Every AI Dungeon Master Forgets: An Honest Map of AI RPGs in 2026

I should get the conflict of interest out of the way first. I'm building an AI RPG engine called Tapestria, which makes me a competitor to most of the tools below, not a neutral reviewer. You should read everything here with that in mind.

That's also exactly why I did this. Before I could say where Tapestria fits, I had to know what already exists and how it performs once real people play it for months, not for the five minutes a demo video shows. So I went through the subreddits, the launch threads, the angry posts, and the quiet ones, for every AI dungeon master I could locate. Around two dozen tools, from the household names down to single-developer projects with empty forums.

The communities disagree about pricing, about which model writes the best prose, about whether any of this should exist at all. But there is one complaint that shows up everywhere, on every platform that has enough players to generate real feedback. The AI forgets.

That single problem turns out to be the most useful lens for the whole category, so I'm going to start there, then walk through the tools themselves.

The one problem every AI RPG has in common

Memory is the unsolved problem of AI roleplay in 2026, and it is very much a present-day complaint.

The pattern repeats with almost funny consistency. NPCs forget who you are or contradict what they told you an hour ago. Inventory and quests vanish. A character you killed shows up alive. Two towns merge into one. Resume a campaign after a week away and the world has quietly rewritten itself.

The receipts are specific. Latitude, the studio behind AI Dungeon, runs its own player surveys, and memory has landed dead last among the things players rate. Everweave players describe the dungeon master developing something close to amnesia over a long campaign. Friends & Fables works fine for a single evening, then degrades over multi-week play to the point that veterans joke their AI game master has gone senile. Voyage was built specifically to fix this, with a persistence layer the studio spent years on, and its own developers concede in the community that the world still contradicts itself.

Here's the part that matters for anyone trying to choose a tool. None of the platforms that handle memory well do it with a smarter language model. They do it with scaffolding bolted on around the model. SillyTavern users maintain lorebooks and run a second small model whose only job is to catch continuity errors. AI Realm players write "GM Guides" and NPC cards by hand. AI Dungeon has a cottage industry of community scripts that patch the gaps. The tools that ask you to do that homework can hold a world together. The tools that promise the AI will just remember cannot.

So when you read a comparison that ranks these tools by prose quality or feature count, it's measuring the wrong thing. The real question is simple. Can it hold a multi-week campaign together without turning you into its unpaid database administrator?

How to tell these tools apart

Strip away the marketing and almost every tool sits somewhere on a single line.

At one end are the freeform engines. The AI reads what you type and continues the story, with no enforced rules and no hard record of the world's state. Maximum creative freedom, and maximum drift. At the other end are the structured engines, where a deterministic layer tracks stats, inventory, and locations and decides outcomes, while the AI only narrates. Less freedom, far more consistency.

The reason this line is the right one to use is that it predicts the failure modes. Freeform tools drift and let you win because an agreeable language model is guessing the numbers. Structured tools fight drift but feel rigid, and their rules engines have their own bugs. Where a tool sits here tells you more than any feature list, so I'll flag it for each one below.

AI Dungeon: the pioneer that still sets the baseline

AI Dungeon, live since 2019, is where this whole category started, and it remains the reference point everyone else is measured against. It sits at the freeform end. There is no ruleset, just open-ended storytelling that goes anywhere.

It has the largest community in the space, and the mood there is a study in coping. The loudest running complaint, beyond memory, is that the AI is too eager to please. It puts words in your mouth, escalates scenes you didn't ask to escalate, and lets a master blacksmith praise a terrible sword because the model has no domain logic to know better. The favorite community joke is the catalog of "AI tang," the tells that give a language model away: every tavern smelling faintly of ozone, every default character named Elara, the same paragraph echoing back dozens of times across a long session.

AI Dungeon vs Voyage: Latitude's structured do-over

This is the comparison people search for most, because Voyage is made by the same studio, and it tells you exactly what AI Dungeon's creators learned.

Voyage, in invite-only beta through 2026, is Latitude conceding in public that pure freeform hits a ceiling. It runs on a "World Engine" that tracks state, gives you real progression, dice luck, and voiced narration. On the architectural line, it moved decisively toward the structured end. That alone is the most important signal in the category. The company that proved freeform AI storytelling could go viral spent years building the structure it originally lacked.

The reception is genuine enthusiasm undercut by friction. Players call it fresh and promising, then run into three walls. The persistence still breaks, with one popular thread cataloguing a hundred-guard castle that rendered empty and NPCs that lose their data when selected. The usage caps bite hard, since a weekly pool means some players get two days of play a week. And the single biggest frustration is something subtle: you cannot edit the AI's text yourself. You have to argue with a separate "Narrator" that often claims it fixed a problem it didn't touch. Plenty of players have gone back to AI Dungeon precisely because it lets them edit freely.

Friends & Fables: the most ambitious TTRPG simulator

Friends & Fables, running since 2021, is the most mature attempt to run actual Dungeons & Dragons 5e through an AI game master named Franz. It sits toward the structured end, with a purpose-built database holding game state instead of relying on the chat log, and a single turn can fan out into many specialized model calls. It supports multiplayer up to six, battle maps, world-building, and voice.

Players love the concept and the worldbuilding, and they are clear-eyed about where it cracks. Franz's rules enforcement is improvised and buggy. Combat skips turns, a 2d6 becomes a 6d6, known spells get reported as unknown. The consensus is that it's a world simulator more than a D&D simulator. And the memory problem hits hard in long campaigns specifically: a three-hour session can accumulate misgendered characters, party members who were never there, and a forgotten boss fight. The community's coping strategy is to treat Franz as a co-dungeon-master you constantly correct, not an authority. Disillusioned long-time subscribers tend to downgrade or wait for the studio's rebuilt successor, which they refer to as Craft.

Everweave and AI Realm: the clearest migration in the space

These two belong together because players walk directly from one to the other.

Everweave is the mobile-first option, and its most-upvoted community post is a structured teardown arguing it's an "AI wrapper" rather than a real D&D simulator. The critique is precise: a language model hallucinates armor class, crams every fight into thirty square feet of imaginary space, ignores ranged attacks, and produces unearned party wipes because it's predicting text instead of running rules. On top of that came the memory failures, an "Empty Bag" glitch where inventory disappears mid-fight, a tendency for every campaign to collapse into the same seed plot regardless of what players do, and a late-April 2026 update that reset attributes and that the community largely hated. Add message quotas that don't roll over, so players burn paid messages arguing with a dungeon master that made the error, and you get an exodus.

The exodus has a named destination, AI Realm, and it is the clearest cross-platform pattern in the space. Departing Everweave players call AI Realm the thing they wanted Everweave to be, and AI Realm even runs a welcome thread for them. AI Realm's pitch is model choice, you pick the language model, and "unlimited messages," which reads as a direct shot at Everweave's quotas. Its weakness is the familiar one. Memory holds only if you do the manual note-taking, and the dungeon master is agreeable enough that it won't let you lose.

Old Greg's Tavern: the multiplayer dark horse

The surprise of the long tail. Old Greg's Tavern has a small but real, organic community, and it was built for the thing most of these tools ignore: playing remotely with friends, up to five or six, with one person able to cover the party. Players find it charming and fast to start, and praise it over raw ChatGPT or AI Dungeon for keeping a plot moving forward.

It breaks in the now-predictable ways. Memory drift merges towns and rewrites companion backstories after a break. Large groups past about ten turns start to feel like parallel solo games rather than one shared session. And the credit pricing draws steady complaints. It's sticky early and fragile over time, which by now you'll recognize as the house pattern.

SillyTavern: not a product, a refuge

SillyTavern is the odd one out, and the most important tool to understand. It isn't a hosted AI dungeon master you buy. It's a free, self-hosted, open-source front end where you bring your own model, and its community, well over a hundred thousand strong, dwarfs every other here.

It matters for two reasons. First, it's the refuge. When people flee Character.AI, JanitorAI, or a hosted tool that's degrading, SillyTavern is the consensus place they land, drawn by total control and uncensored content. Second, it's where the real answer to the memory problem is being worked out in the open. The community runs lorebooks, vectorized memory, and small side models acting as continuity auditors. That tooling is the most advanced approach to persistence anywhere in the space.

The catch is the entry tax. Setting it up is fiddly enough that the friction is the single most cited reason people don't use it, which is why one-click forks that hide the complexity get celebrated when they appear. SillyTavern shows you the ceiling for what's possible and the price you pay to reach it manually.

AI RPGs with real, enforced rules

If what you want is a game that holds you to real mechanics, where a failed roll means you failed and the dungeon master cannot talk its way out of the math, a small group of younger tools is built for exactly that. They make the cleanest version of the argument running through this whole article, that the rules belong in code, where they execute the same way every time, instead of in a language model's good intentions.

RoleForge is a solo-first tool you can play now in a free alpha. It runs a deterministic rules engine where the dice decide the outcome and the AI is kept to narrating what happens. LoreKeeper is further along and heavier on simulation, with a full server-side 5e engine that tracks real spell mechanics and conditions, plus an arena mode for combat. LoreWeaver AI is the most explicit about the idea, with a "Loom Engine" that lets you script rules as visual if-then logic running in real code, so a given rule fires the same way every time instead of drifting.

The tradeoff with all three is maturity. They ask you to accept rougher edges and smaller worlds in return for rules that mean something. If you have ever been let down by an agreeable dungeon master that never lets you lose, that tradeoff is the entire point.

Can ChatGPT or Gemini just be your dungeon master?

This question deserves its own answer, because it's the alternative many players never leave for a branded app. The honest answer in 2026 is that a raw frontier model gets you surprisingly far. A common path looks like this. A player gives up on ChatGPT, which keeps hallucinating and losing details, switches to Gemini 2.5 Pro with a good community prompt, and finds that a solo campaign suddenly holds together.

The ceiling is the same ceiling as everywhere else. A raw model has no enforced rules and no real memory beyond its context window, so it drifts and it lets you win. What a dedicated tool is supposed to add is the scaffolding. The ones that don't add meaningful scaffolding are, fairly, just a wrapper around a model you could prompt yourself.

Can you play an AI RPG in a shared, persistent world?

This one is worth pulling out on its own, because the word multiplayer hides a real difference, and it's the difference I care most about.

Every tool above that offers multiplayer means roughly the same thing by it. A small private party, usually up to five or six friends, shares one instanced adventure. You and your group get your own copy of a story. When the session ends, that copy is yours alone. Nobody else is in it, and nothing you did echoes outward to anyone else. Even at that modest scope the seams show, since multiplayer sentiment in the communities is thin or buggy, and large groups in Old Greg's Tavern slide into parallel solo games inside the same room.

What none of the mainstream tools offer is a persistent shared world. By that I mean one living world that many players inhabit at the same time, where the consequences of one party's actions carry forward into what another party finds later, and the world accumulates its own history instead of resetting each session. That is a categorically harder thing to build, which is most of the reason nobody has. A handful of smaller tools gesture at shared scenarios or group voting, but none runs a true persistent world that players inhabit together.

I'm calling this gap out deliberately, because it's the gap Tapestria is built around.

The comparison, in one table

Approach is where the tool sits on the freeform-to-structured line. Memory is how it fights drift. Multiplayer almost always means a small private party sharing one session, so I've noted where that's what it is. Stage and pricing are as of mid-2026 and change often, so verify before relying on a number.

Platform	Approach	Memory approach	Multiplayer	Pricing model	Stage
AI Dungeon	Freeform	Context work plus community scripts	Party co-op (instanced)	Subscription tiers	Mature
Voyage	Structured	World Engine state, still drifts	Party co-op (early)	Free beta, paid planned	Beta
Friends & Fables	Structured (5e)	Game-state database, fails long-term	Party co-op, up to 6	Free tier plus subscription	Mature
Everweave	Light structure	Weak, story-loop drift	Effectively solo	Non-rolling message quotas	Live
AI Realm	Light structure	Manual notes plus model choice	Party co-op (thin)	Unlimited-message subscription	Live
Old Greg's Tavern	Light structure	Drifts after breaks	Party co-op, up to 5 to 6	Credits plus subscription	Live
SillyTavern	Freeform plus tooling	Best tooling, fully manual	Single-operator	Free, you pay for compute	Mature (self-host)
RoleForge	Deterministic	Structural	Solo (co-op on roadmap)	Free alpha	Alpha
LoreKeeper	Full 5e engine	Structured 5e state	Party co-op, up to 6	Free tier plus subscription	Live
LoreWeaver AI	Deterministic scripting	Authored lore cards	Solo	Tiered, paid caps	Early
Tapestria	Deterministic plus graph world	Persistent state at the core	Persistent shared world	To be announced	In development

Where Tapestria fits, and why I'm building it this way

Here is the part where my conflict of interest becomes the point.

Everything above led me to the same conclusion, and it's the bet Tapestria is built on. The memory problem will not be solved by waiting for a better language model. It is structural, so it needs a structural answer, and this is where Tapestria takes a different road from the rest of the field.

Look at how the other tools with persistence are put together. At the center sits an AI doing the storytelling, and persistence gets wrapped around that center after the fact, a memory layer, a lorebook, a state database bolted on to keep the model from forgetting. The model is the foundation, and the world is a patch applied to keep it honest. This holds even for the tools that handle memory well. Their scaffolding is good work, and it is still scaffolding, propping up something that was never designed to stand on its own.

Tapestria is built the other way around. At the center sits a persistent, consistent record of the world, every location, character, item, and relationship, held as authoritative state. The game is an AI layer wrapped around that core. The language model reads the world and narrates it, and anything it tries to change passes through a deterministic rules engine that decides the outcome, so the dice and the rules are real and the model cannot quietly bend them. Memory is a first-class citizen here, the ground the rest of the game stands on rather than a feature stapled to the side.

This inversion is what makes a shared world possible at all, and that shared world is the part I set out to build first. You cannot drop many players into one living world if that world exists only as an AI's fuzzy recollection of the last few thousand words. The world has to be real, authoritative, and persistent, which is exactly the core Tapestria stands on. Many players move through the same living world at once, and what they do stays in it. A faction one group topples stays toppled for the next group that passes through. A town that burns is burned. The world grows from the sum of what everyone does in it, the way a long-running multiplayer world accretes its own history over years rather than starting fresh each time someone presses play. The graph is what makes that tractable. When the whole world is a structure of entities and relationships rather than a transcript, thousands of threads can run through it at once without the thing dissolving into contradiction. Many threads, one tapestry, which is where the name comes from.

If the direction sounds familiar, it should. It's the same road Latitude took building Voyage, the same instinct behind Friends & Fables' state database and RoleForge's rules engine and that one-line open-source thesis of forcing the AI to narrate and not decide. The whole category is converging on the idea that you separate the storyteller from the bookkeeper. Tapestria takes that a step further, making the bookkeeper the foundation and the storyteller the layer on top, and keeping the books as a graph of entities and relationships, which holds a living, shared world together far better than a flat log or a single table could.

I want to be honest about the stage, since the rest of this article earned its credibility by being honest. Tapestria is in development. It does not have a community of players to quote, because it isn't in their hands yet. So I'm not going to tell you it already beats the tools above. I'm telling you what it's designed to do about the one problem all of them share, and asking you to judge it when it ships, not before.

What I took away from a week of reading

I started this because I wanted a map. What I got was a single, clarifying thread running through every community in the space.

These tools are good at the thing language models are good at, conjuring a scene on demand. They are bad at the thing games require, remembering what happened and meaning it. Every honest player in every community knows this, works around it, and waits for someone to fix it properly. The marketing pages are about prose and features. The forums are about memory.

That's the gap worth building into. It's why I read the communities instead of the landing pages, and it's the bet I'm making with what I build next. If you want to watch that bet play out, in public and with the same honesty I tried to bring here, the rest of the devlog is on the site.

Frequently asked questions

Which AI RPG has the best memory?

None has solved it. Every tool drifts over long campaigns. The ones that hold a world together rely on manual scaffolding like lorebooks, notes, or a deterministic rules engine, rather than a smarter model.

AI Dungeon vs Voyage, which is better?

AI Dungeon is freeform and lets you edit the text freely. Voyage, from the same studio, adds a structured World Engine and persistence, but it locks editing behind a Narrator and caps how much you can play. Voyage is the more structured successor, AI Dungeon the more flexible original.

Can ChatGPT or Gemini be your dungeon master?

Yes, and a frontier model with a good prompt goes a long way, with Gemini 2.5 Pro a common pick. The limit is that a raw model has no enforced rules and no memory beyond its context window, so it drifts and rarely lets you lose.

Is there a persistent, shared multiplayer AI RPG?

The mainstream tools only offer small private party co-op in one session. A persistent world that many players share, where one group's actions carry into what another finds, is still an open gap in the category.

---

Sources for this piece are the public communities of the platforms discussed, read in June 2026, plus the studios' own public surveys and announcements where noted. Player phrasing has been paraphrased throughout. Prices and product stages reflect mid-2026 and should be re-checked before relying on any single figure.

Lucas Quest builds Tapestria, a shared-world AI RPG with a graph-native world model and a deterministic D&D 5e SRD rules engine. This comparison comes from reading the communities behind these tools first-hand, as someone building in the same space.