Generation (a plain chat)
How it answers
Invents a plausible leader, confidently.
Failure mode
Plausible is not true. Every retelling drifts a little further from your canon, and the drift compounds.
The state layer for AI-run games
Whatever runs your game, Claude as a solo DM, an AI Game Master platform, or you with an assistant at prep time, the campaign only holds together if the world persists between sessions. Grimoire, the campaign manager for TTRPG GMs, is that persistence: typed NPCs, factions, quests, and canon any MCP client reads live.
Free tier includes MCP. We never charge for AI tokens.
An AI D&D campaign is any game where an AI helps run the world: Claude or ChatGPT as a solo DM, an AI Game Master platform, or a human GM prepping with an assistant. Every version shares one failure mode: the AI forgets. The fix is a persistent, structured world state the AI can query instead of re-read.
You know the moment. Session three, and the baron your party buried in session one is suddenly alive and hosting a feast. The anti-magic knight casts a spell. Your homebrew rules quietly regress to Forgotten Realms defaults, because the model stopped remembering your world and started remembering the average of everyone's.
A model generates the next plausible token, and plausible is not the same as true in your world. Context windows end, chats reset, and the summaries that carry a campaign across those resets compress away exactly the detail that mattered.
Uploading your lore helps, until retrieval returns three paragraphs that mention the Compact instead of the one fact you needed. Retrieval returns prose. Grimoire returns typed state.
This page states the thesis; why AI forgets your campaign, the full explainer, proves it mechanism by mechanism.
The architecture question
Same question, three architectures. Watch what each one actually does with it.
The query "Who leads the Tidewrit Compact now?"
How it answers
Invents a plausible leader, confidently.
Failure mode
Plausible is not true. Every retelling drifts a little further from your canon, and the drift compounds.
How it answers
Returns three paragraphs that mention the Compact.
Failure mode
You get prose to re-read, not the fact you needed, and stale uploads answer for the world as it was two arcs ago.
How it answers
Returns the faction record's leader field, plus the relationships that changed it, from the live database.
Failure mode
It only knows what you or your AI recorded. An empty database answers nothing.
Only the third survives session thirty.
Three ways in
Connect over MCP in three clicks. The model opens each session by querying narrative state: open threads, recent recaps, canon facts. It writes the session back as a recap when you finish. Next week the world is still there, in the same state you left it. Free on Grimoire's free tier, free on Claude.ai's free tier: end-to-end free to try.
The platform runs the scene, and the good ones run it well. But none of them is designed to be your permanent world record, and your canon should not live and die inside one vendor's chat history. Keep the world in Grimoire: typed, exportable, yours. Any MCP client can query it, and the campaign outlives whatever you play it on.
The assist-not-replace path: you run the table, the assistant drafts rumors, checks continuity, and answers "what does the duke know" from your records instead of its imagination. That workflow has its own page: AI for Dungeon Masters, not AI as the DM.
Concrete proof
Pick up our campaign where we left off. What's open?
open_threads (3): The ransom note, unanswered 5 sessions quiet
The Tidewrit Compact's empty chair · stakes: leaderless
Vellhollow burns · opened last session
last_recap: S21: the party fired the granary to cover their escape
canon_facts (2): Abbess Maraveth is dead (S21) · the Compact's charter names no successor
Last session you left Vellhollow burning and the Tidewrit Compact leaderless. Three threads are open, and the ransom note has sat unanswered for five sessions. Where do you start?
No pasted recap. No "as we discussed." The model asked the world.
Narrative state is one query of many: the same connection reads entities and their fields, relationships, knowledge graphs, the Campaign Bible, and World Foundations. Everything it reads comes from the typed campaign database underneath.
The difference, row by row
Most systems remember text. Grimoire models the world. Here is what that means across a campaign's life.
| A single chat running your game | The same chat + a persistent state layer | |
|---|---|---|
| Session two | Starts from a summary of a summary | Starts from a narrative-state query: open threads, recent recaps, canon |
| Old canon | Drifts as the context ages out | Read live from typed records on every question |
| Your homebrew rules | Regress toward genre defaults | Pinned in World Foundations and the Campaign Bible |
| Secrets and reveals | Everything is one transcript | Visibility tiers; GM Secrets stay server-side |
| The world, long-term | Lives inside one vendor's chat history | A database you own and export; any MCP client connects |
Rows describe the two architectures as of July 2026, not any specific product; Grimoire is the state layer in the right-hand column, and connecting it takes OAuth, not config files.
A one-shot doesn't need one. Pure improv AI fiction, where contradictions are part of the fun, doesn't need one either; that is a real and legitimate way to play. And if you want a fully packaged play experience with zero setup, an AI Game Master platform will have you rolling faster tonight. The state layer earns its keep when you intend to come back to the same world next week, and the week after that.
Add https://api.ttrpg.bot/mcp as a custom connector in your AI client.
Sign in with OAuth and approve access.
Pick which campaign the AI may see. Scoping is enforced server-side.
The full walkthrough, tool list, and every client's setup path live on the Grimoire MCP page, with the step-by-step in the MCP setup guide.
Grimoire MCP →We never charge for AI tokens. Bring your own client.
FAQ
Yes, and they are good at it for a session or two. The failure is longitudinal: context ends, memory compresses, canon drifts. With a persistent state layer connected over MCP, the model queries your world instead of trying to remember it, and a long campaign holds together.
Three routes, honestly ranked: pasting summaries works for a few sessions and then compounds errors; uploading lore documents helps but retrieval returns prose you re-read rather than facts you query; a typed state layer (what Grimoire is) gives the AI structured records to read and update live, which is the only route that scales past session ten.
No. Grimoire never runs the game, narrates a scene, or generates your world. It is the campaign database underneath whichever AI does: you bring Claude, ChatGPT, or any MCP client, and Grimoire is what that client reads and writes. If you want a packaged AI DM experience, an AI Game Master platform is the faster start; Grimoire is where the world lives so the campaign outlasts any one tool.
What you choose: entities and their fields, relationships, knowledge graphs, open threads, session recaps, the Campaign Bible and World Foundations. Player-scoped access respects visibility tiers, and GM Secrets are enforced server-side.
No. Grimoire's free tier includes MCP, and Claude.ai's free tier supports a custom connector. End-to-end free to try, and we never charge for AI tokens.
Yes. The genre engine reshapes the schema for Sci-Fi, Horror, Post-Apocalyptic, Modern, Superhero, or Historical worlds, so the state layer fits whatever your AI is running.
A snapshot versus a live query. Uploaded docs freeze the world at upload time and answer by similarity search over prose. Grimoire is queried live on every call, so yesterday's session is already in today's answers, as structure rather than text.
Typed entities, knowledge graphs, GM Secrets, and canon that survives the chat ending. Free tier, no credit card, MCP included.