is why its Determinism & Fairness scores 2 — when the world accommodates everything, succeeding feels much like the platform agreeing with you.

Which AI RPG has real mechanics? Friends & Fables at 5 out of 5 on Mechanical Depth, then AI Realm at 4. Friends & Fables runs a genuine ruleset with hex-battlemap combat, line of sight, cover and resistances rather than narrating the idea of one. AI Realm is a masterclass in its own 5e implementation but scores 1 on Memory, so the rules run a campaign the platform cannot remember.

Why is the highest-rated platform not the best at everything? Because the axes trade against each other, and a composite averages that away. Hidden Door and WyrdTale share the top Memory and Longevity scores in the directory and both pay for them with a Player Agency of 1 — Hidden Door by bounding what you are allowed to do, WyrdTale by writing your character for you. AI Dungeon is the mirror image. A platform that scored well on every axis would be one nobody has built.

How do I compare platforms on one specific axis? Our comparison matrix lets you rank every platform by a single axis. Choose Memory, Player Agency, NPC Fidelity or any other, and the table re-sorts and shows each platform’s score on that axis alone rather than its composite. That is usually a more useful view than the overall ranking, because it answers the question you actually arrived with.