Feature Request: Persistent Underlord (AI Agent) memory per project

Currently, when I reload Descript, the Underlord agent loses all memory of the actions it previously took and the context of our conversation within that project. This makes it difficult to:

Understand or learn from the steps the agent took to reach a specific result

Continue a complex, multi-step editing conversation after a reload

Audit or reproduce agent actions after the fact

Requested improvement: Persist agent conversation history and a record of actions taken per project/composition, so that reloading the app (or returning to a project later) doesn't erase that context. Even a lightweight action log tied to the project would help users understand what was done and why.