DEV Community

Rushank Savant
Rushank Savant

Posted on

Day 6: Give Your AI a Memory — Building Chatbots that Actually Remember 🧠

#ai #langchain #python #memory

LLMs are stateless. Every time you send a message, it’s like the first time the AI has ever met you. Today, we’re going to fix that. We’re giving our agents a "notebook" so they can remember who you are and what you’ve discussed.

🐟 The "Goldfish Memory" Problem

In a standard chain, the flow is a one-way street: Input → AI → Output. Once the response is sent, the AI's "brain" resets.

To create a real conversation, we need to feed the past messages back into the AI every single time you hit enter. LangChain makes this incredibly easy by managing the "notebook" for us.

📜 1. The Simple Transcript: ConversationBufferMemory

This is the most straightforward memory type. It stores everything—every "Hi" and every "Bye"—in a raw list.

  • Best for: Short conversations where you need exact wording.

  • The Catch: If the chat gets too long, it will eventually exceed the AI's "Context Window" (the amount of text it can read at once).

from langchain.memory import ConversationBufferMemory

memory = ConversationBufferMemory()

# Manually simulating a chat
memory.save_context({"input": "Hi, I'm Gemini"}, {"output": "Hello Gemini! How can I help?"})
memory.save_context({"input": "What's my name?"}, {"output": "Your name is Gemini."})

# Peek inside the notebook
print(memory.load_memory_variables({}))
Enter fullscreen mode Exit fullscreen mode

✂️ 2. The Efficient Pro: ConversationSummaryMemory

As your chat grows, a raw transcript becomes expensive and slow. Summary Memory uses an LLM to summarize the conversation as it happens.

Instead of remembering 50 pages of text, the AI keeps a 1-paragraph "story so far."

from langchain_openai import ChatOpenAI
from langchain.memory import ConversationSummaryMemory

llm = ChatOpenAI(model="gpt-4o-mini")
memory = ConversationSummaryMemory(llm=llm)

# The AI summarizes the interaction in the background, 
# keeping your 'notebook' slim and focused!
Enter fullscreen mode Exit fullscreen mode

🏗️ 3. The "Window" Trick: ConversationBufferWindowMemory

If you only care about what happened in the last 5 minutes, you use Window Memory. It only keeps the last k interactions. This is a great way to keep your token costs predictable.

from langchain.memory import ConversationBufferWindowMemory

# 'k=2' means it only remembers the last 2 turns of the conversation
memory = ConversationBufferWindowMemory(k=2)

memory.save_context({"input": "My favorite color is Blue"}, {"output": "Got it!"})
memory.save_context({"input": "I live in New York"}, {"output": "Nice city!"})
memory.save_context({"input": "I am a coder"}, {"output": "That's cool!"})

# Now, if we ask "What is my favorite color?", it will have forgotten!
# It only remembers 'I live in New York' and 'I am a coder'.
print(memory.load_memory_variables({}))
Enter fullscreen mode Exit fullscreen mode

🎯 Day 6 Summary

Today, we graduated from building scripts to building Companions. We covered:

Statelessness: Why LLMs naturally forget.

Buffer Memory: Storing the full transcript.

Summary Memory: Condensing the history to save space.

Window Memory: Keeping only the most recent messages.

Your Homework: If you were building an AI Customer Support bot, which memory type would you choose? Think about the balance between "knowing the user's specific problem" and "not getting confused by a 2-hour long chat."

See you tomorrow! ☕

Top comments (0)