As AI agents become capable of handling tasks across multiple sessions, they need more than a large context window. The article explains that retrieval and memory solve two different problems. Retrieval brings in external information that the model needs at the moment—such as documentation, databases, policies or code—while memory preserves information the agent has learned or generated through previous interactions. Without these mechanisms, long-running agents may repeatedly ask the same questions, forget previous decisions or fail to use information they encountered earlier.
Retrieval is essentially the agent's connection to outside knowledge. A typical retrieval-augmented generation (RAG) system breaks documents into chunks, converts them into embeddings and stores them in a searchable index. When a user asks something, the system finds the most relevant pieces and places them into the model's context. This information is generally shared across users—for example, everyone querying a company's product documentation can access the same underlying knowledge base.
Memory is different because it records the agent's own history. Short-term memory maintains the current session, while long-term memory can preserve information across sessions, such as a customer's preferences, previous decisions or facts learned during earlier interactions. The article gives a support-agent example: remembering that a particular customer prefers email is memory, while looking up the company's latest shipping policy is retrieval. Both may enter the same prompt, but they originate from different sources and serve different purposes.
The article's main conclusion is that effective agentic systems generally need both retrieval and memory. Retrieval without memory makes an agent repeatedly rediscover previous conclusions and prevents meaningful personalization; memory without retrieval leaves it disconnected from current external information. The challenge is deciding what to retrieve, what is worth remembering, when information has become stale, and how to resolve conflicts between the two. Ultimately, good agent architecture is about giving the model the right context at the right time without flooding it with unnecessary information.
