Memory

A base LLM has no memory. Each conversation starts fresh. Everything the model 'knows' about you or your task exists only within the current context window, and disappears when you close the chat. This is fine for one-off questions but breaks down for personal assistants, long-running projects, or agentic tasks that span multiple sessions.

Memory in AI systems usually means one of several things. In-context memory is just the conversation history being fed back into the prompt: it's temporary and bounded by the context window. External memory is a separate store, often a database, where facts and history are saved and retrieved when needed, similar to how RAG retrieves documents. This is how products like ChatGPT's memory feature or Claude Projects work: they store key information about you and pull it back when relevant.

For builders designing agents, memory architecture is a real design decision. Do you store every message? Just extracted facts? User preferences? Task state? How do you retrieve the right memory at the right time without overloading the context? Memory poisoning is also a real concern: if an attacker can write to an agent's memory store, they can influence future behavior. Agent memory is one of the richer engineering problems in the agentic tooling space.