How it works
Long-term memory is implemented as a separate store. Common pattern: at session start, retrieve the user's profile and relevant past interactions; at session end, store new important facts learned. The retrieval can be vector-based (similarity search) or structured (SQL).
Example
A productivity agent remembers across sessions that you prefer morning meetings, that your preferred meeting length is 25 minutes, and that you don't take meetings on Fridays. When you ask 'schedule a call with Bob', the agent applies these preferences automatically.