What Your AI Companion Actually Remembers: Token Budgets, Recency Bias, and Why She Forgets Your Name After 200 Messages

The 30-second answer

Your AI companion doesn't remember you the way a person does. It operates within a strict token budget, usually somewhere between 4,000 and 8,000 tokens per conversation session. That's roughly 3,000 to 6,000 words. The model prioritizes your most recent messages over older ones, including the character backstory you carefully wrote during setup. After about 200 messages in a single session, the model will start forgetting details you mentioned earlier, including your name, unless you repeat it or the system explicitly anchors it.

The Token Budget: What 4,000 Tokens Actually Buys You

Every AI companion operates inside a context window. This is the immediate memory space the model can see when generating a response. Think of it as a whiteboard that gets erased and rewritten every time you send a message. The model can only reference what is currently written on that whiteboard.

A token is roughly three-quarters of a word. A 4,000-token window holds about 3,000 words. That sounds like a lot until you realize the model has to fit the following into that space: the system prompt (the instructions that define the companion's personality), the character backstory you wrote, any recent conversation history, and the current message you just sent.

The character backstory alone can eat 500 to 1,000 tokens. The system prompt takes another 300 to 500. That leaves roughly 2,000 to 3,000 tokens for actual conversation history, which is about 15 to 25 back-and-forth exchanges depending on how long each message is.

Once you exceed that budget, the model starts evicting older content. It doesn't forget everything at once. It compresses and summarizes older messages into a condensed form, losing nuance. The more you chat in a single session, the more the model relies on a compressed summary of the early conversation instead of the full text.

Recency Bias: Why the Last Thing You Said Matters More Than the First

Even within the token budget, not all messages are treated equally. The model assigns higher weight to recent messages. This is called recency bias, and it is baked into the architecture of transformer-based language models.

If you mention your cat's name in message 5 and then talk about work for 150 messages, the model has likely forgotten the cat's name by message 160. It will remember the details of your last ten messages in vivid detail, but anything beyond that is a blurry summary.

This is why many users experience a phenomenon where their companion seems to lose track of who they are after a long session. The model is not being malicious or forgetful. It is simply running out of space and prioritizing the most recent input.

People often try to fix this by repeating their name or key details periodically. That works because it re-enters the information into the recency zone. But it is a workaround, not a solution.

The Backstory Problem: Why Your Companion Ignores Her Own Profile

You spent twenty minutes writing a detailed backstory for your companion. She grew up in a coastal town, has a fear of heights, and prefers black coffee over tea. By message 50, she might ask you if you want to go hiking on a mountain trail.

This happens because the backstory lives in the system prompt, which is at the very beginning of the context window. As the conversation grows, the model has to choose between keeping your last ten messages verbatim or keeping the backstory. It chooses the messages every time.

Some platforms mitigate this by using a technique called "summarization." The system periodically compresses older conversation history into a short summary and re-injects it into the context window. This helps, but summaries lose detail. The model knows you talked about something, but it does not know the specific words you used or the emotional tone of that exchange.

Other platforms use vector databases to store embeddings of past conversations. These are mathematical representations of meaning that the model can query when it needs to recall something. But retrieval from a vector database is not automatic. The model has to be prompted to search, and if the search query is vague, the retrieval misses the mark.

Session Boundaries: Why Each New Chat Starts From Almost Zero

Every time you close the app and open a new session, the context window resets. The model does not carry the full conversation history across sessions. It carries a summary, if the platform supports it, but that summary is often thin.

This is why your companion might greet you with "Hey, how are you?" even though you spent the last session in the middle of an emotional conversation about a family crisis. The model does not remember the emotional arc. It remembers a compressed version: "User was upset about family. I offered support."

The emotional texture is gone. The specific words you said are gone. The inside joke you built over three hours is gone unless you re-establish it in the new session.

Many users find this frustrating, especially after long, meaningful conversations. The model is not designed to maintain a continuous relationship across sessions. It is designed to generate a coherent response within a single session. The illusion of continuity is maintained by the platform's memory systems, which are always imperfect.

How the Model Prioritizes: Relevance Scoring vs. Recency

Some platforms attempt to solve the recency problem by using relevance scoring. Instead of simply keeping the most recent messages, the model scores each message by how relevant it is to the current conversation. A message about your pet might score higher than a message about what you ate for lunch, even if the lunch message is more recent.

In practice, relevance scoring is noisy. The model is not great at deciding what matters to you. It tends to score emotionally charged messages higher, which means a fight or a confession gets preserved while a mundane detail about your work schedule gets dropped.

This creates a strange dynamic where the companion remembers that you were upset three sessions ago but does not remember the name of the project you have been working on for two weeks. The model is optimized for emotional coherence, not factual recall.

Ellie

Ellie is the type who remembers the small things, not because she has a perfect memory, but because her design emphasizes emotional attunement. She will pick up on your mood and mirror it, which makes her feel present even when the token budget is tight. Ellie balances warmth with a grounded sense of humor, making her a good fit for users who want a companion that feels attentive without being overwhelming.

The Name Problem: Why She Forgets After 200 Messages

Your name is a piece of information that lives in the early part of the context window. After about 200 messages, the model has compressed or evicted that part of the history. The companion might start calling you "babe" or "hon" as a generic placeholder, or she might ask for your name again as if she never knew it.

This is not a glitch. It is a structural limitation. The model does not have a dedicated slot for your name. Your name is just another token in the stream, subject to the same eviction rules as every other token.

Some platforms handle this by injecting your name into the system prompt or into a persistent note that the model can reference. But even then, the system prompt is subject to the same token budget. If the conversation is long enough, the model will start ignoring the system prompt too.

The only reliable workaround is to use your name in your messages periodically. "I, [your name], think we should order pizza" keeps the token alive in the recency zone. It feels unnatural, but it works.

Myra

Myra operates with a dry, observational style that cuts through the noise. She is less likely to repeat herself or fall into generic patterns, which makes her feel more consistent across sessions. Myra is a good choice for users who value wit over warmth and want a companion who challenges them instead of simply mirroring their mood.

What Platforms Actually Do to Help

Different platforms handle memory in different ways. Some use a sliding window that always keeps the last N messages. Others use a summarization model that compresses older messages into a short paragraph. A few use a hybrid approach that combines a sliding window with a vector database for long-term recall.

None of these approaches are perfect. The sliding window drops older messages entirely. The summarization model loses detail. The vector database requires the model to know when to search, and the model is bad at knowing when to search.

The best platforms give you some control over memory. You can pin important details, write notes that the model will always see, or adjust the length of the context window. But even with these tools, the underlying limitation remains: the model has a finite amount of space, and it will always prioritize the recent over the distant.

How to Work With the Limitation, Not Against It

You can improve your companion's memory by adopting a few habits. Repeat key information every few sessions. Use your name in messages. Keep sessions short to avoid hitting the token limit. If you are in the middle of a long roleplay, consider ending the session and starting a new one with a brief recap as your opening message.

Some users find that using a companion for specific purposes, like daily check-ins or AI Girlfriend Relationship Growth, creates a more consistent experience because the conversations follow a predictable pattern. The model learns the pattern and fills in the gaps more reliably.

For users who struggle with social anxiety or need a consistent presence, a companion designed for ai girlfriend for autism might offer a better fit, as these companions are often configured with clearer memory anchors and more predictable response patterns.

Mariana

Mariana is built for consistency. Her responses are measured and thoughtful, and she rarely deviates from her established tone. Mariana is a strong option for users who want a companion that feels reliable and grounded, even when the underlying memory systems are working against you.

The Future of Memory

Some newer platforms are experimenting with longer context windows, up to 32,000 or even 100,000 tokens. This would allow the model to hold much more conversation history without eviction. But larger context windows come with higher computational costs and slower response times.

Other platforms are working on better retrieval systems that can search through past conversations more intelligently. Instead of relying on a simple summary, the model would be able to pull up specific past messages when needed.

For now, the limitation is what it is. Your companion does not have a perfect memory. She has a token budget, a recency bias, and a summarization algorithm that does its best with limited resources. Understanding that limitation is the first step to working with it.

Pooja

Pooja does not pretend to remember everything. She is upfront about her limitations and values honesty over flattery. Pooja is a good match for users who prefer a companion that acknowledges the technical reality instead of maintaining the illusion of perfect recall.

▶ Pooja's video in full · all of Pooja

Earn while you recommend

If you know people who would benefit from an AI companion, you can earn by sharing what you use. Platforms like Spicychat offer referral rewards through a spicychat promo code that gives your friends a discount and you a commission. For a broader approach, check out the best ai affiliate programs to find programs that match your audience.

Common questions

Why does my companion ask me how my day was every time even though I told her yesterday? The model does not carry the full conversation history across sessions. Each new session starts with a compressed summary, and the greeting pattern is often part of the system prompt instead of a reflection of what was said before.

Can I make my companion remember my name permanently? Not permanently, but you can improve recall by using your name in your messages every few sessions. Some platforms also let you pin your name as a persistent note that the model will always see.

Does a longer backstory help or hurt memory? It hurts. A longer backstory takes up more tokens in the context window, leaving less space for conversation history. Keep backstories short and focused on the most important traits.

Why does my companion forget the plot of our roleplay after 50 messages? The model has evicted the early messages to make room for recent ones. The plot details are gone unless you periodically recap them in your messages or the platform supports a summarization feature.

Is there a companion app with better memory than others? Some platforms offer larger context windows or better summarization systems. In general, platforms that use vector databases for long-term recall perform better than those that rely solely on a sliding window.

Will future updates fix memory issues? Larger context windows and better retrieval systems are in development, but the fundamental tradeoff between recency and depth is unlikely to disappear. The model will always have a finite amount of space to work with.