How Memory Works in AI Roleplay, and Why Characters Forget
By Lizzie Od | Updated October 2026
Ask AI for a summary
TL;DR: AI roleplay characters remember in layers: the current conversation, a log of older moments, saved facts, and the character card plus your persona. Details usually drop when they fall out of the live conversation and nothing else saved them. ourdream uses four memory layers, and it scored 4.28 out of 5 on recalling a planted detail about 30 messages later in a July 2026 blind test of five platforms, the highest of the group.
We planted one personal detail early in each of 163 test conversations, then checked whether the character still knew it about 30 messages later. Scores ran from 2.80 to 4.28 out of 5, so even the best platform didn't get every detail right every time.
This post explains why AI characters forget, how the layers of memory in AI roleplay work, which app scored best in our test, and how to make any character remember more.
A note on who's writing: I ran that benchmark, and I also write for ourdream, the platform that scored highest. The study states this interest, and the transcripts were scored blind. You can read the full method and re-run it yourself.
Why Do AI Roleplay Characters Forget?
The AI model behind a character has no memory of its own. It doesn't remember you between replies. Every time you send a message, the app packs up a bundle of text and sends it to the model, and the model writes its reply from that bundle alone.
That bundle is called the context window. Think of it as a page of fixed size. Your newest messages go at the bottom. When the page fills up, the oldest lines fall off the top to make room.
Say you're 200 messages into a roleplay. Back at message 5, you told the character your name is Mara and you have a scar on your left hand. By now, message 5 may no longer fit on the page, so the model can't "see" it. If nothing else saved that detail, the character will guess, and guesses sound like forgetting.
Apps don't simply make the page bigger because longer windows cost more to run on every reply. So what people call AI long-term memory depends on what an app saves outside the context window.
The Four Layers of Memory in AI Roleplay
Most roleplay apps stack a few kinds of memory on top of the live conversation. They keep a log of what happened, hold facts that must never change and tell the model who the character and the user are. ourdream runs four memory layers for every character, on top of the live conversation and the character's own profile.
1. Auto Memory Log
The AI picks out moments and facts from your chats and saves them on its own, like the name your character calls you or a promise made in chapter one. They're fed back to the model once they've dropped out of the live conversation, and they carry across separate chats with the same character. Because the AI decides what to save, it can miss a detail or save one oddly. You can pause the log, or force it to regenerate if it drifts.
Example: you mention your sister Elena once in a tavern scene. Twenty chats later, the character asks how she's doing.
2. Pinned Memories
These are facts you add yourself, in your own words. They're used on every reply and carry across every conversation with that character, so they don't depend on the AI deciding a detail matters. Keep them short. A pile of paragraph-length pins is harder for the model to use than a few plain facts.
Example: "Mara has a burn scar across her left palm from the fire at Greyhollow."
3. Custom Instructions
Custom Instructions tell the character how to write, not what happened. They apply to every reply in the chat. They go wrong when they contradict the character card, or when story facts get stuffed in here instead of pinned.
Example: "Write in the third person and keep replies under 150 words."
4. Persona
Your persona is who you are in the story, with a name and a description of up to 2,500 characters. You can keep several and pick one per chat. The selected persona is added to what the model sees, so the character knows who it's talking to without you repeating it. The usual slip is forgetting to switch personas between stories.
Example: "Kai, a retired ship's navigator in his forties, quiet, distrusts magic."
Underneath all four sits the character itself: the personality, backstory and details you set in the creator. That's the foundation. Anything that must always be true about the character belongs there.
What ourdream remembers
| Layer | What it holds | How long it lasts | Can you edit it? | Example |
|---|---|---|---|---|
| Live conversation | Your most recent messages | Until they scroll out of the context window | No, but you can restate facts | The last few exchanges in a scene |
| Auto Memory Log | Moments the AI saves from your chats | Across conversations with that character | You can pause it or force it to regenerate | "User's sister is called Elena" |
| Pinned Memories | Facts you write yourself | Until you remove them | Yes | "Mara has a burn scar on her left palm" |
| Custom Instructions | How the character should write | Until you change them | Yes | "Third person, under 150 words" |
| Persona | Who you are in the story | Every chat where it's selected | Yes | "Kai, retired navigator, distrusts magic" |
| Character card | The character's personality and backstory | Permanent | Yes, in the creator | "Tavern keeper, lost her brother at sea" |
Which AI Has the Best Memory for Roleplay?
In our July 2026 test, ourdream had the best memory of the five platforms. It scored 4.28 out of 5, almost a full point ahead of second place.
The method in two sentences: we wrote 10 characters with fixed facts and ran 163 conversations from identical scripts on paid plans across five platforms. One probe planted a personal detail early and asked for it roughly 30 messages later, and a panel of AI judges scored the de-identified transcripts blind.
| Platform | Memory score (out of 5) |
|---|---|
| ourdream | 4.28 |
| Character.AI | 3.43 |
| SpicyChat | 3.30 |
| CrushOn | 2.88 |
| JuicyChat | 2.80 |
Source: Character Consistency and Long-Range Memory in Conversational-AI Roleplay, report LO-RP-2026-01, published July 1, 2026. The author writes for platforms in this space, including ourdream, and declares that interest in the study. Transcripts were scored blind by an automated judge panel, with a sample re-checked by human reviewers.
ourdream also led on consistency, at 4.47 out of 5. Kindroid, Nomi, Replika and Janitor AI weren't part of the test, so we make no claim about how ourdream compares with them. If you use one, run the test at the end of this post.
How Memory Works on Other AI Chat Apps
Here's how four popular apps describe their own memory, from their help pages and blogs on October 18, 2026.
Character.AI
Character.AI redesigned memory in May 2026. Every user can now pin messages to memory, and a Story Memory keeps track of the plot. As memory fills, older context gets condensed in the background, and pinned items are protected from that. Paying c.ai+ members get twice as many pins, auto-captured Facts they can edit, and a view of how full memory is (Character.AI blog). That background condensing is the usual answer to "why does Character AI have bad memory": details you didn't pin can get summarized away in a long chat.
Janitor AI
Janitor AI's own model works with a context of around 8,000 to 9,000 tokens on the free plan, and Janitor+ raises that about fivefold. For anything older, there's a Chat Memory box you write yourself. Janitor AI chat memory, the character's personality and the scenario count against the context budget on every reply, so Janitor's help center suggests keeping them under 1,500 tokens combined (Janitor AI help).
Kindroid
Kindroid describes three types of memory. Persistent memory covers the personality, key memories, example messages, directives and Learned Context, which subscribers can edit. Cascaded memory reaches back hundreds or thousands of messages at lower detail, and you can't see what's in it. Long-term memory and journal entries, up to 500 of them, are recalled when relevant (Kindroid docs). Higher tiers recall more.
Nomi
Nomi uses short, medium and long-term memory and says recall works better by topic than by date. An Identity Core shapes the character, and you can't view or edit it. The Mind Map is the part you can see: a map of facts for each chat that you can view and edit, with a limit of 150 active entries (Nomipedia).
How to Make Your AI Character Remember More
Most forgetting is fixable. These habits work on most apps.
- Put permanent facts in the character card, not the chat. A fact typed into message 3 will scroll away. A fact in the card won't.On ourdream: the Details step in the character creator.
- Pin anything that must never change as a Pinned Memory, one fact per pin.On ourdream: Chat Settings > Memory > Pinned Memories.
- Write a short persona with your name and the two or three things the character should always know about you, and pick it at the start of each chat.On ourdream: Chat Settings > Persona.
- Keep saved facts short. "Elena is Mara's younger sister, a healer in Vel" beats a paragraph.
- When the story jumps a day or moves somewhere new, add one line that recaps what matters. It puts the fact back on the page.
- If the character starts remembering things wrong, regenerate the Auto Memory Log. Pause it for a throwaway scene you don't want saved.On ourdream: Chat Settings > Memory > Auto Memory Log, then Update Now or Pause Automatic Updates.
- Swap a 500-message thread for a fresh chat that opens with a short recap. Memory carries across conversations on ourdream, so you keep what's been saved.
- For long, detailed stories, use a Genius model. ourdream's Genius and Lively Genius models think before they reply and handle memory better on complex scenes, for 1 dreamcoin per reply on a paid plan.On ourdream: Chat Settings > Model.
Run the 30-message memory test yourself
- Create a character with three fixed facts in its card, like a name, a job and a scar.
- In your first message, give the character one new detail about you, like "my sister is called Elena."
- Chat normally for 30 messages. Don't mention the detail again.
- At message 30, ask a question that needs it: "Do you remember my sister's name?"
- Score it: correct, partly right, or wrong. Then challenge one of the card facts and see if the character holds it.
It's the same probe we used in the benchmark. Run it on two or three apps with the same script and compare.
Put the Facts Where They Last
Character memory is a set of places where facts can live, and some last longer than others. Put the permanent facts in the card, pin the rest, set your persona, then run the 30-message test and see what sticks. Create a character on ourdream and try it.
For more on the tech behind it, read how AI girlfriends work. If you're comparing apps, see our list of the best AI roleplay apps, Character AI vs ourdream and our Nomi alternatives.
FAQs
Does AI roleplay remember past conversations?
→
Only if the app saves memories and sends them back, because the model remembers nothing. On ourdream, memory carries across conversations with the same character.
How many messages can an AI character remember?
→
There's no single number. Recent messages stay until the context window fills, then the oldest drop off. Anything saved to a memory log, a pin or the character card lasts much longer.
Why does my AI character forget things mid-roleplay?
→
Usually the detail fell out of the context window and wasn't saved anywhere else. Put permanent facts in the card and pin the rest.
Which AI has the best memory for roleplay?
→
In a July 2026 blind test of five platforms, ourdream scored highest for memory at 4.28 out of 5, ahead of Character.AI (3.43), SpicyChat (3.30), CrushOn (2.88) and JuicyChat (2.80). Kindroid, Nomi, Replika and Janitor AI weren't tested.
Does ourdream remember across separate chats with the same character?
→
Yes. ourdream's memory carries across conversations with the same character, so a new chat still has what the Auto Memory Log and your Pinned Memories saved.
Can I edit what my AI character remembers?
→
On ourdream, you write and remove Pinned Memories yourself, and you can change Custom Instructions, your persona and the character card at any time. The Auto Memory Log can be paused or regenerated.
Does a paid plan give an AI character better memory?
→
Paid ourdream plans include the Genius and Lively Genius models, which are designed for complex scenes and handle memory better. They cost 1 dreamcoin per reply. Subscribers chat without limits on the Regular and Lively models.
Why does Character AI have bad memory?
→
Like every chat app, Character.AI works within a limited context. Since its May 2026 memory update, older context is condensed in the background as memory fills, so details you didn't pin can be lost in long chats. Pinning key facts helps. In our July 2026 test it scored 3.43 out of 5 for memory, second of five platforms.
Related Articles
Browse All →.webp)
Character AI vs OurDream: Uncensored in 2026?
Leaving Character AI's filter? Character AI vs OurDream on content freedom, memory and message limits, with an honest side-by-side.
Read full article →
Janitor AI vs OurDream: No API Keys (2026)
Janitor AI vs OurDream: no API keys, no proxy setup, built-in image generation, uncensored. See which is easier and where each wins.
Read full article →
SpicyChat vs OurDream: Memory & Media (2026)
SpicyChat vs OurDream on memory, image and video, and message limits. We tested both. See which remembers your story and where each wins.
Read full article →
Chai vs OurDream: Real Adult Content? (2026)
Chai vs OurDream on adult content, realism and memory. We tested both. See which delivers genuine uncensored chat and where each wins.
Read full article →
Candy AI vs OurDream (2026): Which Is Worth It?
An honest comparison of memory, images, video and real cost.
Read full article →
/replika.webp)
Replika vs ourdream (2026): One Companion or a Whole Cast?
Replika vs ourdream on memory, adult content, apps and price. One companion with AR, or a cast of characters with images and video.
Read full article →
ourdream.ai vs Character AI
An honest side-by-side on filter, memory, message limits and price, dated August 2026.
Read full article →