Agent memory

Agent memory is how an AI agent retains information so it can act coherently across many steps — and often return to that information in later sessions. There are two levels: short-term memory, the current working context of a given task, and long-term memory, knowledge persisted externally — in a database, file, or notebook — that the agent draws on selectively when it is needed.

It is easy to confuse this with the context window, but they are not the same. The context window is a hard limit on how many tokens the model sees at once; once it is exceeded, content simply falls out of reach. Agent memory sits one layer above: it is the logic deciding what to store outside the model, and what to reload into the window and when, so that important conclusions are not lost. Long-term memory is often implemented through document retrieval, which pulls only the relevant fragments of earlier knowledge into the context.

In an enterprise deployment, memory determines whether an agent "remembers" a client's preferences from a previous conversation or starts from scratch every time. It is designed deliberately: what to persist, how long to keep it, and how to protect the data — because stored memory becomes a lasting record of information about users and processes.