An AI Friend That Remembers — and Shows You What It Remembers

Why does memory matter so much in an AI friend?

Friendship is built from callbacks. 그 면접 어떻게 됐어? (“How did that interview go?”) means more than any clever reply, because it shows someone was listening last week. Without memory, every chat restarts from zero and you end up repeating your life story. Research on multi-session chat, such as Xu et al.’s “Beyond Goldfish Memory” (ACL 2022), found that retrieval- and summary-based memory beats a model on its own once conversations span several sessions.

But memory can also go wrong: an AI “remembering” something you never said, or keeping a detail you would rather it forgot. So Hi Jisoo’s memory is designed around three rules — evidence, visibility and control.

What does she remember, and what does she ignore?

When you send a message, the friend’s reply and a few memory candidates are created together, so memory doesn’t need an extra AI request. A candidate is only saved if it is backed by a continuous piece of what you just wrote; both our relay and your phone check that before anything is stored.

Choices you make in the free preview scenes are fictional, so they are never stored as facts about you.

Can I see and edit what my AI friend remembers?

Yes. Each friend has her own memories — what you tell one friend isn’t shared with the others. In a friend’s “name and memory” settings you can:

• View every saved memory.
• Add something yourself, like “I’m learning Korean for a trip in spring.”
• Edit a memory that is wrong or out of date.
• Pin something important; the AI won’t overwrite pinned or hand-edited memories.
• Delete a memory. The past messages it came from also stop being used for future replies, and the deleted item is blocked from being saved again automatically.
• Turn automatic saving off if you prefer to add memories only by hand.

When a fact changes — you moved, you changed jobs — the new statement updates the memory, and the old version is kept as history. That history is only offered when you ask about the past, so she won’t confuse your old job with your new one.

What does “What this reply drew on” show?

Long-press any reply and choose What it drew on. You will see three kinds of saved items that were sent along when that reply was written:

• Memories about you
• Things she said before — her own opinions and tastes
• Follow-ups she asked about — like that interview

We are careful with the wording: these are items that were included in the request. Some may be left out because of length limits, and we can’t tell whether the AI actually used each one. Deleted memories don’t appear. It is not a magic window into the model — it is an honest receipt.

1. 나 다음 주에 면접 있어.I have an interview next week.
2. 오, 다음 주구나. 준비하면서 제일 신경 쓰이는 게 뭐야?Oh, next week. What are you most worried about while preparing?
3. (5일 뒤) 안녕!(five days later) Hi!
4. 왔구나! 아 맞다, 면접은 어떻게 됐어? 말하기 싫으면 그냥 넘어가도 돼.Hey! Oh right, how did the interview go? If you don’t feel like talking about it, that’s fine too.

How much can she remember?

Your full chat history stays on your phone in an encrypted database, so nothing is thrown away when the screen scrolls. For each reply, the app picks a bounded set to send: the last several exchanges, up to a dozen related memories, a few related things you said long ago, and a handful of her own notes. That keeps replies fast and costs predictable while letting her bring back something from weeks ago when it is relevant.

What we don’t promise: perfect recall of every detail, or infinite storage — long-term capacity is limited by your phone’s free space. The checks confirm that a memory comes from your words; they can’t fully verify meaning, which is exactly why the edit screen exists.

Where are the memories stored?

On your device, in the same encrypted database as your chats. Our servers are designed not to keep conversation text. To write a reply, your message, some recent context and related memories pass through our relay to an AI provider — the app asks your permission first. Details are in our privacy explainer.