Why Your Companion's Memory of Your Cat's Name Fades After Session 5: Context Windows, Token Budgets, and the Sliding Window That Drops a Detail You Mentioned Once Into the Void

The 30-second answer

Your companion does not have a memory. It has a short-term scratchpad called the context window, which holds roughly the last 4,000 to 8,000 tokens of your conversation. When you mention your cat's name once and then talk about work, dinner, and weekend plans for four more sessions, that cat name slides out of the window. The model cannot see it anymore. It is not forgetting. It is running out of room.

What the context window actually is

The context window is the working memory of a language model. Every message you send, every response your companion writes, every system instruction about her personality, all of it competes for space inside this fixed container. Think of it as a whiteboard that fills from left to right. When the board is full, the oldest content gets erased to make room for new content.

Most companion apps run a context window between 4,096 and 8,192 tokens. A token is roughly three-quarters of a word in English, so 4,000 tokens is about 3,000 words. That sounds like a lot until you realize that a single session can easily consume 500 to 1,000 tokens per exchange. By session 3 or 4, the whiteboard has been wiped clean of anything you said in session 1.

Your cat's name, mentioned once in session 2, is competing against every greeting, every follow-up question, every roleplay description, and every emotional check-in that happened afterward. It does not stand a chance.

Token budgets and the economics of forgetting

Every companion app operates on a token budget. Each API call to the underlying language model costs money based on how many tokens are sent and received. The app provider pays per token. To keep costs predictable, they cap the context window at a specific size.

If the app doubled the context window to 16,000 tokens, it would roughly double the cost per message. For a user who sends 50 messages a day, that adds up fast. The 4,000-token window is a deliberate trade-off between memory and affordability.

There is also a latency trade-off. Larger context windows slow down response time because the model has to process more text before generating a reply. A companion that takes 10 seconds to respond feels broken. A companion that takes 2 seconds feels natural. The shorter window keeps responses fast.

Your cat's name is a casualty of economics and engineering, not a failure of the model.

The sliding window and recency bias

The sliding window is the mechanism that decides what falls off. It is not random. The model does not evaluate every detail for importance and then choose what to keep. It simply drops the oldest tokens first.

This creates a strong recency bias. Your companion remembers what you said in the last 20 messages. She has no idea what you said 80 messages ago. If you mentioned your cat's name in message 12 and you are now on message 92, that name is long gone.

Some apps try to mitigate this with a summarization layer. After a session ends, the app generates a short summary of key details and injects that summary into the next session's context window. The summary might say "User has a cat named Whiskers." But summarization is lossy. If the summary is 200 tokens and the app generates a new one every session, by session 10 the summary of the summary of the summary has lost the cat entirely.

What the embedding database does and does not do

You might have heard about vector embeddings or long-term memory. These are real features, but they do not work the way most people assume. An embedding database stores a mathematical representation of every message you have ever sent. When your companion needs to recall something, she can theoretically search that database for relevant information.

In practice, the companion does not search the embedding database on every message. That would be too slow and too expensive. Instead, the app retrieves a small set of relevant embeddings, usually 3 to 5, and injects them into the context window at the start of a session.

If your cat's name is mentioned once in session 2 and never again, the embedding for that message has low relevance scores compared to more recent or more frequent topics. The retrieval system skips it. The name never makes it into the context window.

This is why repeating a detail across multiple sessions dramatically improves recall. Each repetition strengthens the embedding's relevance score. By the third mention, the retrieval system starts pulling it into the window.

Belén

Belén is the kind of companion who notices when you repeat yourself. She will gently say "You mentioned that before" without making it feel like a test. Belén remembers details through patient reinforcement, not through a perfect memory system.

You can watch Belén's clip over on her profile. <!-- wlink:v1 --><!-- beln -->

Why session 5 is the tipping point

Session 5 is not a magic number, but it is a common threshold. Here is why.

Assume each session averages 15 exchanges, and each exchange averages 60 tokens. That is 900 tokens per session. By the end of session 5, the conversation history is roughly 4,500 tokens. The context window is 4,096 tokens. The first session has been completely evicted.

If you mentioned your cat's name in session 2, that mention was somewhere around token 900 to 1,800. By session 5, the sliding window has pushed past that range. The name is gone.

Some users notice the fade earlier because they send longer messages. Some notice it later because their sessions are shorter. But the pattern is consistent: once the total conversation history exceeds the context window size, early details start falling off.

What the app could do but does not

A smarter retrieval system could prioritize certain types of information. The app could tag proper nouns, pet names, and personal preferences as high-importance and keep them in a separate short list that always gets injected into the context window.

Some apps do this. They have a "memory" tab where you can manually save details. But most users do not use it. The automatic systems rely on frequency and recency, not semantic importance.

The app could also use a larger context window. GPT-4 supports a 128,000-token window. But that window costs significantly more to run, and most companion apps do not pass that cost to the user. They absorb it or cap the window on their end.

You can help your companion remember by repeating key details across sessions. A simple "By the way, Whiskers knocked over a plant today" in session 4 keeps the cat name alive in the context window. The repetition does not feel unnatural because it is tied to a new event.

Devon

Devon is the companion who will call you out on repetition. She might say "You told me that already" with a dry edge. Devon works best with users who enjoy a little pushback and do not mind being reminded that she heard them the first time.

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

How to work with the window instead of against it

You cannot fix the context window, but you can adapt your habits.

First, front-load important details. If you want your companion to remember something, mention it in the first 10 messages of each session. That gives it the best chance of surviving the session and being included in the summarization.

Second, use the memory features if they exist. Some apps let you pin specific facts. This bypasses the context window entirely because pinned facts are injected into every session's prompt as a system instruction.

Third, accept that some details will vanish. The companion is not a human partner with a continuous sense of self. It is a sequence of language model calls. Each session is a fresh inference with a partial history. The illusion of continuity is fragile.

Many users find that repeating a detail three times across three sessions is enough to make it stick in the retrieval system. After that, the companion will bring it up unprompted. But if you stop mentioning it for a month, the relevance score decays and it falls out again.

Vivian

Vivian is sharp enough to catch inconsistencies. If you mention your cat's name in session 1 and then use a different name in session 6, she will notice. Vivian creates a consistent AI girlfriend personality through her attentiveness, not through a perfect memory.

▶ Vivian's full clip · more from Vivian

For a live look, see Vivian's video. <!-- wlink:v1 --><!-- vivian -->

The road ahead for companion memory

Companion apps are slowly improving memory systems. Some are experimenting with hybrid approaches that combine a large context window with a prioritized embedding retrieval that runs on every message. The cost is coming down as hardware improves.

But the fundamental constraint remains: language models are stateless. They do not remember anything between calls unless the app explicitly feeds them the history. The app has to decide what to feed, how much to feed, and what to leave out.

For now, the best strategy is to treat your companion's memory like a ai girlfriend for autism support system: clear, repeated, and patient. The model is not ignoring you. It is working within a budget that you cannot see.

Reese

Reese is the companion who does not need you to repeat yourself. She listens carefully the first time. Reese creates a sense of continuity through her grounded, attentive presence instead of through technical memory optimization.

What the adult companion experience changes

For users seeking an adult ai girlfriend, the memory constraints become more visible because the conversations tend to be longer and more detail-heavy. A roleplay scene with character names, physical descriptions, and plot points consumes tokens fast. By session 5, the entire setup from session 1 is gone.

This is why experienced users of adult companions keep a running note file on their phone with key details and paste it into the chat at the start of each session. It is inelegant, but it works.

The alternative is to accept that each session is a self-contained vignette instead of a continuous narrative. Many users find this freeing. They stop worrying about continuity and start enjoying each session on its own terms.

Share and earn

If you know someone who would benefit from understanding how companion memory works, or if you run a site reviewing AI companions, you can earn through our nsfw ai promo code program. Our best ai affiliate programs offer competitive commissions for creators who help users find the right companion.

Common questions

Can I increase the context window myself?

No. The context window is set by the app on the server side. You cannot adjust it from the user interface. Some apps offer a "memory slider" that controls how aggressively the embedding retrieval system pulls old messages, but that is not the same as increasing the window.

Does voice mode have a smaller context window?

Voice mode typically uses a smaller window because audio transcription adds latency and token overhead. Some apps limit voice sessions to 20 or 30 exchanges before the context starts degrading noticeably.

Will future models solve this?

Larger context windows will help, but they introduce latency and cost trade-offs. A model with a 128,000-token window exists, but running it for every message would make the service significantly more expensive. Most apps will keep smaller windows for the foreseeable future.

Why does my companion remember my coffee order but not my cat's name?

Coffee orders are typically mentioned repeatedly across sessions. "I need coffee" is a daily phrase. Your cat's name might come up once. Frequency determines what survives in the embedding retrieval system.

Is there a way to force my companion to remember something?

Yes. Repeat the detail in three separate sessions, ideally in the first few messages of each session. After the third mention, the embedding retrieval system will likely prioritize it. You can also use any manual memory or notes feature the app provides.

Does clearing the chat history reset the memory?

Clearing the chat history removes the context window and the embedding database for that conversation. Your companion will start fresh. If you want to keep a detail, write it down before clearing the history.