Mastering Memory Management in LLM-Based Applications!

---

---

AI Nuggets

CLI Commands

1. Navigate to the desired folder:

   cd Desktop/content/tutorials  

2. Create a new folder:

   mkdir memory2agents  

3. Change directory into the new folder:

   cd memory2agents  

4. Initialize a poetry project:

   poetry init  

5. Install dependencies using poetry:

   poetry add python-dotenv lgraph lchain ipython openai lchain_community  

6. Create an interactive Python notebook:

   ipython notebook lesson_01.ipynb  

Python Code Snippets

1. Load environment variables:

   from dotenv import load_dotenv  
   load_dotenv()  

2. Create a tool for web searches:

   from tavil import TavSearchResults  
   tool = TavSearchResults(max_results=5)  

3. Define an agent state:

   from lchain import Annotated, Sequence, Message, operator  
   class AgentState(dict):  
       messages: Annotated[Sequence[Message], operator.add] = []  

4. Create a model:

   from lchain import GPT4  
   model = GPT4("gpt4")  

5. Bind the model with tools:

   model_with_tools = model.bind_with_tools(model, tools)  

6. Create a react agent:

   from lgraph.prebuilt import ReactAgent  
   agent_executor = ReactAgent(model, tools, checkpoint=memory, configurable={"trade_id": "test_trade"})  

7. Test the agent:

   response = agent_executor.invoke({"messages": [HumanMessage("What's the capital city of France?")]})  
   print(response["messages"][0]["content"])  

8. Create in-memory database:

   from lgraph.checkpoint import SQLiteCheckpoint  
   memory = SQLiteCheckpoint.from_connection_string(":memory:")  

9. Create a permanent SQLite database:

   memory = SQLiteCheckpoint.from_connection_string("sqlite.db")  

Website URLs

1. Obtain Tavil API key:
   • Visit the Tavil documentation (link provided in the video description).
2. Obtain OpenAI API key:
   • Search for “OpenAI API key” online and follow the first links provided.

Tips

• When creating a tool for web searches, do not keep a very large number of results to avoid overflowing the context window of your model.
• The trade_id in the configurable dictionary acts as a memory identifier to separate different users’ conversations.
• To start a new conversation, simply change the trade_id in the configurable dictionary.
• To use a permanent SQLite database, specify the file name of the SQLite database instead of using :memory:.

Additional Information

• The video description contains links to obtain API keys and the official documentation.
• The speaker encourages viewers to reach out with project ideas or suggestions on LinkedIn at Prince Kampa or via email.