Short-term vs long-term memory

Models have no memory — every call is a stranger. An agent's memory is whatever you put back into the context.

Two kinds:

• Short-term: append the conversation history (simple; grows and costs more);
• Long-term: store key facts in a vector store and retrieve them when needed (cheaper, scales).

Rule of thumb: start short-term, add long-term when volume or cross-session recall demands it.

One line to remember:

Memory isn't "storing" — it's "retrieving next time".