Extra memory for an Agentic AI w/Chroma

Scenario: Building an AI assistant with Chroma memory

• Goal: Use Chroma to provide long-term, embedding-based memory for conversations.

• The agent evolves its responses based on prior interactions, enabling personalized, context-aware conversations.

Step-by-Step Code Implementation

1. Install Required Libraries

!pip install chromadb langchain openai

• Purpose:

  • chromadb: Manages the memory system, storing embeddings and contexts.

  • langchain: Integrates Chroma memory with LLMs.

  • openai: Connects to OpenAI’s GPT-3.5 or GPT-4 for conversation.

• What Happens: Libraries are installed for managing memory and building the agent.

2. Import Libraries

from langchain.chains import ConversationChain
from langchain.memory import ChromaMemory
from langchain.llms import OpenAI
from chromadb.config import Settings

• Purpose:

  • ConversationChain: Handles the flow of conversation using memory.

  • ChromaMemory: Embedding-based memory system that retains context.

  • OpenAI: Powers the LLM for natural language processing.

• What Happens: These components will be used to integrate the memory system with the LLM.

3. Initialize Chroma Memory

# Configure Chroma settings
chroma_settings = Settings(
    persist_directory="chroma_memory",  # Directory to save memory
    chroma_db_impl="duckdb+parquet",   # Database type
    anonymized_telemetry=False
)

# Create Chroma memory
memory = ChromaMemory(settings=chroma_settings)

• Purpose:

  • Configures Chroma to store memory in a persistent directory (chroma_memory).

  • Enables embedding-based storage for long-term memory retention.

• What Happens:

  • A Chroma database is created to store conversation embeddings.

  • Memory will persist even after the session ends.

4. Initialize the LLM and Conversation Chain

# Set up the LLM (OpenAI GPT)
llm = OpenAI(model="gpt-3.5-turbo", temperature=0.7)

# Create a conversation chain with memory
conversation = ConversationChain(
    llm=llm,
    memory=memory,
    verbose=True
)

• Purpose:

  • llm: Defines the language model (GPT-3.5 or GPT-4).

  • conversation: Combines the LLM and Chroma memory for context-aware conversations.

• What Happens:

  • The AI agent is now capable of retaining memory between queries.

  • Verbose mode provides detailed logs of how memory is used during interactions.

5. Engage in a Conversation

# First interaction
response_1 = conversation.run("Hi, I love Python programming!")
print(response_1)

# Second interaction
response_2 = conversation.run("What did I just tell you about my interests?")
print(response_2)

# Third interaction
response_3 = conversation.run("Can you suggest a Python library for data visualization?")
print(response_3)

• Purpose:

  • Simulates a conversation where the agent retains memory and adapts responses based on past interactions.

• What Happens:

  • First Interaction: The user’s preference for Python programming is stored in Chroma memory.

  • Second Interaction: The agent recalls and references the stored memory.

  • Third Interaction: The agent provides relevant suggestions based on the context.

Sample Output

Verbose Logs

> Entering new ConversationChain...
Memory retrieved: []
Thought: Respond to the user's greeting.
Output: Hi! Python programming is awesome. How can I assist you?
> Finished chain.

> Entering new ConversationChain...
Memory retrieved: [{'user_input': 'Hi, I love Python programming!', 'output': 'Hi! Python programming is awesome. How can I assist you?'}]
Thought: The user mentioned loving Python programming earlier. I should reference it.
Output: You mentioned that you love Python programming. How can I assist you further?
> Finished chain.

> Entering new ConversationChain...
Memory retrieved: [{'user_input': 'Hi, I love Python programming!', 'output': 'Hi! Python programming is awesome. How can I assist you?'}, 
                   {'user_input': 'What did I just tell you about my interests?', 'output': 'You mentioned that you love Python programming. How can I assist you further?'}]
Thought: The user is asking for a Python library. Recall their interest in programming.
Output: For data visualization, I recommend using Matplotlib or Seaborn. They are powerful Python libraries for creating charts and graphs.
> Finished chain.

Final Output

Agent Response 1: Hi! Python programming is awesome. How can I assist you?
Agent Response 2: You mentioned that you love Python programming. How can I assist you further?
Agent Response 3: For data visualization, I recommend using Matplotlib or Seaborn. They are powerful Python libraries for creating charts and graphs.

Step-by-Step Explanation of Impact

1. First Interaction:

   • The user mentions they love Python programming.

   • This input is embedded and stored in Chroma memory for future reference.

2. Second Interaction:

   • The agent retrieves the stored memory.

   • It references the user’s interest in Python programming to deliver a personalized response.

3. Third Interaction:

   • The agent uses both memory and the context of the current query to suggest relevant Python libraries.

Why This Matters

• Before Chroma: The agent would forget the user’s preferences after each query.

• After Chroma: The agent retains and evolves its understanding, enabling more natural, personalized conversations.