It is distinct from two things it is often confused with. Context windows are working memory: everything in them evaporates when the session ends. And retrieval over documents (RAG) recalls what the organization wrote, but not what the agents themselves learned by doing the work.
Useful agent memory splits WHAT from WHERE. Knowledge (a decision, a correction, a learned constraint) ranks by relevance, importance, and recency. Location (which file, which module, which doc) is a map that spares the agent re-exploring territory it has already covered.
The capture side matters as much as recall: memory written automatically from everything becomes noise, and memory that depends on someone remembering to save it becomes empty. The durable pattern wires capture into the process itself, at the steps where decisions actually happen.
Memory also needs upkeep to stay useful. Knowledge goes stale: a decision is reversed, a file moves. A good memory system lets the next run correct an entry rather than add a contradicting one, and lets a forgotten entry come back when the forgetting was a mistake. Without that upkeep, the same question is answered again and again, and the answers drift apart.