Conversation Memory

Conversation Memory Management

LLMs are stateless per API call — applications must manage conversation history explicitly.

# 1. In-memory sliding window
class ConversationManager:
    def __init__(self, max_turns: int = 10):
        self.history   = []
        self.max_turns = max_turns

    def add(self, role: str, content: str):
        self.history.append({'role': role, 'content': content})
        if len(self.history) > self.max_turns * 2:
            self.history = self.history[-self.max_turns * 2:]

    def get(self) -> list:
        return self.history

# 2. Redis-backed persistent memory
import redis, json
r = redis.Redis(decode_responses=True)

def load_history(session_id: str) -> list:
    return json.loads(r.get('chat:' + session_id) or '[]')

def save_history(session_id: str, history: list):
    r.setex('chat:' + session_id, 3600, json.dumps(history))

# 3. Summarisation memory (for very long conversations)
def summarise_history(history: list) -> str:
    prompt = 'Summarise this conversation in 3 sentences: ' + str(history)
    return llm.complete(prompt)

# When history gets too long:
# summary = summarise_history(old_history)
# new_history = [{'role':'system','content':'Previous: ' + summary}]