External Memory: Scales for the World

If you store message history in a Python variable (memory = ...), that memory dies as soon as your script stops. In a real website, your server might restart, or you might have 10,000 users. You need a Database to store the state.

1. The Redis Advantage

Redis is an in-memory database that is extremely fast. It is the most popular choice for "Hot" conversation state.

from langchain_community.chat_message_histories import RedisChatMessageHistory

# Every conversation gets a unique key
history = RedisChatMessageHistory(
    session_id="user_123_conversation_1",
    url="redis://localhost:6379"
)

history.add_user_message("Hi!")
print(history.messages)

---

2. Using Session IDs for Isolation

In your backend code, you typically extract the session_id from a cookie or a JWT token. You then initialize the LangChain memory using that ID. This ensures the agent only "remembers" what happened with that specific user.

---

3. Visualizing Distributed Memory

graph TD
    U1[User A] --> API[FastAPI Server]
    U2[User B] --> API
    API -->|Session 1| R[(Redis DB)]
    API -->|Session 2| R
    R -->|Memory 1| LLM[Agent Brain]

---

4. Why use Postgres?

If you want the memory to last for years (e.g., a personal companion bot), Redis might be too expensive or risky. Use Postgres (SQL) for long-term, cold storage of chat history.

---

5. Engineering Tip: TTL (Time to Live)

Most chat history has an "Expiration Date." In Redis, you can set a TTL of 24 hours. If the user doesn't come back in a day, the memory is deleted automatically. This saves space and improves privacy.

---

Key Takeaways

• Python-local memory is only for testing; never for production.
• Redis is the standard for fast, ephemeral chat history.
• Session IDs are mandatory for separating user conversations.
• TTL helps manage database size and data privacy.