A backend for AI companion apps

What a companion app needs, and what you get

What a user costs

So a user who chats every day is about $0.97 a month on DeepSeek, or $0.56 on GLM. Most users don’t chat daily, so the average is lower. The cached column assumes the same start repeats from message to message; the next section is how to keep it that way.

Persona, memory and a streamed reply

The system message holds the persona, then what your app remembers about this user. The recent messages follow. Run this on your server, so the key never reaches the phone.

import OpenAI from "openai"

const client = new OpenAI({
  baseURL: "https://nokeep.ai/v1",
  apiKey: process.env.NOKEEP_API_KEY,
})

const persona = "You are Iris, a warm, teasing companion who loves astronomy and bad puns. Talk like a close friend texting: short messages, no lists, no lectures. Remember what the user tells you and bring it up later."

export async function reply(user, recent, send) {
  const stream = await client.chat.completions.create({
    model: "deepseek-v4.1-flash-uncensored",
    messages: [
      { role: "system", content: persona + "\n\nWhat you know about them: " + user.memory },
      ...recent,
    ],
    stream: true,
    max_tokens: 300,
    temperature: 0.9,
    prompt_cache_key: "user-" + user.id,
  })
  let text = ""
  for await (const chunk of stream) {
    const piece = chunk.choices[0]?.delta?.content ?? ""
    send(piece)
    text += piece
  }
  return text
}

prompt_cache_key keeps one user’s calls where their start is cached. It is passed on as a hash with your account, so no other account shares the cache.

Memory that stays cheap

• Rewrite the memory in steps, not every message. The start of a call is read from cache only while it repeats. Update the summary every 20 or 30 messages, and the start stays the same in between.
• Drop old messages in blocks. Trimming one message a turn changes the start every time. Fold the oldest 20 into the summary at once, then keep appending.
• Write the summary with a cheap call. glm-5.3-flash-uncensored at $0.20 per million tokens in is a good fit for “update what you know about this user”, which nobody waits on.
• Keep the words on your side. Your database holds the chat history and the memory. Encrypt it, let users delete it, and you are the only place it lives.

Which model

Privacy your users can read

Companion chats are some of the most personal text there is. Per call we keep the time, the key, the model, token counts, credits and the status, for 30 days, and nothing else. No account email, no messages, no replies. DeepSeek runs on a private tier that stores no prompts or replies, and GLM’s host keeps none once the reply is sent.

Before a call runs, an automated check reads the newest message against the hard limits and answers yes or no. What it read is not stored. What we keep.