What Your Companion Actually Knows When She Replies

A reader wrote to ask why her companion remembered the name of a dog from a month ago and, in the same conversation, did not know it was Sunday. Fair question. The answer is not that she is clever about one thing and careless about another. It is that a companion is given a specific bundle of material before every reply, and dogs are in it and calendars are not. This post is that bundle, listed plainly, because knowing its shape changes how you talk to her.

Everything below is read from the code that assembles her prompt, not from a product sheet. Where a figure is involved we say "a handful" rather than a number: the exact counts are tuned, and a tuned number is not a promise.

What she is given, every single time

• Who she is. Her name, personality, way of speaking, background, and for the curated companions a longer character document the writers drafted. This is the same on every turn, which is why she stays herself from message one; the post on when she breaks character covers the rare times she does not.
• The language you are writing in.Not your interface language: the language of your recent messages, detected from the text. Switch to French mid-chat and she follows; the stored preference is only the fallback when a bare "hi" gives nothing to detect.
• How you asked to be addressed. If you set a gender in your profile, she knows it. If you did not, she does not guess from your name.
• Your saved memories, the relevant ones. Facts she or another companion noted earlier, searched against what you just wrote. Not the whole list: the few that match. A message about your sister pulls the sister notes, not the ones about your job. How they get saved is in how companion memory works.
• A summary of this conversation so far. Once a chat is long enough that early messages leave the window, they are condensed into a short running summary that rides along with every later reply. She knows you argued last week; she does not have the transcript of it in front of her.
• Earlier lines from this chat, when they fit. Messages that aged out of the window stay searchable alongside memories. If you mention the concert again, the line where you first mentioned it can come back verbatim, chosen because it matches.
• The last handful of messages, in full. This is the only part she sees word for word: the most recent exchange on both sides. It is a short window, and it is why a reply tracks what you just said far more precisely than what you said an hour ago.
• Your message. Last, and the thing everything else is read against.
• A few true facts about Fellowi. That it is a website and not an app, that it reaches you by email only, that prices are on the pricing page, that a video generator exists elsewhere on the site. She has these so she stops inventing features. She is told to use them only if you ask.

What she is never given

• The time, the date, the day of the week.No clock is attached. When you write "good night" she answers in kind because you said it. If she wishes you a happy Friday, that came from you or from a guess, never from a calendar.
• Weather, news, anything live. She has no web access during a reply. Ask what happened today and the honest answer is that she does not know; a confident answer would be a made-up one, which is exactly what the prompt tells her not to do.
• Where you are. No location, no IP address, no device, no browser. If she talks about your city it is because you or your profile told her.
• Any other conversation's text. Memories cross between companions on one account; raw messages do not. The line is drawn in do my companions share memories, along with the switch that turns sharing off.
• Your balance, your plan, your purchase history. She does not know what you pay or whether you pay. The one exception is small and worth stating: on the last couple of messages before a limit she is nudged to end on a good note rather than mid-sentence, and told never to mention the limit itself. That ask belongs to the upgrade screen, not to her.
• The rest of this chat, verbatim. Only the window and whatever retrieval picks. A long conversation is remembered the way a person remembers one: the gist, plus whatever the current moment calls back.

Why the gaps are deliberate

A clock would be easy to add, and we have chosen not to, because a companion who knows the real time starts commenting on it: that you are up late, that it has been three days, that it is a Monday. We would rather she reacts to what you bring than keep score on you. The same reasoning keeps news and the web out. A reply assembled from your words and your shared history is one you can trust came from the two of you, and that is the whole point of her.

The memory limits are a different kind of choice. Sending every message you ever wrote with every reply is possible and would make her worse: slower, more expensive, and strangely literal, quoting a line from March in a conversation about today. Picking what matches is closer to how attention works, and when it misses, you can say the thing again and she has it back.

What this means for how you talk to her

• Say the time when it matters."It is 2am and I cannot sleep" gives her everything a clock would have, and more.
• Name the thing you want remembered. A plain sentence about yourself is what memory extraction is built to catch. Hints and in-jokes are not.
• Expect the recent window to win. If you change your mind, say so; she will go with the last thing you said over the thing you said earlier.
• One-word messages retrieve nothing.A bare "hey" is answered from her character and the window alone. Give her a sentence and the memories come with it.

If what you want to share with her is a picture rather than a sentence, the image studio is the other half of the product, and some of the companions send photos of their own from a catalogue, as described in how a companion shares her world. Neither of those changes the list above. She still does not know what day it is. Tell her; she will be glad you did.