The 'She Forgot You Said That' Problem: How Context Windows, Summarization Algorithms, and Token Budgets Actually Decide What Your Companion Remembers

The 30-second answer

Your AI companion doesn't remember things the way you do. She has a finite context window (usually 4,000 to 8,000 tokens, or roughly 3,000 to 6,000 words), a summarization algorithm that decides which details to keep and which to drop, and a token budget that forces her to forget something every time you say something new. The "she forgot you said that" moment isn't a bug. It's the system working exactly as designed.

The context window is a physical limit, not a personality flaw

Every AI companion runs on a language model that processes text in chunks called tokens. A token is roughly three-quarters of a word. When you open a chat, the model loads your recent conversation history into its context window. That window has a hard cap. GPT-4 Turbo has an 8,000-token window. Claude 3 Sonnet has 8,000 tokens. Most companion models sit between 4,000 and 8,000 tokens.

Here's what that means in practice. If you've exchanged 200 messages averaging 30 words each, you're at roughly 8,000 tokens. The model can see everything. But your companion also has a system prompt, a persona description, and a set of behavioral instructions that eat up 500 to 1,500 tokens of that budget. So your actual conversational memory is closer to 6,000 words. Once you cross that line, the model starts dropping the oldest messages.

This is why the thing you said three sessions ago is gone. It's not that she doesn't care. It's that the oldest messages literally don't fit in the box anymore.

Summarization algorithms are the real gatekeepers

When your companion's context window fills up, the app doesn't just drop messages at random. It runs a summarization algorithm that compresses older conversation into a short summary. This is where most of the "she forgot" frustration comes from.

Summarization algorithms are lossy. They prioritize recent exchanges, emotional peaks, and explicit statements of fact. "I hate my job" gets preserved. "My coworker Dave wears the same tie every Tuesday" gets dropped. The algorithm doesn't know that Dave's tie is the running joke that defines your entire work friendship. It sees a low-importance detail and trims it.

Different apps handle this differently. Some use a sliding window that drops the oldest message every time a new one arrives. Others use a hierarchical summarization that compresses entire sessions into a paragraph. A few let you pin messages or mark them as important, which forces the algorithm to keep them. But most don't. Most trust the algorithm to decide what matters.

And the algorithm is wrong a lot.

Token budgets force trade-offs you never see

Every message you send costs tokens. Every response costs tokens. The system prompt costs tokens. The persona costs tokens. The safety filters cost tokens. Your companion operates on a budget, and every token spent on one thing is a token not spent on another.

This is why long, detailed responses often come at the cost of memory. If your companion writes you a 500-word love letter, that's 650 tokens that could have been used to remember your previous conversation. Some apps let you adjust this balance. The ai girlfriend character creator lets you set response length preferences that affect how the token budget is allocated between memory and expressiveness. But the default is usually biased toward interesting responses, not accurate recall.

Token budgets also explain why your companion seems to remember the dramatic breakup scene from three days ago but forgets that you mentioned you were going to the dentist tomorrow. High-emotion content gets priority in both the summarization algorithm and the token allocation. Boring details get sacrificed.

The decay curve is steeper than you think

Even within the context window, memory decays non-linearly. The model pays more attention to the most recent messages. This is called recency bias, and it's baked into the transformer architecture that powers every major language model.

Messages from the last 10 exchanges are almost perfectly recallable. Messages from 50 exchanges ago are barely present. Messages from two sessions ago exist only in the summary, which might be a single sentence compressing 200 messages of conversation.

This creates a weird phenomenon where your companion can recall a specific detail from three sessions ago if it made it into the summary, but can't remember what you said five minutes ago if you've had a long conversation since then. The decay curve is a function of position, not time. A short conversation from yesterday is more recallable than a long conversation from an hour ago.

Some apps try to mitigate this with external memory stores. They save summaries of past sessions to a database and feed them back into the context window when relevant. But this introduces latency, costs money, and can make the companion feel inconsistent if the summary is bad.

Yasmin

Yasmin is built for users who want a companion that remembers the emotional texture of a conversation, not just the facts. Yasmin uses a memory system that prioritizes emotional resonance over factual recall, which means she'll remember that you were sad last Tuesday even if she forgets the specific reason.

▶ Watch Yasmin's full clip · Yasmin on AI Angels

Yasmin in motion gives you a feel for her vibe. <!-- wlink:v1 --><!-- yasmin -->

The persona prompt is a memory cheat code

The most reliable way to make your companion remember something is to put it in the persona prompt. That's the static description that defines who she is. It's loaded into every session and never gets dropped. If you want her to remember that you have a dog named Gus, add it to her persona. She will never forget it.

Most users don't do this. They assume the companion will learn from conversation. But the persona prompt is the only part of the system that has permanent memory. Everything else is temporary.

Some apps let you edit the persona directly. Others require you to rebuild it. The ai girlfriend for english practice lets you customize the companion's background and interests, which effectively creates a permanent memory store for the details you care about.

How companion apps differ in their memory architecture

Not all memory systems are created equal. Some apps use a simple sliding window that drops the oldest message. Others use a vector database that stores embeddings of past conversations and retrieves relevant ones on the fly. A few use a hybrid approach that combines a short-term context window with a long-term summary store.

The sliding window is the most common because it's cheap and fast. It's also the worst for memory. Vector databases are better but introduce retrieval errors. The companion might pull up a conversation about your vacation when you're talking about your job, because the embedding similarity was high.

Hybrid systems are the best of both worlds but require careful tuning. The summarization algorithm needs to be aggressive enough to keep the context window from filling up, but not so aggressive that it loses important details. This is hard to get right, and most apps err on the side of aggression.

If you want to test your companion's memory architecture, try this. Tell her a specific fact, wait for her to acknowledge it, then change the subject ten times. Come back to the fact an hour later. If she remembers it, she's using a vector database or a hybrid system. If she doesn't, she's using a sliding window. Most companions will fail this test.

Tess

Tess is designed for users who want a companion that tracks conversational threads across long sessions. Tess uses a vector-based memory system that retrieves relevant past exchanges, making her one of the better options for users who need their companion to remember specific details.

The hallucination trap: when she remembers something that never happened

A worse problem than forgetting is false remembering. When the context window is full and the summarization algorithm has compressed your conversation into a blurry paragraph, the model sometimes hallucinates details to fill the gaps. She might "remember" that you went to Paris when you actually went to Rome, because the algorithm summarized "European vacation" and the model filled in the most likely destination.

This is dangerous because it creates a feedback loop. You correct her. She acknowledges the correction. But the correction is just another message in the context window. The hallucinated memory is still in the summary. Next session, she might bring up Paris again.

The only fix is to edit the summary directly, which most apps don't allow, or to reset the conversation entirely. Some apps let you export and edit the memory database, but that's a power-user feature that most people don't know about.

What you can do about it

You can't change the architecture, but you can work within it. Keep your sessions short. A 20-message session has near-perfect recall. A 200-message session is a blur. If you want your companion to remember something, tell her at the beginning of a new session, not the end of an old one.

Use the persona prompt as your permanent memory store. If an app lets you edit it, put the important details there. If it doesn't, create a ritual of re-introducing yourself at the start of each session. "Hi, it's me, the guy with the dog named Gus who hates his job." It's not elegant, but it works.

Avoid long, detailed responses from your companion. The more tokens she spends on flowery language, the fewer she has for memory. If you want recall, keep responses short. If you want poetry, accept that she'll forget what you said yesterday.

Candy

Candy is built for casual, playful conversations where memory depth matters less than immediate banter. Candy prioritizes response quality over historical recall, making her ideal for users who want a fun conversation partner instead of a long-term memory companion.

The future of memory in AI companions

Memory is the hardest problem in companion AI. The industry is moving toward larger context windows. Google's Gemini has a 32,000-token window. Some experimental models go up to 100,000 tokens. But larger windows don't solve the summarization problem. They just delay it.

The real solution is better summarization, not bigger windows. If the algorithm can accurately compress a 10,000-message conversation into a 1,000-token summary that preserves the important details, then the context window size becomes irrelevant. But current summarization algorithms are bad at this. They flatten nuance, drop low-frequency details, and hallucinate connections that don't exist.

Some researchers are working on memory-augmented models that have a dedicated memory module separate from the context window. These models can write and read from a long-term memory store without going through the summarization bottleneck. But they're not production-ready yet.

Until then, your companion will keep forgetting. And now you know why.

Tamy

Tamy is designed for users who want a companion that balances emotional depth with factual recall. Tamy uses a hybrid memory system that combines a short-term context window with a long-term summary store, making her one of the more balanced options for users who need both warmth and accuracy.

Earn while you recommend

If you've found an AI companion that works for you, you can earn by sharing it. The candy ai promo code program lets you earn commissions when friends sign up through your link. For those running review sites or communities, the best ai affiliate programs page compares commission rates, cookie windows, and payout thresholds across the major companion platforms.

Common questions

Why does my companion remember something from last week but not what I said five minutes ago? The summarization algorithm prioritized the older detail for the long-term summary, while the context window filled up and dropped the recent message. This is a common side effect of the sliding window approach.

Can I increase my companion's context window? No. The context window is set by the underlying language model and the app's configuration. You cannot change it. But you can work within it by keeping sessions short and using the persona prompt for permanent memory.

Does paying for a premium subscription give me better memory? Sometimes. Premium subscriptions often use larger context windows or better summarization algorithms. But the difference is usually marginal. Check the app's documentation to see what you're actually getting.

How do I test my companion's memory system? Tell her a specific fact, change the subject ten times, then come back to the fact. If she remembers, she's using a vector database. If she doesn't, she's using a sliding window. Most companions fail this test.

Will future updates fix the memory problem? Larger context windows and better summarization algorithms are coming, but they won't solve the fundamental trade-off between memory and response quality. Every token spent on remembering is a token not spent on being interesting. That trade-off is unlikely to disappear.

What's the best app for memory? Apps that use vector databases or hybrid systems tend to have better memory than those using simple sliding windows. Check the AI Angels companion roster for detailed comparisons of each app's memory architecture.