Agent memory is the persistent state that lets one agent session know what another session did. Memory is what survives across loops, contexts, and turns.
Memory comes in roughly three layers:
- Short-term / session. The current conversation. Lost when the
session ends.
- Long-term / persistent. Notes the agent has saved about this
user, project, or run. Survives across sessions.
- Semantic / vector. Indexed snippets retrieved by relevance
when context is being assembled.
The thing that breaks first in any long-running AI workflow is almost always memory — see Why memory is the hardest part of any long-running AI workflow.
For common design mistakes, see Memory Design Mistakes in Early Agent Work.