Working memory is the context of the current task, passed to the model at each step. Long-term memory keeps information from one session to the next in external storage, read back on demand. Since the model's context is limited, you must choose what to keep.
Example
Long-term memory: user preferences read from a database at the start of each session.