Here's what Quick Desktop told me:
Amazon Quick for Desktop — Data Storage & Non-Persistent VDI Guidance
Prepared for: Partner deployment planning
Context: Non-persistent VDI on Amazon WorkSpaces Applications, using Application Settings Persistence (up to 5 GB profiles)
Date: 9 September 2026
---
## 1. Where Amazon Quick keeps user data
Amazon Quick for desktop is architecturally a thin client over a cloud runtime. The agent itself runs in the Amazon Quick cloud runtime; the desktop app acts as a lightweight local "outpost" that executes local-machine work (file access, browser automation, screenshots, notifications) only when a task requires it.
The practical consequence for a VDI deployment:
| Data category | Where it lives | Persists per-user profile? |
|---|---|---|
| Conversation history | Cloud runtime | No — server-side |
| Long-term memory | Cloud runtime | No — server-side |
| Knowledge graph | Cloud runtime | No — server-side |
| Activity feed | Cloud runtime | No — server-side |
| App settings & preferences | Cloud runtime | No — server-side |
| The user's own local files | Stay on the machine; read only on demand, never copied into Quick's own store | N/A |
| Local cache / config / logs | Local app data | Yes (disposable) |
| Optional local LLM model | Local app data | Yes (large — see §3) |
Because conversation state, memory, knowledge graph, feed, and settings are keyed to the corporate sign-in identity and stored in the cloud, they re-hydrate automatically on whatever machine the user next signs into. This is precisely why the same account and conversations follow a user across the desktop, web, and mobile apps.
Bottom line: the data a user would care about preserving is not held in the local profile — it comes back on sign-in. In a non-persistent model this is favorable: you generally do not need to persist Quick's local profile data to preserve the user experience.
---
## 2. Per-machine installation (recommended for this scenario)
Amazon Quick for desktop now supports per-machine installation, which places the application binaries in a machine-wide location instead of inside each user's profile.
For non-persistent VDI this is the right choice:
- The application itself no longer lands in each roaming/persisted profile, reducing the per-user profile footprint.
- The app is present on the image for every session, so there's no per-user (re)install cost at logon.
---
## 3. The one footprint lever to watch: Local Processing (local LLM)
Amazon Quick has an optional Local Processing feature that downloads a local LLM for offline inference. When enabled, that model download is multi-gigabyte and lands in local/app data.
For a non-persistent deployment with ~5 GB profiles, you should ensure Local Processing is disabled so a large model isn't being downloaded into (or attempting to persist within) each user profile. With it off, the residual local data is just cache, config, and logs — which is well within a 5 GB Application-Settings-Persistence budget and can safely be treated as disposable.
---
## 4. Recommended configuration summary
For non-persistent VDI on Amazon WorkSpaces Applications with Application Settings Persistence:
1. Use per-machine install — keeps binaries off the per-user profile.
2. Disable Local Processing — avoids a multi-GB local model competing for profile space.
3. Do not treat Quick's local profile data as something that must be persisted — user state (history, memory, knowledge graph, settings) is cloud-side and restores on sign-in. Local cache/config/logs are disposable.
4. Existing 5 GB Application Settings Persistence profiles are more than adequate once the above are in place.
---
## 5. Caveat — for an exact byte figure
The in-product guidance does not publish a specific number for the desktop app's local user-profile footprint (cache/config/logs), nor the exact per-user Windows paths (AppData\Local vs Roaming). For a precise sizing figure and path list, that's a packaging/deployment question best confirmed with the Amazon Quick / EUC deployment team. The architectural design above, however, means an exact number is unlikely to be a blocker: with per-machine install and Local Processing off, the persisted-per-user data is small and non-critical.