Every AI companion app I tried wanted the same three things: a monthly subscription, an account, and my conversations living on their server forever. The "memory" was always a black box too — the character would forget something important, and there was no way to see what it remembered, let alone fix it.

So I built the opposite. Xinyu is an AI companion app for Android with no backend. Not "backend I promise to keep private" — there is literally no server. The whole brain is ~4,700 lines of pure-stdlib Python (zero pip dependencies) running inside the APK via Chaquopy, talking straight from your phone to the DeepSeek API with your own key.

Full disclosure: this is my own product, and I sell the license on Gumroad. This post is about how it's built and why.

The design rule: if there's no server, most problems disappear

  • No subscription. There's no server bill, so there's nothing to subscribe to. The app is a one-time $9.99 license. Your only running cost is your own API credit — a few dollars lasts months.
  • No account, no uploads. Chats, memories, character data: all plain files in the app's private storage on the phone. I couldn't read your conversations if I wanted to.
  • Licensing without a license server. Activation codes are RSA-signed and bound to a machine code derived from the device. I sign codes with a small offline tool; the app verifies the signature locally. Clock-rollback tricks are handled by persisting the max timestamp the app has ever seen. Old-school shareware mechanics, and they work fine.

Memory you can open like a notebook

The feature I care about most. Everything the companion remembers about you lives in a notebook inside the app:

  • Entries are created from your conversations, each with a timeline of how it changed.
  • When a memory is updated, the old version is struck through, not silently overwritten — you can see what it used to believe about you.
  • You can add, edit, delete, and mark entries sensitive. Sensitive entries never surface in group chats.

Anyone who has hand-curated a lorebook in SillyTavern knows why this matters: memory you can't inspect eventually becomes memory you can't trust. (The app imports SillyTavern character cards, by the way — PNG chara/ccv3 and JSON.)

Group chats need a director, not a round-robin

My first version of group chat had every character reply to every message in turn. It read like a meeting, not a conversation.

Now there's a director step: one extra LLM call per group message decides who speaks (usually 1–3 characters), who stays silent, and in what order. Non-speakers still get their mood appraised from what happened. Each character also has a mood and affection state shaped by group events, plus a slow-burning "what's on their mind" thread that develops over days and leaks into later chats. The difference in feel is hard to overstate — silence turns out to be half of believable group dynamics.

The unglamorous parts I'm proud of

  • The APK is built with a hand-rolled pipeline (kotlinc → d8 → aapt2 → zipalign → apksigner) because Gradle simply would not run in my build environment. It works, it's scripted, and I never want to talk about it again.
  • Voice is pluggable: cloud TTS if you add a SiliconFlow key, Android's native TTS otherwise.
  • The English edition reuses the exact same brain with a ui_lang switch — two translation dictionaries applied at render time, and growth-stage names translated at the API layer so the internal logic still keys off the original terms.

Honest limitations

  • Android only, sideloaded APK — no Play Store (solo dev, that route isn't open to me right now).
  • Activation is manual: the app shows a machine code, you email it to me, I send your code back. The 7-day free trial works the same way — free, but not instant.
  • The interface and main companion are fully English; the built-in preset characters' story content is still Chinese (translation is on the roadmap). Imported cards are unaffected, and every character can be renamed.

Try it

If you build with LLMs, I'm curious: how are you handling long-term memory in your projects — and would your users trust it more if they could open it and edit it, like a notebook? That bet is the whole product, and I'd love to hear where it breaks.