---
title: Can AI remember a long roleplay? How AI story memory works
subtitle: What the model actually reads on every turn, why characters forget, and the four settings that decide how much of your story survives.
description: "How AI story memory works in roleplay apps: what the model reads each turn, why it forgets after about 15 turns on a 4K limit, and what to change."
author: Yumina
date: 2026-11-14
updated: 2026-11-15
category: guides
tags: AI roleplay, AI memory, story memory, interactive fiction
lang: en
url: https://yumina.io/blog/can-ai-remember-a-long-roleplay
canonical: https://yumina.io/blog/can-ai-remember-a-long-roleplay
---

# Can AI remember a long roleplay? How AI story memory works

*What the model actually reads on every turn, why characters forget, and the four settings that decide how much of your story survives.*

Yes, an AI can remember a long roleplay, but only the part the app sends back to the model on each turn. As of November 2026 most apps send between 4,000 and 32,000 tokens, which is 15 to 120 turns. Everything older is gone unless the app does something about it.

## What does the AI read on every turn?

![What the AI reads on every turn: the world, then the conversation](/blog-assets/can-ai-remember-a-long-roleplay/story-memory.svg)

A language model has no memory between messages. It is a function: text in, next words out. So on every turn the app assembles one long text and sends it. On Yumina that text has two parts.

The world comes first: who the AI is playing, what the setting is, the rules of the game, the current numbers. The creator wrote it, it is the same size every turn, and it is never cut.

The conversation comes second: your messages and the AI's replies, newest at the bottom. This part grows by one exchange per turn. Each reply costs a little more the longer you play, so every app puts a ceiling on it.

The ceiling is the story memory. It is measured in tokens, the units a model reads; one token is about three quarters of an English word.

## Why does the AI forget?

When the conversation grows past the ceiling, the oldest exchanges drop off the top. Nothing marks the cut. The AI still answers with confidence, from the text it can see, so a name you gave it in scene one is now a name it has never heard. The character is answering from a transcript with the first pages missing.

At a 4,000-token ceiling, the common free tier, that happens after about 15 turns of 250 words. At 64,000 tokens, Yumina's free plan, after about 190.

## Which settings decide what the AI remembers?

On Yumina the settings live under Settings > AI Config. Other apps expose some of them under different names.

### Story memory

On Yumina, memory starts at 64,000 tokens on the free plan and has no cap with your own API key. That covers the world the creator wrote plus your conversation. How much of it goes to the conversation is a setting you can raise or lower. A longer history is not always better, because it distracts the model from the world and from your latest message, especially on smaller models.

### The recap

When a scene falls out of story memory, a second, smaller model folds it into a short chapter summary that stays in the prompt. The AI keeps the events after it loses the exact words, such as the deal you made in scene three. It is on by default at Standard, with Lean and Detailed settings, and you can turn it off. The first catch-up is held to a fixed size so it never costs more than one normal reply.

### Pinned notes

Things the AI must keep no matter what. Select the sentences in any message and pin them, such as a secret or a promise. Pinned notes sit in the world part of the prompt, so they survive every cut.

### The world's own size

The world is never trimmed, and a world that stores its state as numbers, health 40, trust 2, location "the docks", keeps it in a few dozen tokens instead of pages of prose. That is the other reason state survives better than transcript.

## How can you see what the AI remembers right now?

Open the + menu in the chat and choose Memory. It shows three groups: the messages the AI reads word for word, the ones folded into the recap, and the ones it no longer sees. If a detail you care about sits in the third group, pin it or raise the story memory.

## Does bringing your own key change this?

Yes. With a key from OpenRouter, Anthropic, OpenAI or Google, Yumina applies no cap, and the limit is the model's own window, 128,000 tokens or more on current models. The same holds for a model running on your own computer through Ollama or LM Studio. You pay the provider per token instead of spending mushies, so the only cost of a long memory is the provider's per-token price.

## How much do other AI roleplay apps remember?

Most apps publish the ceiling only on the pricing page, if at all. Checked in October 2026: AI Dungeon gives 4,000 tokens free and sells 8,000, 16,000 and 32,000. Janitor AI gives 8,000 to 9,000 free. Chub AI sells 60,000 on its top tier. Character.AI has never published its number. It counts memory in pins instead, 15 on the free plan and 30 on a paid one. [AI Dungeon alternatives compared by memory and price](/blog/ai-dungeon-alternatives) has the full table.

## FAQ

### Can AI remember a long roleplay?

Only the part the app sends back to the model on each turn. Most apps send 4,000 to 32,000 tokens, which is 15 to 120 turns.

### Why does my AI character forget things?

When the conversation grows past the app's memory limit, the oldest exchanges drop off, and nothing marks the cut. At a 4,000-token limit that happens after about 15 turns of 250 words.

### How much does Yumina remember?

Memory starts at 64,000 tokens on the free plan, about 190 turns, and has no cap with your own key. Scenes older than that are folded into a recap that stays in the prompt.

### Does a longer memory always help?

No. A very long history distracts the model from the world and from your latest message, especially on smaller models.

## Sources

Checked 14 November 2026. Yumina settings and limits are from [Choosing an AI model](https://docs.yumina.io/guide/choosing-an-ai-model) and [Does Yumina remember my story?](https://docs.yumina.io/help/does-yumina-remember-my-story). Turn counts assume 250 words per exchange at 0.75 words per token. Other apps' figures are from their published plans, October 2026. Character.AI's pins are from the plan table in "About (c.ai) lite" on support.character.ai, checked 15 November 2026.
