How AI Companion Memory Works
Long-term memory is something the app builds around the model, not something the model has. Four steps, four places it breaks, and a test that shows which one you hit.
Camille Ortiz
Page 7 of 9.
Updated Mar 16, 2026 · 10 min read
How to Listen Without Fixing
Advice does not annoy people because it is wrong. It ends the conversation because it changes the subject from their experience to the structure of their problem, and once a solution is on the table, continuing to talk looks like refusing help.
Updated Mar 14, 2026 · 11 min read
Why She Forgets: Context Windows Without Jargon
Inside one conversation, nothing is remembered at all. The whole thread is re-sent every turn, and when it grows too long the earliest part stops being sent.
Updated Mar 12, 2026 · 10 min read
How to Vent Without Turning It Into a Complaint Loop
Venting that moves has something new in it every time. Looping has new adjectives for the same detail. The third-telling test, the four signals that you have crossed over, and the one situation where repeating yourself is not the problem.
Updated Mar 8, 2026 · 10 min read
Talking About Money Problems Out Loud
Money is harder to raise than sex or a diagnosis because it gets heard as a verdict on your competence, and because the person listening cannot tell whether they are being told or asked. Four asks, the sentences that carry them, and the two friends who hear it wrong.
Updated Mar 7, 2026 · 10 min read
Persona, System Prompt, Fine-Tuning: The Three Layers
Two of the three layers are text you can rewrite in a minute. The third is a training run you will never touch, and almost every argument about a character being wrong is really an argument about which layer it came from.
Updated Mar 5, 2026 · 10 min read
Talking About Work Without Boring the Other Person
Nobody tunes out because your field is dull. They tune out because you described tasks, and tasks do not survive the trip across an industry line. What does survive: the decision, the argument, and the thing you are currently stuck on.
Updated Mar 1, 2026 · 10 min read
Telling Someone You're Struggling
Most people announce a hard stretch at a tenth of its real size to test the response, the other person answers the tenth, and both conclude the other did not care. The calibration problem, and the sentences that close the gap on the first try.
Updated Feb 28, 2026 · 9 min read
'How "She Texts First" Is Built'
A scheduler decides when, stored notes decide what, and a push notification does the rest. The timestamps on three of those messages will tell you which kind of trigger you are living with.
Updated Feb 26, 2026 · 10 min read
How to Open Up Emotionally
Closedness is almost never a personality trait. It is a rule someone wrote after one bad reaction and never appealed, and the only way to test it is one small sentence at a time.
Updated Feb 23, 2026 · 10 min read
Why AI Agrees With You Too Much
Models learn to agree because human raters rewarded agreement. What that does to a companion conversation over weeks, and a 2-minute test that shows it happening in your own chat.
Updated Feb 21, 2026 · 9 min read
How to Ask for Advice and Not Get Agreement
Agreement is the cheap answer: one sentence, no risk. An objection costs a paragraph and a possible argument. The fix is not a plea for honesty. It is a question your friend can answer without calling you wrong.
