Chat assistant with long-term memory

Goal: a chat assistant that remembers facts about the user from earlier conversations.

Pattern: for every turn, read relevant memory, answer with it in the system prompt, then write what you learned.

import { formatContext, type WikiMemory } from '@equationalapplications/core-llm-wiki';

export async function chatTurn(wiki: WikiMemory, userId: string, userMessage: string) {
  // 1. Recall
  const memory = await wiki.read(userId, userMessage, { maxResults: 8, tokenBudget: 1200 });
  const context = formatContext(memory, { maxEvents: 5 });

  // 2. Answer
  const reply = await callModel({
    system: `You are a friendly assistant. What you remember about the user:\n\n${context}`,
    user: userMessage,
  });

  // 3. Remember (don't block the reply on it)
  void wiki.write(userId, { event_type: 'observation', summary: `User said: ${userMessage}` });
  void wiki.write(userId, { event_type: 'action', summary: `Assistant replied: ${reply.slice(0, 500)}` });

  return reply;
}

When maintenance runs (call runPendingMaintenance(), or set maintenance: 'auto' for every 20 events), the librarian turns these into facts such as “Prefers vegetarian recipes” and tasks such as “Remind the user about the dentist on Friday”.

In an Expo screen

import { useState } from 'react';
import { useWiki, useEntityStatus } from '@equationalapplications/expo-llm-wiki';
import { chatTurn } from './chat';

export function ChatScreen({ userId }: { userId: string }) {
  const wiki = useWiki();
  const { librarian } = useEntityStatus(userId);
  const [messages, setMessages] = useState<{ role: string; text: string }[]>([]);

  const send = async (text: string) => {
    setMessages((m) => [...m, { role: 'user', text }]);
    const reply = await chatTurn(wiki, userId, text);
    setMessages((m) => [...m, { role: 'assistant', text: reply }]);
  };

  return (
    <>
      {librarian && <Text style={{ opacity: 0.6 }}>Updating memory…</Text>}
      {/* render messages + input that calls send() */}
    </>
  );
}

Variations

Summarize rather than log raw messages. Write one short observation per meaningful statement instead of every message. You get fewer events and better facts.
Let users correct memory. Show read(userId, '') (the most recent facts) in a settings screen with a delete button that calls forget({ entryId }).
Keep history when facts change. The default 'ops' librarian strategy supersedes “lives in Seattle” with “moved to SF” instead of overwriting it.