A language model on its own has no persistent memory, each call is independent unless something explicitly feeds it prior context. AI agent memory is the mechanism that closes that gap: storing relevant facts, past decisions, or user preferences somewhere retrievable, and pulling the right pieces back in when they're relevant to a new task.

Short-term memory covers the current conversation or task, what was just said a few turns ago. Long-term memory persists across sessions entirely, an agent remembering that a particular vendor always sends invoices with a certain quirk, months after first encountering it.

Frequently Asked Questions

What's the difference between short-term and long-term agent memory?

Short-term memory holds context within one active task or conversation. Long-term memory persists across separate sessions entirely, letting an agent build up institutional knowledge over time rather than relearning the same thing every time it's invoked.

Why does agent memory usually rely on a vector database?

Because retrieving the right memory later requires finding what's semantically relevant to the current situation, not just an exact keyword match, and vector databases are built specifically for that kind of similarity search over stored information.

What happens if an agent's memory is wrong or outdated?

A stale or incorrect memory can bias a future decision the same way an outdated assumption would for a person. Well-designed agent memory systems need a way to update or expire information, not just accumulate it indefinitely.

Is agent memory the same thing as a bigger context window?

No. A context window is how much text a model can process in one call. Memory is a separate system for storing and selectively retrieving information across many calls over time, which a large context window alone doesn't provide.