Build a chatbot with memory

import os
 
from openai import OpenAI
 
client = OpenAI(base_url="https://app.syntic.ai/v1", api_key=os.environ["SYNTIC_API_KEY"])
 
BASE_SYSTEM = """
You are Lingotrail's conversation tutor. Chat with the learner in their target language
at their level, correct one mistake per reply (gently, after your answer), and keep replies short.
""".strip()
 
 
def chat_turn(chat: dict, user_text: str) -> str:
    chat["turns"].append({"role": "user", "content": user_text})
    reply = client.chat.completions.create(
        model="Amara S1.2",
        stream=False,
        max_tokens=400,
        messages=[{"role": "system", "content": build_system(chat)}] + chat["turns"],
    )
    text = reply.choices[0].message.content or ""
    chat["turns"].append({"role": "assistant", "content": text})
    chat["last_prompt_tokens"] = reply.usage.prompt_tokens
    return text

KEEP_RECENT = 12          # turns sent verbatim
SUMMARISE_AT = 30         # start folding when the chat has this many turns
 
 
def compact(chat: dict) -> None:
    if len(chat["turns"]) < SUMMARISE_AT:
        return
    old, recent = chat["turns"][:-KEEP_RECENT], chat["turns"][-KEEP_RECENT:]
    transcript = "\n".join(f"{t['role']}: {t['content']}" for t in old)
    r = client.chat.completions.create(
        model="Amara Quick",
        stream=False,
        max_tokens=300,
        messages=[
            {"role": "system", "content": (
                "Update the running summary of a tutoring chat. Keep: topics practised, "
                "recurring mistakes, words the learner struggled with, anything they said about themselves. "
                "Drop greetings and small talk. Under 150 words."
            )},
            {"role": "user", "content": f"Current summary:\n{chat.get('summary') or '(none)'}\n\nNew turns:\n{transcript}"},
        ],
    )
    chat["summary"] = r.choices[0].message.content
    chat["turns"] = recent

import json
 
FACT_RULES = """
You maintain a short profile of a language learner for their tutor.
Given the existing facts and a new conversation, return the updated list as JSON:
{"facts": ["...", "..."]}
Keep only facts useful in future lessons: goals, level, native language, interests, preferences,
persistent mistakes. Remove facts the learner has corrected. At most 15 facts, each under 20 words.
Never store health details, passwords, payment data or other people's personal information.
""".strip()
 
 
def update_facts(user: dict, transcript: str) -> None:
    r = client.chat.completions.create(
        model="Amara Quick",
        stream=False,
        max_tokens=500,
        messages=[
            {"role": "system", "content": FACT_RULES},
            {"role": "user", "content": json.dumps({"existing_facts": user.get("facts", []), "conversation": transcript})},
        ],
    )
    text = r.choices[0].message.content or ""
    try:
        facts = json.loads(text[text.index("{"): text.rindex("}") + 1])["facts"]
    except (ValueError, KeyError):
        return                      # keep the old facts rather than store something broken
    user["facts"] = [f for f in facts if isinstance(f, str)][:15]

def build_system(chat: dict) -> str:
    parts = [BASE_SYSTEM]
    if chat["user"].get("facts"):
        parts.append("What you know about this learner:\n- " + "\n- ".join(chat["user"]["facts"]))
    if chat.get("summary"):
        parts.append("Earlier in this conversation:\n" + chat["summary"])
    return "\n\n".join(parts)