AI Agent Memory Architectures: Robust Conversational AI

Core Problem:* AI agent demos often fail after a few turns due to lack of memory, hindering reasoning and leading to brittle interactions. The post emphasizes that prompt engineering alone is insufficient; robust memory mechanisms are crucial.
Memory Hierarchy:* The PhiloAgents course (Lesson 3) addresses this by implementing a memory architecture comprising short-term (conversational flow), semantic (factual knowledge via agentic RAG), episodic (past experiences), and procedural memory.
Short-Term Memory Implementation:* Short-term memory is managed using LangGraph's state management, persisting conversation context and recent messages. The `PhilosopherState` class stores static (philosopher's attributes) and dynamic context (messages, summary, RAG context). State persistence to MongoDB enables reuse across processes and supports RESTful APIs for multiple users, differentiated by `thread_id`.
Long-Term Memory Implementation (Agentic RAG):* Long-term memory leverages RAG, involving ingestion and retrieval phases. The RAG pipeline includes document extraction, cleaning, chunking, deduplication (using MinHash LSH), embedding, and loading into MongoDB with a vector index. The agent dynamically decides when to query semantic memory using LangChain tools.
Database Choice:* MongoDB is used as the agentic-ready database due to its support for unstructured collections combining text and vectors, reducing infrastructure overhead and enabling scalability.
Deduplication:* MinHash is used to deduplicate documents from Wikipedia and SEP. The `deduplicate_documents` function uses MinHash and Locality Sensitive Hashing (LSH) to identify similar document pairs.

Source:

https://www.linkedin.com/posts/pauliusztin_heres-why-98-agent-demos-break-after-3-activity-7327626410951364610-6cd0