What is the difference between RAG and memory for AI agents?

RAG (Retrieval-Augmented Generation) and memory are two different architectures that are frequently confused because both inject external content into an LLM's context window. They solve fundamentally different problems: RAG retrieves static knowledge to answer a query; memory persists and evolves user-specific state across sessions.

How RAG works

RAG indexes a corpus of documents — PDFs, wikis, support articles — as vector embeddings in a database. At query time, the user's input is embedded and the top-K semantically similar chunks are retrieved and injected into the prompt. The retrieved content is the same regardless of who is asking. RAG answers the question: what does the knowledge base say about this topic?

How memory works

Memory is a read-write store attached to a specific user or agent. When a user says 'I'm vegetarian' or 'my deadline is Friday,' a memory system extracts and stores that fact. On the next session, those facts are retrieved and injected into the prompt — not because they are semantically similar to the query, but because they are known facts about this user. Memory answers the question: what do I know about this specific person right now?

The key architectural differences

A concrete example

A customer support bot using RAG retrieves your return policy when a user asks how to return an item — the same answer for every user. A memory-augmented bot also knows that this specific user already attempted a return last Tuesday, has a premium account, and previously expressed frustration about shipping delays. That context comes from memory, not documents.

When to use which

Use RAG when the information is the same for all users and lives in documents: internal wikis, product documentation, compliance policies. Use memory when the information is specific to a user and changes over time: preferences, ongoing tasks, past decisions. Most production systems need both — RAG for universal knowledge, memory for personalization. Mem0 provides the memory layer, with a hybrid vector and graph store that handles extraction, deduplication, and retrieval across sessions.

from mem0 import Memory

m = Memory()

# Store a user-specific fact
m.add("I prefer dark mode and use Python 3.11", user_id="alice")

# Retrieve relevant facts at query time
results = m.search("user environment preferences", user_id="alice")
# Returns: [{"memory": "Prefers dark mode, uses Python 3.11", "score": 0.94}]