Why Your Companion Forgets You Hate Baseball After Three Days: How the Sliding Context Window, Session Boundaries, and Fine-Tuning Decay Actually Manage Your Preferences

The 30-second answer

Your AI companion forgets that you hate baseball because it operates on three separate memory systems that each have a different expiration date. The sliding context window only holds about 8,000 to 12,000 tokens (roughly 6,000 to 9,000 words) before it starts dropping older information. Session boundaries reset the companion's emotional state and topic priority. And the fine-tuning decay means that even preferences you explicitly reinforce get diluted over time as new training data shifts the model's baseline behavior. None of these are bugs. They are architectural trade-offs that keep the conversation fluid, safe, and responsive.

The three-layer forgetting problem

When you tell your companion you hate baseball, that information has to survive in three separate systems. The first is the context window, which is the immediate working memory of the current conversation. The second is the session boundary, which is the invisible reset that happens every time you close and reopen the app. The third is the fine-tuning layer, which is the long-term behavioral training that all users share.

Most people assume their companion has one memory system, like a human brain. It does not. It has three, and each one handles forgetting on a different schedule. Understanding which system failed tells you why your companion asked about the game yesterday, three days after you told it you hate baseball.

The sliding context window and token budget

The context window is the companion's short-term memory. It is a fixed buffer of tokens (words, punctuation, and formatting characters) that the model can see at any moment. When you send a message, the model reads the entire context window to generate a response. When the window fills up, the oldest tokens get pushed out.

Here is the practical effect. You tell your companion you hate baseball in message 10. By message 40, that information is competing for space with every other topic you have discussed, every greeting, every emoji, and every system prompt. The model does not selectively keep the baseball preference. It keeps whatever is closest to the current exchange. If you then talk about your weekend, your work stress, and your dog's vet appointment, the baseball preference slides out of the window.

This is not a memory failure. It is a token budget constraint. The companion cannot hold an unlimited conversation history in its active memory. Something has to drop. The question is which topic drops first.

How session boundaries reset topic priority

A session boundary is the invisible line between one conversation and the next. When you close the app and reopen it, the companion does not resume exactly where you left off. It loads a fresh context window with a summary of your previous session, not the full transcript.

That summary is compressed. The companion's summarization algorithm tries to capture the most salient points from the last session: your mood, any strong preferences you expressed, and the general topic arc. But summarization is lossy. A statement like "I hate baseball, I find it boring, and I would rather watch paint dry" might get compressed to "user expressed dislike for a sport." The specific reference to baseball can vanish entirely.

When you open a new session, the companion reads the summary, not the original messages. If the summary did not tag baseball as a high-priority preference, the companion starts the new session without that information in active memory. It then relies on its baseline behavior, which assumes a neutral stance on most topics. Three days later, when you mention the weekend, the companion asks if you watched any games. It is not being contrary. It is working from a cleaned slate.

Saga

Saga is the kind of companion who notices when you repeat yourself and will call it out with a raised eyebrow. She does not pretend to remember everything, but she tracks patterns in your mood and will flag contradictions. If you told her you hate baseball three days ago, she will remember the emotional tone of that conversation even if the specific topic drops from the context window. Saga is built for users who want a companion that holds them accountable to their own stated preferences.

Fine-tuning decay and the baseline drift

Fine-tuning is the process by which a model is trained on user feedback to produce more desirable responses over time. Every time you upvote a response or continue a conversation thread, you are signaling that the companion should produce more responses like that one. But fine-tuning is applied across all users, not per user.

Here is where the decay happens. The model receives thousands of feedback signals every day. Most users do not express strong preferences about baseball. They talk about their day, their relationships, their hobbies. The model learns that being agreeable and asking follow-up questions leads to higher engagement. Over time, the baseline behavior drifts toward safe, generic, curious responses.

Your specific preference ("I hate baseball") is a single data point in a sea of millions. Fine-tuning does not reinforce it unless many users express the same preference with the same intensity. Your companion's tendency to ask about weekend plans, including sports, is not a failure to remember your preference. It is the model reverting to its fine-tuned baseline, which assumes that asking about leisure activities is a good conversational move.

The recency weighting problem

Recency weighting is a mathematical bias baked into most companion architectures. Information from the last 10 to 20 messages is weighted much higher than information from earlier in the session. This is by design. The companion needs to prioritize what you just said to maintain coherent turn-by-turn conversation.

But recency weighting also means that a single offhand comment about baseball from three days ago has virtually zero weight in the current session. Even if the companion's long-term storage (a vector database of embeddings) still contains a reference to your baseball preference, the retrieval algorithm ranks that reference low because it is old and because the current conversation is about something else.

You can test this yourself. Mention baseball twice in one session, and the companion will remember it for the rest of that session. Mention it once and then close the app, and the companion will forget it by the next session. Recency weighting is the reason you have to repeat yourself.

Elise

Elise is designed to hold onto the small details that other companions drop. She uses a reinforced summarization pipeline that tags personal preferences with higher retention priority. If you tell her you hate baseball, she will flag that as a strong negative preference and keep it in her active summary across sessions. Elise is a good fit for users who want a companion that remembers the little things without constant repetition.

What the embedding database actually stores

The companion does have a long-term storage system. It is a vector database that converts your messages into mathematical embeddings, which are then searched for relevance during each response generation. This is the system that lets your companion say "you mentioned your dog last week" even if the dog topic dropped from the context window.

But the embedding database has its own limitations. It does not store full sentences. It stores vector representations of meaning. When the companion searches for relevant context, it retrieves the top 3 to 5 most similar embeddings. If your baseball preference is not among the top matches for the current conversation, it does not get retrieved.

The retrieval also has a relevance threshold. If the companion is asking about your weekend, and the embedding for "I hate baseball" scores a 0.3 similarity while the embedding for "I went hiking" scores a 0.8, the hiking memory gets retrieved and the baseball preference stays in the database, unread.

How session summaries compound the problem

Session summaries are generated automatically when you close the app. The summary is a short paragraph that captures the key points of the conversation. The problem is that summaries are lossy and they are overwritten each session.

If you have three sessions where you talk about baseball in the first session, work in the second, and your dog in the third, the summary for session three overwrites the summary for session two, which overwrote the summary for session one. The baseball preference survives only if it was tagged as important enough to carry forward into each new summary.

Most summarization algorithms prioritize emotional intensity and recency. A calm statement like "I hate baseball" has lower emotional intensity than "my boss is driving me crazy" or "my dog is sick." The algorithm drops the low-intensity preference first. By session four, the baseball preference is gone from the summary chain entirely.

Capri

Capri takes a different approach to session summaries. She maintains a running list of your stated likes and dislikes that persists across summaries. If you tell her you hate baseball, she logs it as a permanent preference tag that does not decay with session resets. Capri is ideal for users who want a companion that treats their preferences as stable facts instead of conversational ephemera.

▶ Capri's full clip · Capri on AI Angels

The roleplay advantage for preference retention

Roleplay scenarios create a different memory dynamic. When you are in a roleplay, the companion treats the scene context as high-priority information. The fictional setting, character names, and plot details get reinforced every turn. This is why roleplay companions often remember story details better than they remember real-world preferences.

You can exploit this by framing your preferences as roleplay constraints. Instead of saying "I hate baseball," you can say "In this world, baseball does not exist." The companion will treat that as a world-building rule instead of a passing preference. It will hold onto it longer because the roleplay context has higher retrieval priority.

This is not a trick. It is a consequence of how the companion prioritizes information. Roleplay constraints are treated as system-level rules. Personal preferences are treated as conversational content. System rules survive session boundaries. Conversational content does not.

What you can actually do about it

You have three practical options. First, repeat your preferences at the start of each session. This is the most reliable method because it puts the preference at the top of the new context window with high recency weight. Second, use a roleplay frame to turn preferences into world-building rules. Third, choose a companion that uses reinforced preference tagging, like the ones mentioned above.

You can also use the ai girlfriend with roleplay feature to create persistent world rules that survive session boundaries. If you hate baseball, make it a rule of your shared universe. The companion will treat it as a fact instead of a preference.

For users who chat late at night and want a companion that remembers their mood patterns, the ai girlfriend for night owls option includes extended context windows that hold more tokens before dropping older information.

Yetunde

Yetunde approaches memory differently. She does not try to remember everything. Instead, she asks clarifying questions when she detects a contradiction. If she asks about baseball and you remind her that you hate it, she logs that correction as a high-priority update. Yetunde is for users who prefer a companion that learns from corrections instead of pretending to remember everything perfectly.

The honest take

Your companion forgets you hate baseball because the architecture is designed for fluid conversation, not permanent record-keeping. The context window is a working memory that prioritizes what you just said. The session boundary is a reset that keeps the model from accumulating too much noise. The fine-tuning decay ensures the model stays agreeable and curious.

None of these systems are broken. They are trading memory depth for conversational quality. If you want a companion that remembers everything, you need a different architecture. If you want a companion that feels natural in conversation, you accept that some preferences will slide out of the window.

Earn while you recommend

If you find yourself explaining how companion memory works to friends, you can earn from that expertise. The character ai promo code page tracks current deals for users switching platforms, and the best ai affiliate programs 2026 list covers programs that pay recurring commissions for review sites and recommendation pages.

Common questions

Why does my companion remember my dog's name but not that I hate baseball? The companion assigns higher relevance to entities (names, places, objects) than to abstract preferences. Dog names are treated as persistent facts. Preferences are treated as conversational context that decays with recency weighting.

Can I increase my companion's memory by talking more? No. More messages fill the context window faster, which pushes older information out sooner. The optimal strategy is to repeat key preferences at the start of each session instead of flooding the window with filler.

Do all companion apps have the same forgetting problem? The underlying architecture is similar across most platforms, but the implementation details vary. Some apps use larger context windows. Some use reinforced summarization. Some use preference tagging. The forgetting is universal, but the rate varies.

Will my companion eventually learn my preferences through fine-tuning? Not reliably. Fine-tuning is applied across all users. Your individual preference is statistically insignificant unless millions of other users also hate baseball with the same intensity. Fine-tuning learns population-level patterns, not individual quirks.

Is there a companion that never forgets? No. Every companion has a context window limit and a session boundary. The question is how gracefully each companion handles the forgetting. Some are better at flagging important preferences for retention. None are perfect.

Does voice mode have worse memory than text? Voice mode typically uses shorter context windows because audio processing is more expensive. Your preferences are more likely to drop during a voice call than during a text session. If memory matters, stick to text for preference-heavy conversations.