Short-term vs long-term agent memory

Short-term memory

Short-term, or working, memory is the context of the task in progress: the conversation so far, intermediate results, the plan. It usually lives in the model’s context window or a scratchpad, and it is thrown away when the task ends. Its limit is size, since a long task can outgrow the window.

Long-term memory

Long-term memory outlives the session. It stores what should carry over: a user’s preferences, decisions already made, facts established once. It lives in a store the agent reads from and writes to, and the challenge is retrieval: pulling back the few relevant items without flooding the context.

Using both

A capable agent uses short-term memory to stay coherent now and long-term memory to stay useful later. They are not interchangeable. Treating the context window as long-term memory loses everything at session end; treating a long-term store as working memory uses it for a job the context window already handles. For a customer-facing agent there is a third input neither covers: what the customer is doing in your product right now, read fresh before each reply.