Cua Driver Adds Privacy-Conscious Cross-Session Action History
- Source
- Cua
- Date
1/ Today, we're releasing the first open-source Computer History - now in early preview for Cua Driver on macOS, Windows, and Linux. It gives agents an encrypted, local record of actions they took through Cua Driver, so new sessions can recover useful context from earlier work.
2/ An agent's chat history helps within one conversation. Computer History survives a reset, a new session, or a different compatible agent. The agent can query which apps and Cua capabilities were used and whether actions succeeded. It still asks when the user's intent is missing.
3/ Test prompt on macOS: “Move the white pawn from e2 to e4.” With history off, the agent tried three routes that failed. With history on, it reused the successful route, completed the task with 33.3% fewer actions, and had no failed routes.
4/ Test prompt on Windows: “Continue the prior Paint task after context was reset.” Computer History gave the agent app attribution, the prior action sequence, and confirmed outcomes. It did not store the drawing, screenshots, typed text, raw arguments, titles, or paths.
- Computer History gives agents an encrypted local record of Cua Driver actions that survives a reset, a new session or a different compatible . Agents can query which apps and capabilities were used and whether actions succeeded. It is off by default.
- In Cua's macOS test prompt to move a chess pawn from e2 to e4, the agent without history tried three routes that failed. With history on it reused the working route, used 33.3% fewer actions and had no failed routes. This is one demo prompt, not a .
- It records fixed metadata only: session lifecycle, Cua actions, target app identity, outcomes and writer health. It excludes screenshots, typed text, keystrokes, clipboard contents, window titles, URLs, raw arguments and results, accessibility trees and paths.
- History is encrypted before it reaches disk using CBOR Sequence, COSE_Encrypt0 and CloudEvents, with the key held in the OS credential store. Cua says no history enters product telemetry, and users can pause recording, disable it, or delete the history and its key.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
Gives computer-use agents cross-session memory of what worked without storing sensitive screen content, a concrete privacy/utility tradeoff worth copying in other agent memory designs.
Checking sign-in…
Loading comments…