Skip to content

9. Memory Layer (Mem0)

The Memory Layer is an optional module in the ML-CB-B-identifier ecosystem designed to augment your interactions with an LLM (Large Language Model) or agent by persisting long-term context. It utilizes Mem0, a memory layer for AI assistants.

Core Concept

When integrating this classification model with chat agents (like a veterinary assistant bot), the agent needs to remember previous conversations, user preferences, or specific facts about the predicted breeds.

The Memory Layer provides a ChromaDB-backed vector store to handle this context persistence.

Directory Structure

memory/                 # Mem0 vector store module
├── __init__.py
└── service.py          # The core MemoryService class
outputs/
└── memory/             # Local ChromaDB SQLite storage (persisted data)

Scoping Memories

To ensure privacy and relevance, all memories are strictly scoped. When storing or retrieving a memory, you must provide: - user_id: The ID of the human interacting with the system. - agent_id: The ID of the specific AI agent. - run_id: The session or thread identifier.

This prevents cross-contamination of context between different users or different assistant workflows.

API Endpoints

The FastAPI webapp automatically mounts endpoints for the memory layer (located in webapp/server.py). You can test these via the "Memory" tab in the UI, or hit them programmatically:

1. Store a Memory

curl -X POST http://localhost:8000/api/memory \
  -H "Content-Type: application/json" \
  -d '{
    "text": "The user is primarily interested in Gir cattle and lives in Gujarat.",
    "user_id": "user_123",
    "agent_id": "vet_bot",
    "run_id": "session_001"
  }'

2. Retrieve Relevant Context

When the user asks a question, query the memory to inject context into the LLM prompt: 

curl -X POST http://localhost:8000/api/memory/search \
  -H "Content-Type: application/json" \
  -d '{
    "query": "Which breed was I asking about earlier?",
    "user_id": "user_123",
    "agent_id": "vet_bot",
    "run_id": "session_001",
    "limit": 5
  }'
This returns a list of the most semantically relevant memories based on vector similarity.

Integrating with the Classifier

A typical pipeline utilizing both the model and the memory layer looks like this:

  1. User uploads an image.
  2. The ML model predicts Sahiwal cattle.
  3. Agent queries memory: search("User's history with Sahiwal cattle").
  4. Memory returns: "User previously treated a Sahiwal for a tick infection."
  5. Agent response: "This is a Sahiwal. Do you want me to pull up the tick treatment plan we discussed last month?"
  6. Agent stores memory: store("User uploaded a new photo of their Sahiwal on Sept 6.")

Configuration

No extra configuration is required. The vector store is created automatically in outputs/memory/ the first time you invoke the service. Ensure you have the required dependencies (mem0ai, chromadb) installed via requirements.txt.