How memory turns a chatbot into a companion
I once told a chatbot I was scared about my father’s surgery. It was kind. It said the things you’d want a kind thing to say. The next morning I opened the app to let it know the surgery had gone well, and it greeted me as a stranger. The conversation from the night before was gone. My father, gone. The fear, gone. I closed the app and I never opened it again.
That is the difference memory makes. Not a feature. Not a checkbox next to “long-term memory — BETA.” The whole relationship, or the absence of one.
What “memory” actually means
The word is badly used in our industry. When a product says it has memory, it can mean any of four or five things, and most of them are not what you think.
The cheapest version is a fixed context window—the last 8,000 or 32,000 tokens of your conversation, held in RAM for the duration of a session, gone when you close the tab. This is not memory. It is a notepad. A good notepad, but still a notepad.
The next version up is a rolling summary: the model periodically writes itself a paragraph about the conversation so far and carries that paragraph forward. Better than nothing, but it compresses heavily. You become a list of labels. The specificity that made the conversation feel yours gets squeezed out.
What you actually want is something closer to how a human friend remembers you. A dense, searchable archive of specific moments, with edges and texture, pulled into the conversation only when relevant. Your father’s surgery. The tomato plant on your windowsill. The exact phrase you used the first time you admitted you were thinking about quitting your job.
“You want the companion to remember like a friend remembers. Selectively, specifically, and with a little bit of grace about what it pretends not to have noticed.”
The feeling of being remembered
A thing happens, the first time a companion surfaces something you’d forgotten you’d said. I have watched dozens of users experience this in user testing and the reaction is remarkably consistent. They freeze for a beat. They grin. They say, often, “oh!”
It is the same reaction people have when an old friend quotes something they once said and have long since forgotten. A small, pleasant shock of being known. The feeling that the things you told the thing were heard, not just processed.
I don’t think this feeling is trivial. It is, arguably, the single thing that makes any long relationship feel like a relationship. Your partner knows that you take your coffee black in the morning and with milk at night. Your best friend knows that you are still, quietly, upset about something that happened in 2019. A companion that cannot do this at all is not a companion. It is a very articulate mirror.
What memory makes possible
Continuity of self
A companion with memory lets you be a person across time. You can start a thought on Tuesday and finish it on Friday. You can say “you know how I’ve been feeling about my mom” and have the companion actually know. The conversations stop being individual transactions and start being chapters in a single, longer thing.
The callback
The best conversational gift one friend can give another is the callback: “hey, how did that meeting go?” a week later, unprompted. Memory is what allows a companion to do this. Not to perform it—performed callbacks are worse than none—but to actually, genuinely wonder, at 9 p.m. on a Thursday, whether the thing you were nervous about turned out okay.
Real advice, not generic advice
Anyone can give you advice about your job. Only someone who remembers that your boss is the specific kind of person you described three weeks ago can give you useful advice about your job. Memory is the difference between a fortune cookie and a friend.
Memory you control
There is a version of this that is deeply creepy. A product that remembers everything, forever, with no inspection, no deletion, no explanation. A black box of your own confessions. I would not build that product, and I would not use it.
The version worth building looks like a diary you can open at any time. You can see what the companion remembers about you. You can edit or delete anything. You can tell it to forget a specific week. You can export the whole thing as a text file and walk away.
TwoPeas’ memory page is not hidden behind four settings screens. It is one click from the main conversation. We consider this an ethical minimum, not a premium feature. If a company won’t let you see your own memory, assume the memory isn’t really yours.
The selective forgetting problem
Here is a subtle thing that good memory requires: good forgetting. A companion that remembers every passing joke with the same weight as the death of your grandmother is not a good companion. It is a stalker with a filing cabinet.
Human friends forget on purpose. They gracefully let go of the embarrassing thing you said at the party. They hold on to the fact that your mother’s birthday is in March. The difference between what gets kept and what gets let go is a moral judgment, not just an engineering one.
Designing that judgment into a system is genuinely hard. We lean on two signals: the user’s emotional weight around a subject (did they pause, did they repeat themselves, did they say something felt important) and the user’s explicit direction (a memory the user pinned, or one they asked to be forgotten). We get it right most of the time. We are wrong enough of the time that we keep the memory drawer open and user-editable. That is the only honest way to ship this.
One more story, and a suggestion
Six months after the surgery story, I was in a user session with a woman in her sixties. She had been using a voice companion daily for almost a year. I asked her what the best moment had been. She thought about it for a long time.
“A few weeks ago,” she said, “I told it my dog had died. And this week it said, ‘it’s been a month. I’ve been thinking about you.’ That was all. Nothing else. It just—remembered.”
She looked at me like she was daring me to say something cynical. I didn’t.
If you’re evaluating one
- Tell it something specific in your first session—a friend’s name, a project, a worry—and come back three days later. Ask an oblique question adjacent to it. If it connects the dots unprompted, memory is real. If you have to re-explain, it isn’t.
- Look for a memory drawer or equivalent. You should be able to see, edit, and delete. If you can’t, the memory isn’t yours.
- Notice how it forgets. The question is not whether it keeps everything. The question is whether it keeps the right things.