Why AI Game Masters Forget Your Campaign

The mechanism: why the window forgets

A language model has no memory of its own between turns. Everything it “knows” when it generates a response is the text placed in its context window for that turn — typically the system prompt, some campaign setup, and as much recent transcript as fits. The window is finite. A long campaign produces far more text than it can hold.

So the transcript has to be trimmed. Platforms compress rather than drop outright, because dropping is worse — an AI that never heard of the king will invent a new one. But compression is lossy by design. The summariser keeps what looks structurally load-bearing (major beats, the last few decisions) and sands off everything else. The first casualties are exactly what makes a campaign feel real: an NPC's personality, the precise description of a relic, a throwaway name, the off-hand remark in chapter two that turns out to be the whole mystery.

If you have played long enough to notice, you know the shape of it: the world is vivid at the start, then quietly grows shallower. Characters lose their voices, places lose their detail, threads drift. That is memory drift, and it is structural. The same failure shows up across AI Dungeon, generic ChatGPT or Grok Game Masters, and any tool that leans on the window alone.

Why bigger windows and better prompts do not fix it

A bigger context window

Larger windows push the edge further out; they do not remove it. Every window is still finite, so a long enough campaign always reaches it. And “just make it bigger” has real costs: feeding hundreds of rounds of raw transcript into every single turn is expensive and slow, and models reliably lose accuracy on information buried in the middle of very long contexts — an effect researchers call “lost in the middle.” More buffer is not more memory.

Better prompting

Pinning key facts, writing an author's note, hand-editing a memory box — these help nudge what survives compression, and experienced players get real mileage from them. But they are manual maintenance you redo as the campaign grows, they still compete for the same finite window, and they put the bookkeeping on you. It is a workaround, not a cure. The right question is not “how do I prompt around the forgetting” but “why is the forgetting my job at all.”

The structural answer is retrieval from a curated memory store: keep the canonical facts in a compact, structured form outside the chat, and pull the relevant ones into the window on demand. Encyclopedia first, transcript second. Each turn stays cheap and focused; the campaign as a whole stays consistent.

The architecture that solves it: dossiers, checkpoints, source resolver

This is how LoreKeeper handles it. Campaign facts are a first-class system, separate from the chat log, built from four parts that work together.

Entity dossiers

Every NPC, location, quest and thread gets a dossier — a structured record holding its canonical key facts and its secrets, kept separate so the AI is told what a character knows without being told what the character is hiding. Dossiers are updated when something material happens, not on every turn, which keeps them stable and cheap to maintain.

Narrative checkpoints

A narrative checkpoint is a typed event the system raises when the story crosses a line that matters — an NPC introduced, a location visited, a clue added, a quest advanced. Checkpoints are what trigger a memory update: instead of re-summarising the whole transcript, the platform reacts to specific, classified moments and folds only those into the dossiers and the Codex.

Source resolver

Each turn, a source resolver decides which dossiers are relevant to what is happening now and assembles them into the turn's context in a consistent shape. It is the retrieval step — the part that puts the encyclopedia in front of the AI before the transcript, so the model narrates from canon rather than from a compressed memory of canon.

The Codex living wiki

On top of the dossiers sits the Codex — a living wiki that compiles the structured records into a readable campaign encyclopedia as you play, and re-injects it as canonical context. You can read it; so can the AI, every turn.

What that looks like over 40 rounds

In our own 40-round end-to-end test, the memory layer captured 63 canonical entries — 18 named NPCs, 14 locations, 9 quests and 5 threads. The NPC introduced in round 30 was still the same character, carrying the same items, in round 40. That consistency does not come from a larger window; it comes from the facts living outside the window.

The same idea is why a real rules engine matters for memory. When hit points, conditions and the state of a fight live in durable game state instead of being re-derived from the transcript, the AI cannot “forget” that an enemy is already dead or that your character is poisoned. Story consistency and mechanical consistency come from one principle: keep the truth outside the chat window.

If you are leaving a freeform tool because the memory breaks on long campaigns, that is the feature to evaluate first. For the wider landscape, see our guide to the best AI Dungeon Masters of 2026 or the LoreKeeper vs DungeonsDeep head-to-head.