Context Management

Checkpoints

Run state is saved at milestones. Interruptions and failures resume from the checkpoint, not from zero.

Long-term memory

Knowledge worth keeping across runs is written to memory: reviewer feedback, workflow quirks, decisions that stuck.

Assembled, not accumulated

Each run gets the context it needs, built from instructions, memory, and the business state it references, instead of dragging an ever-growing transcript.

Business state is the source of truth

The center of context is business data and run records, not chat history. When the model misremembers, the system is still right.