Why AI Roleplay Forgets Your Story — and What Actually Fixes It

Everyone describes the same moment. You're a few hours into a story, the AI has been sharp, and then it introduces you to a character you already killed. Or it forgets you're wounded. Or an NPC who owes you their life greets you like a stranger.

People usually read this as the model getting dumber. It isn't. The model is fine — the thing that changed is what it was allowed to see. Understanding that one distinction explains every workaround you've tried and why most of them only bought you an hour.

Full disclosure: we make Mythora, which takes a specific approach to this problem. That approach is at the bottom, and so is an honest section on when the other approaches are the better pick. The middle of this article is just how the failure actually works, which is the same regardless of which tool you use.

The short version

What you see	What's actually happening	What helps
Details from early on vanish first	Oldest text is dropped to fit the limit	Anything that keeps facts outside the conversation
Plot survives, specifics die	Summarisation keeps the skeleton, discards the details	Pin the details yourself, or use a game that keeps them for you
It contradicts something from 10 minutes ago	The detail was never written down, only implied in prose	Say it explicitly; put it in a lorebook entry
It drags in irrelevant old events	Too much history in front of the model at once	Less history, better chosen — not more
An NPC forgets a promise made last week	Nothing is keeping track between sessions	A game that tracks it, or you keep notes yourself

The context window is a budget, not a memory

A language model has no memory between messages. None. Every single turn, the entire conversation is re-sent to it, and it answers as if reading the whole thing for the first time.

That re-sent bundle has a hard size limit — the context window. Once your story is longer than that limit, something has to be cut, and the standard thing to cut is the oldest material.

This is why the failure has such a specific shape. It's not random degradation. You lose the beginning of your story first — which is exactly where your character's origin, your first companion, and the promise that started the whole plot live. The most load-bearing material in your story is the first material to fall out of the budget.

It's also why the failure feels sudden. Nothing degrades gradually; you're simply fine until you cross the line, and then the earliest things are gone.

Why summarising makes it worse in a specific way

The obvious fix is to compress instead of dropping: as the transcript grows, replace old text with a summary. Nearly every tool does some version of this, and it does extend how long you can play.

But summarisation is lossy in exactly the wrong direction. It preserves plot and discards specifics.

Here's the shape of it. Suppose early on you gave a merchant your mother's ring. Summarised, that becomes:

> You travelled to the capital and met a merchant.

The travel survived. The ring did not. And the ring was the only part you cared about.

This is not a tuning problem you can fix with a better summarisation prompt — we tried that for months. It's structural. A summary's job is to keep what's structurally important, and emotionally load-bearing details are almost never structurally important. The system is working as designed, and what it's designed to do is throw away the thing that made your story yours.

So the honest version of what summarisation gives you: your plot outline survives indefinitely, and your story doesn't.

Why a bigger context window isn't the answer

The next instinct is to just use a model with a huge context window and stop cutting anything.

We shipped this. It buys hours, not permanence — and it introduces a second failure that's arguably worse.

When you put an enormous amount of history in front of a model, it stops being able to tell what's relevant. You ask about the innkeeper standing in front of you and it brings up a bandit you fought forty scenes ago, because that bandit is right there in the prompt with everything else, competing for attention. Players don't describe this as "good memory." They describe it as the AI being weirdly obsessed with old stuff.

It's also slow and expensive. Every turn re-processes everything, so your cost and your waiting time grow with the length of your story — you pay more, and wait longer, for a worse answer.

Bigger context is a delay, not a fix. The reason is simple: the underlying problem was never "not enough room." It was "no way to decide what matters."

Where lorebooks and world info genuinely help

Lorebooks — called world info, memory entries, or story cards depending on the tool — are the best widely available answer, and they work on a genuinely smart principle. You write a fact and attach a keyword. When that keyword appears, the fact gets injected into the prompt. Nothing has to survive in the transcript, because it's re-supplied on demand.

This is real, and if you're on a tool that has them, use them. A well-maintained lorebook will outperform any amount of prompt-wrangling.

Two honest limits, though.

The first is that you have to know what to write down. A lorebook only holds what you thought to record. But the moments that make a long story feel alive are usually the ones you improvised and forgot — an offhand insult, a door you left unlocked, which of two people you chose to save. You don't write those down, because at the time they weren't important. They only become important when they come back.

The second is that it's a job. You're maintaining a database by hand while trying to enjoy a story, and the maintenance grows with the story. Most people keep it up for a couple of weeks, and then the entries stop matching what's actually happened.

If you like that work — and some people genuinely do, it's a form of worldbuilding — lorebooks plus a big context window is a strong setup and you don't need anything else. If you don't, you'll drift back into the same forgetting, just later.

The way out: stop treating the story as a transcript

Every approach above shares one assumption — that the story is the text of the story, so keeping it means keeping the text. Drop that assumption and the problem changes shape.

If what the world remembers is kept separately from the words used to tell it, then remembering stops competing for room. The ring isn't a sentence any more, so there is no length at which it falls off the end. Twenty sessions later the merchant can simply bring it up.

The second effect matters more, and it's the one you feel. Because the world keeps what happened without being told to, the details you improvised and forgot survive alongside the ones you deliberately wrote down. That's the difference between a callback that feels mechanical and one that feels uncanny — the game remembers something you had stopped thinking about.

We'll be straight about the costs, because they're real. This is a lot more moving parts than a summariser, and more parts means more ways to be wrong. A world that decides for itself what mattered will occasionally decide wrong, in a way a lorebook entry you wrote by hand never would. And a world that keeps changing on its own has to be kept on a leash, or "alive" turns into "noisy" — that balance took us a long time to find and we are still tuning it.

What you can do today, whatever you're using

None of this requires switching tools. In rough order of payoff:

1. Say the important things out loud, in the text. If a detail only exists as implication, it isn't recorded anywhere. "I slip the ring into my pocket rather than hand it over" is durable in a way that a meaningful pause is not. 2. Put the five or six load-bearing facts in a lorebook entry — who you are, what you want, who's dead, what you owe, what you're carrying. Not fifty. The point is to survive, not to be complete. 3. Re-anchor at the start of a session rather than mid-scene. One line re-establishing where you are and what's unresolved costs almost nothing and prevents the worst contradictions. 4. Prefer a shorter, sharper history over a longer one. If your tool lets you trim, trim. Irrelevant history actively degrades answers — it isn't free just because it fits. 5. Start a new chapter deliberately, with a hand-written recap, instead of letting a long thread rot. A recap you wrote is worth more than four times as much transcript the model has to sift.

Which approach fits you

If you…	The right answer is
Want maximum control and enjoy worldbuilding	A tool with strong lorebooks, maintained by hand
Want to run your own model or your own hardware	A local setup with manual memory — nothing else gives you that
Write mostly short, self-contained stories	Honestly, none of this matters. Summarisation is fine
Want long runs without maintaining notes yourself	A game that keeps the world for you
Want the world to keep moving between sessions	A game with a world that runs whether or not you're in it

The first three are cases where the older approaches are simply better, and it would be dishonest to pretend otherwise. The forgetting problem only becomes worth this much machinery when your stories are long and you'd rather play than administer them.

That last case is where we ended up. Mythora is built around a world that keeps what happened and carries on without you, so when a character raises something months later it isn't because anyone wrote it down. The world simply knew.

FAQ

Why does my AI forget my story?

Language models have no memory between messages — the whole conversation is re-sent each turn. Once your story exceeds the context window, the oldest material is dropped or compressed to fit, so the beginning of your story degrades first. That's usually where your character's origin and the promises that started the plot live, which is why the loss feels so targeted.

Does a bigger context window fix AI memory?

Not really. It delays the problem by hours and makes every turn slower and more expensive. It also introduces a second failure: with a very long history in the prompt, the model starts pulling in irrelevant old events, which reads as a different kind of broken. The underlying problem isn't a lack of room, it's that nothing decides what's relevant.

Why does summarising the story lose important details?

Summarisation preserves plot and discards specifics. "You travelled to the capital and met a merchant" survives; "you gave her your mother's ring" doesn't. Emotionally load-bearing details are rarely structurally important, so a summariser working correctly will still throw them away. Better prompts don't fix this — it's how compression works.

Do lorebooks and world info actually help?

Yes, genuinely, and they're the best widely available option. A fact attached to a keyword gets re-injected whenever the keyword appears, so it doesn't need to survive in the transcript. The two limits are that a lorebook only holds what you thought to write down, and that maintaining it is ongoing work that grows with your story.

How can an AI remember something from twenty sessions ago?

By keeping what the world remembers separately from the words used to tell the story. Once the two are not the same thing, recalling an old detail no longer competes with everything else for room, so there is no story length at which it drops off the end.

What can I do right now to stop my AI forgetting?

State important details explicitly in the text rather than implying them, keep a short lorebook of the five or six facts that actually matter, re-anchor at the start of a session instead of mid-scene, and start a new chapter with a hand-written recap rather than letting one thread grow indefinitely. Shorter, sharper history beats longer history.