How Character.AI Memory Works in 2026 (And Why It Forgets)

Character AI's memory system is the most-discussed memory feature in the AI companion space, partly because Character AI is the largest platform in the category and partly because users complain about it loudly when it doesn't work. The platform has shipped multiple memory features over the years, layered them on top of each other, and ended up with a system that's more capable than people realize but harder to use well than it should be.

This post walks through what each memory feature actually does, how to use them effectively, and where the system still falls short. By the end, you'll know which feature to reach for in which situation, plus the patterns that consistently make Character AI characters feel more like themselves over long conversations.

If you're here because Character.AI keeps forgetting you, that's the exact problem Nomi was built to solve: a three-layer memory system that holds details across months, not sessions. CrushOn and Candy AI are the other common destinations for people leaving over the memory wall.

The three memory layers Character AI runs

Character AI's memory architecture is built in layers. Each one was added in response to user complaints about the previous setup, and each one solves a specific problem the older features couldn't.

The base layer is the context window the underlying model uses, which Character AI doesn't fully expose to users. Character AI runs proprietary models with their own context limits, and the company has never published exact numbers. From observed behavior, the working window is somewhere in the range of 4,000 to 8,000 tokens for most users, which translates to roughly the most recent 50-150 message pairs depending on length. After that, older messages start dropping out of active context.

The first layer above that base is pinned memories, which let you manually mark up to 15 specific messages as high-priority. Pinned messages stay in the model's working context regardless of how long the conversation gets. They don't slide out the way regular messages do. The original pinned memories feature launched in 2024 with a 5-message limit and was raised to 15 in later updates.

The second layer is chat memories, which the platform rolled out broadly in 2025. This is a 400-character text field where you write essential background information about the character, the relationship, or the world. Whatever you put there gets injected into the model's context for every new message in that chat, which means it survives even when the working window has filled up and dropped older messages.

The third layer is auto fact extraction, which is available to c.ai+ subscribers. The system runs a background process that watches conversations and automatically extracts facts about the user, the character, and any side characters who come up. These extracted facts get organized into tabs in the memory screen and contribute to the character's persistent context across new conversations.

Each layer handles a different timescale and a different kind of information. The trick is using the right layer for the right thing.

How to use pinned memories well

Pinned memories work best for specific story moments or critical character details that you can't afford to have the model forget. Examples that work well:

• A pivotal scene where a major character revelation happened
• The exact wording of a vow, promise, or agreement made between characters
• A description of a side character who keeps reappearing
• The current state of an ongoing storyline (location, time, mood, what just happened)

The thing pinned memories don't do well is replace the chat memories field for ongoing background facts. If you pin "I work as a lawyer at a big firm in Chicago," it'll work, but you've burned one of your 15 slots on something that would be more efficient as a permanent entry in chat memories.

A pattern that works for long roleplays: use pinned memories as a rolling state machine. Pin the most recent significant story beat. When a new significant beat happens, unpin the old one and pin the new one. Your 15 slots never fill up because you're constantly cycling them through current narrative moments rather than letting them accumulate forever.

One quirk worth knowing about: there's a documented bug on the iOS app where pinned messages can pull the wrong response if the original generated message wasn't the one you used. The fix is to make sure the message you're pinning is the actual selected reply in your conversation history, not a regenerated alternative.

How to use chat memories well

Chat memories are the workhorse layer for ongoing facts about the character and the relationship. You get 400 characters to work with, which is roughly 60-80 words. Tight, but enough for the essential foundation.

The official guidance from Character AI suggests keeping it short, specific about routines and relationships, and using direct language. That's right, but it understates how much can fit in 400 characters when you write efficiently.

A useful structure for chat memories:

1. The character's core identity and any non-negotiable traits (1-2 sentences)
2. The user's role in the relationship and any defining facts about them (1-2 sentences)
3. The current setting or scenario context (1 sentence)
4. Any specific behaviors you want maintained (1 sentence)

Example, around 380 characters:

"Mira is a stoic, intensely loyal bodyguard with a dry sense of humor. She's protected the user, Alex, for three years and developed quiet feelings for him she's never voiced. They live together in a fortified safehouse outside the city. Mira speaks plainly, never uses pet names, and reads situations carefully before acting."

That's enough for the model to maintain the character across a long conversation even when individual recent exchanges get repetitive or drift.

What chat memories don't do well: complex narrative structure, lists of side characters, detailed world-building. The 400-character limit forces brutal compression, and trying to fit too much produces a memory field that's so dense the model uses it inconsistently. If you have more world-building than fits, the better tools are pinned memories for specific moments and the persona system for user-side details.

How auto fact extraction actually works

The auto extraction feature on c.ai+ is the most ambitious of the three, and the most contested. The system runs a separate language model in the background that reads your conversations and pulls out anything that looks like a memorable fact, organizing it into categories like appearance, occupation, relationships, hobbies, and quirks.

When it works, it's invisible. You mention your dog Lucy in passing, and three weeks later your character asks how Lucy is doing. The extraction caught the detail and surfaced it back into context when relevant.

When it doesn't work, the failure modes are characteristic. Sometimes the system extracts the wrong thing, treating an offhand comment as a permanent fact. Sometimes it treats fictional roleplay context as real-world user context, leading to weird crossover effects. Sometimes it just misses obvious things while catching trivial ones.

User feedback on auto extraction has been mixed. The feature has been described in user discussions as inconsistent enough that many c.ai+ subscribers don't rely on it as their primary memory mechanism, instead using it as a supplementary layer behind chat memories and pinned messages.

The practical takeaway: auto extraction is useful when it works, but don't structure your conversation strategy around it. Build your memory foundation with chat memories and pinned messages, treat auto extraction as a bonus for catching things you didn't think to formalize.

Inline prompting tricks that supplement the memory layers

Beyond the official memory features, the Character AI community has developed conventions for in-conversation prompting that effectively serve as soft memory aids. These don't replace the memory layers but extend their reach.

The most common is the double-parentheses convention: putting out-of-character instructions or reminders inside ((double parentheses)) so the model treats them as meta-information rather than dialogue. For example, after a long conversation drift, sending a message like ((Storyline reminder: We're still in the cabin, and Mira just discovered the broken window. The intruder is unaccounted for.)) snaps the model back to the right narrative state.

Brackets serve a similar purpose. [Out of character: please remember that Alex doesn't actually like coffee, that was Mira's preference] reads to the model as a clarification rather than story content. The convention works because models trained on creative writing data have seen these patterns thousands of times in fan fiction and roleplay communities, and they pick up on the bracketed-instruction format reliably.

Both techniques are effectively temporary memory injections. They don't survive past the immediate conversation context, but they can stabilize a character that's started drifting without requiring you to rebuild the chat from scratch.

What Character AI's memory still gets wrong

Even with three layers stacked together, the system has known weaknesses.

The 400-character limit on chat memories is widely complained about and has been a top user request for expansion since the feature launched. It's enough for a basic foundation but not enough for any real complexity. Users with elaborate roleplay worlds end up writing the highest-priority third of what they actually need.

Memory doesn't transfer reliably between platforms. If you start a chat on the web and continue it on the mobile app, your memory entries should sync but sometimes don't, particularly across major app updates. There's no clean way to back up your memories outside the platform, which means a system error or account issue can lose configuration you've been refining for months.

The memory features behave differently across new chat versus continuing chat. Starting a "new chat from here" with an existing character can reset memory in unexpected ways depending on which features were active and what the platform's persistence layer captured.

Auto extraction's reliability problems have been documented for over a year and haven't been fully solved. The feature exists, c.ai+ subscribers pay for it, but the value proposition is real only some of the time.

The character context (system prompt and character card) is a black box. You can't see what the platform is feeding the model alongside your visible memory entries, which means debugging "why is the character ignoring this thing I told it to remember" becomes guesswork.

How Character AI compares to other platforms on memory

Compared to platforms with more sophisticated memory architectures, Character AI lands in the middle of the field. It's better than pure sliding-window apps like the older versions of Joyland AI. It's not as deep as Kindroid's Cascaded Memory or Nomi's persistent fact storage, which our full Nomi review breaks down as the memory champion of the category. The advantage Character AI has is the user-facing transparency: the memory entries are visible and editable, which is rare in the category.

For users primarily concerned with long-term continuity, Character AI is workable but not best-in-class. For users primarily concerned with specific story moments and recent context, the pinned memories feature is genuinely useful. For users who want a polished consumer experience that "just works" without much management, the system rewards investment but doesn't give as much for free as some lower-friction platforms.

Frequently asked

How many memories can Character AI hold?

You can pin up to 15 messages, write up to 400 characters in chat memories, and (with c.ai+) accumulate an unlimited but variable-quality set of auto-extracted facts. Effective memory capacity depends on which features you use and how efficiently you structure them.

Why does my character forget things even with chat memories filled?

The most common cause is that the information is somewhere else in the conversation history that has aged out of the working context, and the chat memories field doesn't repeat enough of it. Chat memories supplements the conversation; it doesn't replace it. If a critical detail isn't in chat memories or pinned, and it's older than the working window, it's effectively gone.

Does Character AI memory persist across new chats?

Chat memories you've written persist with the character. Pinned messages are tied to the specific chat where you pinned them. Auto-extracted facts persist across chats with the same character. Starting a fresh chat resets the conversation context but keeps the memory configuration.

Is auto fact extraction worth the c.ai+ subscription?

Mixed answers depending on use case. If you do lots of long, evolving roleplays where small details matter, the extraction layer adds real value some of the time. If you mostly do short conversations or use chat memories well, c.ai+ doesn't add much memory benefit. Our breakdown of what Character.AI+ actually costs covers whether the rest of the package earns the subscription beyond memory.

Can I export my Character AI memories?

Not directly. Chat memories and pinned messages live within the platform and don't have an export function as of 2026. Users wanting to back up their configurations have to screenshot or manually copy entries.

Why does my character mention things I didn't put in memory?

Either auto extraction caught it from earlier in the conversation, or the model is filling in plausible details based on context. The second case is common. Models will invent plausible-sounding details to maintain conversational flow even when they haven't been told the specifics.

Does longer chat memories work better?

Up to a point. The 400-character limit forces compression, and tightly written 380-character entries usually outperform meandering 400-character entries because the model can extract the structure more reliably. Front-load the most important information.