Conversation Memory Module (06_conversation_memory)

Stores conversation outputs as embeddings for context-aware retrieval.

Files

• memory_processor.py: Parses LLM's output, builds chunks, and stores them in Pinecone.
• config.py: Configuration for embedding model and Pinecone namespace.
• error_handling.py: Error handling for Pinecone operations (only needed for full pipeline, not for test_multi_retrieval.py)

Features

• Deduplication: Checks similarity before storing. If a memory is >95% similar to an existing one, it's skipped.
• Session Parsing: Automatically parses LLM's Query/Answer/Evidence format.
• Unified Retrieval: Designed to work with 04_vectoredb/retriever_multi.py.

Usage

1. Storing Memory

from memory_processor import store_memory_sync

raw_output = '''
Query: "What is the patient's age?"
Answer: The patient is 45 years old.
Evidence: Patient is a 45-year-old male.
'''

# Automatically checks for duplicates before storing
result = store_memory_sync(raw_output, pinecone_index)
if result:
    print(f"Stored: {result['chunk_id']}")
else:
    print("Skipped (duplicate or error)")

2. Retrieval Configuration

Adjust thresholds in 04_vectoredb/config.py:

• MEMORY_MIN_SCORE: Minimum impact (default 0.4)
• MEMORY_BOOST: Ranking boost (default 0.05)

Integration Flow

1. LLM generates answer
2. memory_processor parses & embeds
3. Checks Pinecone for duplicates (>0.95)
4. Stores if unique
5. Future queries retrieve this memory via retriever_multi.py