AI roleplay apps with long-term memory, compared
Updated
The AI roleplay apps with the most developed memory save facts outside the chat and bring them back when they matter. Kindroid and Nomi document layered memory for long conversations with one character. AI Dungeon, Voyage, Friends & Fables, and Hidden Door track the state of an adventure, and NovelAI and SillyTavern let you manage memory by hand. Character.AI added Story Memory and automatic Facts in 2026. Lunelle AI is built around long-term memory for character-driven stories on a phone.
We make Lunelle AI, one of the apps on this page. Every other app is described from its own help center, documentation, blog, or app store listing, and the sources are listed at the end. Prices are in US dollars per month. Memory features change often, so the details here are as of October 29, 2026.
How AI roleplay apps remember
Before each reply, an AI roleplay app sends the model a block of text: the character’s description, any notes, and as much of the recent conversation as fits. That limit is the context window. Once a story outgrows it, the oldest messages drop out, and anything that mattered in them is lost unless the app kept it some other way. Apps handle this in four ways, often combined:
- Notes you write that are sent with every reply, such as a backstory, a memory box, or a pinned summary of the plot.
- Automatic summaries or extracted facts, which compress older events into a few lines.
- Retrieval, where saved memories or lorebook entries are searched and the relevant ones are added to the reply. Some apps match by keyword and others by meaning.
- A larger context window, which fits more of the story at once and delays the problem.
Each method has a weakness. Notes take up room in the context and help only as long as you keep them current. A summary can point to the right moment and still drop the detail you needed. Retrieval keeps the detail but can miss a fact when your question doesn’t match how it was stored. A bigger context window helps less than its size suggests: in “Lost in the Middle”, a 2024 study, models used information at the start and end of a long input much more reliably than information in the middle.
Research on long conversations shows the combined effect. In LoCoMo, a 2024 benchmark of conversations that run up to 35 sessions, models with long context windows or retrieval did better than plain models but still fell well behind people, and questions about time and the order of events were especially hard.
What you can see matters too. When an app shows what it saved, you can fix a wrong memory before it spreads through the story. Our guide to why AI characters forget explains the context window in more detail.
The apps at a glance
| App | Built for | How it remembers | You can see and edit memory | Price per month |
|---|---|---|---|---|
| Lunelle AI | Long, character-driven stories on a phone | Long-term memory of events, choices, and relationships across sessions | Not yet | Free; Unlimited Messages from $14.99 |
| Kindroid | One-on-one companion chat | Backstory and key memories, Learned Context, Cascaded Memory, long-term memory, journals | Yes, except Cascaded Memory | Free tier; Standard $13.99 on the web |
| Nomi | One-on-one companion chat | Short-, medium-, and long-term memory, Shared Notes, Mind Map | Mind Map and Shared Notes | Free; Standard $15.99 |
| Replika | One-on-one companion chat | Memory tab plus patterns from your history | You can add and delete memories | Free; paid plans |
| AI Dungeon | Text adventures | Memory Bank of automatic summaries, Plot Essentials, Story Cards | Plot Essentials and Story Cards | Free; $14.99 to $99.99 |
| Voyage | RPG worlds, solo or with friends | World Engine tracks characters, places, inventory, and relationships | — | Free to start; paid plans |
| Friends & Fables | D&D 5e campaigns | Event memories linked to characters and places, Working Context | Working Context; memories can be forgotten | Free; $19.95 to $39.95 |
| Hidden Door | Guided stories in classic worlds | Narrative world model that retrieves earlier chapters | — | Free to play |
| NovelAI | Co-writing fiction | Memory, Author’s Note, Lorebook | Yes | Free trial; $10 to $25 |
| SillyTavern | Running your own setup | World Info, summaries, vector recall | Yes | Free software; you pay for the model |
| Character.AI | Character chat | Chat Memories, Story Memory, pins, Facts, Lorebook | Yes | Free; c.ai+ $9.99 |
| Janitor AI | Character chat | Chat Memory with optional summaries, Scripts lorebook | Yes | Free; Janitor+ $12.99 |
A dash means the app’s own material doesn’t describe this.
Lunelle AI
Lunelle AI is our app: interactive fiction for iOS and Android, built for long stories. Characters remember events, the choices you made, and how relationships in the story have changed, across sessions, so an adventure picks up where you left it days later. They keep the same personality and facts over long plays, and several characters can share a scene and react to you and to each other.
You can play the curated cast or adventures other players shared, or create your own characters and adventures by describing them to the creator in a chat. AI image generation is built in. Lunelle AI is text only, with no voice calls, no multiplayer, and no web version. It is free to download, with an optional Unlimited Messages subscription from $14.99 a month. Our page on long-term memory describes what it keeps track of.
Pick it if you want long, character-driven stories on your phone, with a recurring cast.
Companion apps with layered memory
Kindroid, Nomi, and Replika are AI companion apps, built for an ongoing relationship with one character more than for stories with a plot. They are on this page because they publish more about their memory systems than most apps do.
Kindroid
Kindroid splits memory into three kinds. Persistent memory is what you write: backstory, key memories, example messages, and directives, sent with every reply. Subscribers also get Learned Context, three short notes the character keeps up to date on its own (growth and relationship, important facts, and ongoing threads), limited to 800, 1,200, and 800 characters. When a note fills up, the least important items are dropped.
Two automatic layers cover older parts of the chat. Cascaded Memory, for subscribers, lets a character recall hundreds or thousands of messages back, with less detail the further back it goes, and you can’t see what it holds. Long-term memory has no size limit, and a memory comes back when the app judges it relevant. Kindroid says this recall is less reliable than Cascaded Memory.
Journals work like a lorebook. Each entry is 20 to 500 characters and is recalled when one of its keyphrases appears in your message, with up to 500 entries per journal. You can tap any reply to see which memories and journal entries were recalled for it, and tell the app to rely on one of them less.
Kindroid added journals for group chats on October 28, 2026. It runs on the web, iOS, and Android. Standard costs $13.99 a month on the web or $15.99 in the app, and the Ultra and MAX add-ons raise the context and memory limits.
Nomi
Nomi describes short-, medium-, and long-term memory layers. In March 2025 it said medium-term recall reached more than 1,000 messages back, and it describes long-term memory as unlimited. Shared Notes hold what you want the character to keep in mind, such as backstory and important facts.
The Mind Map, updated in October 2025, is a view of what a Nomi remembers: people, places, topics, and goals, shown as a graph and a searchable table that you can edit. Each chat and each group chat has its own map, and the first one forms after about 500 messages. A separate Identity Core, where a Nomi refines its own personality, can’t be viewed or edited. Nomi runs on the web, iOS, and Android, and Standard costs $15.99 a month.
Replika
Replika keeps a Memory tab with memories it saved on its own and ones you added. You can add and delete them, though Replika notes that deleting memories can affect what the character recalls. It also draws on patterns from your overall conversation history. The Ultra plan shows which memories were used. Replika doesn’t support group chats.
Adventure and RPG apps
These apps treat memory as part of running a game: tracking what happened, where you are, and which threads are still open.
AI Dungeon
AI Dungeon’s Memory System summarizes earlier actions automatically, stores the summaries in a Memory Bank, and adds the most relevant ones when the whole adventure no longer fits in the context. The bank holds 25 memories on the free plan and up to 800 on the top plan, and when it fills, the least-used memories are dropped. Plot Essentials holds facts you want in every reply, and Story Cards describe characters, places, and items. Memory belongs to one adventure, and a new adventure starts with an empty bank.
Context is 4,000 tokens on the free plan, 8,000 on Journey ($14.99 a month), 16,000 on Legend ($29.99), and 32,000 on Mythic ($49.99). AI Dungeon runs on the web, iOS, and Android.
Voyage
Voyage is Latitude’s RPG platform alongside AI Dungeon. Its World Engine tracks characters’ backstories and motives, locations, history, inventory, health, and relationships, and feeds them back into the story. In April 2026, Latitude said a plan decides how much you can play, not the context length. Since August 2026, Mythic and higher plans get more advanced memory in Voyage’s Unlimited mode. Voyage runs on the web, iOS, and Android.
Friends & Fables
Friends & Fables runs D&D 5e campaigns with an AI game master called Franz. Since an update in June 2025, it saves memories per event and links each one to the characters and locations involved, so returning to a place brings back what happened there. Recent memories are always included, and older ones are searched when they become relevant.
Before each reply, Franz assembles a Working Context from memories, lore, and character sheets, which you can see and edit. The editable blocks in it are capped by plan, from 4,000 characters on the free plan to 12,500 on the top plan. You can’t edit individual memories, but you can tell Franz to forget them. Memory belongs to a campaign. Friends & Fables runs in the browser, with paid plans at $19.95, $29.95, and $39.95 a month.
Hidden Door
Hidden Door runs stories in classic and licensed worlds with a narrator that works like a game master. An October 2026 update changed how each turn is built: the narrator looks up the world, the history of your story, and facts from earlier chapters, and keeps a list of open threads to bring back later. Hidden Door doesn’t describe any way to view or edit what it remembers. Playing is free and runs in the browser.
Apps where you manage memory yourself
NovelAI and SillyTavern give you the most control. Beyond the recent chat, they keep only what you set up, and in return you decide exactly what the AI sees.
NovelAI
NovelAI has a Memory field placed at the top of the context, an Author’s Note for style and direction, and a Lorebook of entries that are added when their keywords come up or are set to always appear. Context settings let you give each part its own token budget. The Opus plan ($25 a month) has a context of 28,672 tokens. Tablet ($10) and Scroll ($15) have smaller ones. NovelAI runs on the web.
SillyTavern
SillyTavern is free software you install yourself and connect to an AI model you pay for separately. It has World Info (lorebooks triggered by keywords), a Summarize extension, and vector storage that searches older messages in the current chat and adds the relevant ones back. Most of these work within one chat, and lorebooks can be reused across chats. It runs on Windows, Mac, Linux, and Android.
Chub and Agnai, two other roleplay platforms, have similar tools: Chub has a chat-level memory summary and lorebooks, and Agnai has keyword-triggered memory books.
Character chat apps
Character chat apps center on a large library of characters made by the community. Their memory tools are mostly per-chat notes, pins, and summaries, with automatic facts on some paid plans.
Character.AI
Character.AI has added memory tools in stages. Chat Memories, from May 2025, is a 400-character field per chat for fixed facts. Since May 2026, every user can write Story Memory and pin important messages, 15 pins on the free plan and 30 on paid plans. With c.ai+ ($9.99 a month), the app also records Facts about your persona and the characters automatically, and you can edit, disable, or copy them into a new chat. The subscription page also lists a Lorebook, with advanced features for c.ai+.
Character.AI filed for Chapter 11 on October 12, 2026, and is being sold. Our guide on backing up your Character.AI chats and characters explains how to save your stories while the sale is pending.
Janitor AI
Janitor AI’s Chat Memory is a block of text sent with every reply. You can write it yourself or have the app summarize the chat since the last update, and since August 2026 you can choose whether a summary replaces the older messages it covers. Scripts work as a lorebook triggered by keywords. Janitor’s own model reads about 8,000 to 9,000 tokens, and Janitor+ ($12.99 a month) gives five times the free context. Janitor runs on the web, and its iOS and Android apps have been in public beta since February 2026.
Talkie, PolyBuzz, and Chai
Talkie and PolyBuzz say in their FAQs that memory capacity is limited and that characters can forget older parts of a conversation. PolyBuzz’s membership page lists “Permanent Memory” and “Long Memory” without explaining how they work. Chai gives character creators a Memory field. None of the three publishes details about how much a character keeps over a long story.
How to test an app’s memory yourself
We found no published study that measures the memory of these apps directly, so the practical way to compare them is to test them yourself. Researchers who test AI memory plant facts early in a long conversation and ask about them much later, using several kinds of questions. LongMemEval, a benchmark presented at ICLR 2025, tests five abilities and found that commercial chat assistants and long-context models lost about 30 percent of their accuracy when information had to be remembered across a long history. You can run a smaller version of the same test in any app, with questions that follow its five abilities:
- Early in a story, plant a few facts in plain words: a name, a promise, an object someone carries, and two events in a clear order.
- Play on for several sessions, or a few hundred messages.
- Ask, in character, about each fact, using different words from the ones you planted. Research on long contexts found that models do worse when a question doesn’t repeat the original wording.
- Ask a question that needs two of the facts at once, such as what the character who made the promise was carrying.
- Ask which of the two events happened first.
- Change a fact partway through, such as a character moving cities, and check that the app uses the new one.
- Ask about something that never happened, and see whether the character says so or invents an answer.
Watch the character’s personality as well. A character who remembers every fact but has drifted into a different voice has a consistency problem, which our guide to keeping AI characters consistent covers.
Common questions
Which AI roleplay app has the best memory?
It depends on the kind of story. For one ongoing relationship with a character, Kindroid and Nomi publish the most detailed memory systems. For adventures, AI Dungeon, Voyage, and Friends & Fables track the state of the game. For full control, NovelAI and SillyTavern let you decide what the AI sees. Lunelle AI is built for long stories with a recurring cast on a phone.
What is the difference between a context window and memory?
The context window is the amount of recent text the AI reads before each reply. Memory features save information from outside that window, as notes, summaries, or searchable entries, and add it back when it is needed. Every app has a context window, and memory is what keeps a story going after it fills up.
Does a bigger context window fix forgetting?
It delays it. A larger window fits more of the story, but research such as “Lost in the Middle” found that models use the middle of a long input less reliably than the start and end. In RULER, a 2024 benchmark of 17 models that all claimed at least 32,000 tokens of context, only about half performed well at that length. Saved memories and summaries are still needed for long stories.
Can I see what an AI character remembers?
In some apps. Kindroid shows which memories were recalled for each reply, Nomi has the Mind Map, Character.AI shows Facts and memory usage, Replika has a Memory tab, and Friends & Fables shows its Working Context. In NovelAI and SillyTavern, you write the memory yourself, so you always see it.
If you’re switching apps, see apps like Character.AI for long stories and AI Dungeon alternatives.
Sources
- Kindroid Help Center, Memory, updated October 28, 2026, and Subscriptions
- Nomi, Major memory update, March 11, 2025, Mind Map 2.0, October 9, 2025, and Identity Core, December 12, 2024; prices from the US App Store listing
- Replika Help Center, How does Replika’s memory work?, What is Replika Ultra?, and Choosing a Subscription
- AI Dungeon Help, The Memory System, Plot Essentials, and Memberships & Benefits
- Latitude, Voyage isn’t AI Dungeon 2.0, April 16, 2026, Introducing Voyage Expanded Beta, April 22, 2026, and Voyage Unlimited, August 12, 2026
- Friends & Fables, Memories V2, June 3, 2025, and Franz 2.0: Working Context, August 28, 2025
- Hidden Door, Inside our agentic story harness, October 2, 2026
- NovelAI documentation, Story Settings, Lorebook, and Subscription
- SillyTavern documentation, Chat Vectorization, Smart Context, and Installation
- Chub documentation, Just Chatting and Lorebooks; Agnai, memory guide
- Character.AI, Helping Characters Remember What Matters Most, May 19, 2025, Smarter Memory for Smarter Chats, May 21, 2026, c.ai+ plans, and An Update on Character.ai, October 12, 2026
- Janitor AI Help Center, Chat Memory & Context Management and FAQ: Subscription; Janitor AI, changelog, August 19, 2026 and The Janitor App is Now in Public Beta, February 7, 2026
- Talkie, FAQ; PolyBuzz, FAQ and Membership; Chai, Web Lite
- Maharana et al., Evaluating Very Long-Term Conversational Memory of LLM Agents (LoCoMo), ACL 2024
- Wu et al., LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory, ICLR 2025
- Liu et al., Lost in the Middle: How Language Models Use Long Contexts, Transactions of the ACL, 2024
- Hsieh et al., RULER: What’s the Real Context Size of Your Long-Context Language Models?, COLM 2024
- Modarressi et al., NoLiMa: Long-Context Evaluation Beyond Literal Matching, ICML 2025
Kindroid, Nomi, Replika, AI Dungeon, Voyage, Friends & Fables, Hidden Door, NovelAI, SillyTavern, Chub, Agnai, Character.AI, Janitor AI, Talkie, PolyBuzz, and Chai are trademarks of their owners. Lunelle AI is not affiliated with any of them.