loading
loading
A working-memory space where the agent writes intermediate notes it can read back during a task.
The scratchpad is a designated chunk of context (or an external file/tool) where the agent jots reasoning, partial results, and to-dos as it works, then re-reads them to stay on track across many steps. It externalizes state so the agent doesn't have to hold everything in its head, which is what keeps long multi-step tasks coherent. Example: an agent solving a 12-step refactor keeps a running checklist on the scratchpad, ticking items off so it never loses its place.
Plainly
Think of Scratchpad as a simple recipe for doing the work better. A working-memory space where the agent writes intermediate notes it can read back during a task.
In practice
Use it when you need a repeatable method instead of guessing from vibes. In practice, define the owner, input, output, and failure mode before you rely on it.