Know what happens to your work.

The desktop app stores project files on your computer by default, with conversations and related records in a local database. Projects can also use local folders. Importing a file does not automatically upload the whole project to a model; automatic AI tagging after import is disabled by default in the desktop configuration.

Storage and AI processing are separate steps. Cloud AI, mobile transfers, shared team features, remote sync and online tools may send relevant content off your device when you use them. The desktop app also keeps local conversation history and runtime logs.

The iOS app keeps originals in its local application directory, which the current implementation excludes from system backups. Cloud image delivery temporarily stores a relay copy. After the computer acknowledges receipt, the service attempts to delete that file content early, while keeping its delivery status record.

The current code sets a 30-day expiry threshold for remaining image deliveries, processed by a daily cleanup job. This is not a guarantee of immediate, complete deletion at exactly 30 days. Storage deletion failures require later cleanup or an object-storage lifecycle fallback. The phone original, received computer file and cloud relay copy are managed separately; deleting one does not delete every copy.

No. In the current code, staged downloads supplied by the computer to the phone (PULL) have a 7-day expiry threshold; staged uploads (PUSH) have a 30-day threshold. Cleanup jobs process expired transfers. These are relay rules, not a backup service: confirm that important material has been saved on the receiving device.

No. Connecting to a local service and connecting to a cloud service are different paths. Cloud Office add-in conversations and sync data are stored on the server. A sync acknowledgement can remove the delivery copy without deleting the underlying conversation. The website’s Skill drafting assistant also saves conversation content.

Shared team and firm memory is stored on the server. Configuring remote sync, such as Git, can create remote copies too. Accounts, orders and usage records are online service data. We therefore do not describe the entire Community Edition as a zero-retention service.

Keep recoverable copies of local files, project databases and personal or project memory before migrating or cleaning up. Signing back into the website is not the same as restoring everything from your old computer.

Deleting local files, shared cloud content, a website account and external sync copies are separate operations. Mobile relay storage is not a long-term backup. Check diagnostic logs for case content before sharing them.

AI WorkDeck uses existing model APIs and does not train its own foundation model on your case material. Community Edition memory records relevant facts, preferences and experience as readable Markdown content for later context. It does not train or change model weights.

This statement concerns AI WorkDeck itself. OpenRouter, model providers and other tools process requests under their own policies, endpoints and actual privacy settings. No training and zero retention are different promises; this does not guarantee that every external service discards all content.

Personal and project memory is stored as Markdown content and indexes in the desktop app’s local database, with read, edit and delete operations. A memory document is not necessarily a separate .md file on disk. Enabling relevant Git sync can also save it to your configured remote.

Team and firm memory is shared through the website and stored on the server. Team managers control team memory; managers in the head team control writes to firm memory. Read access follows organization membership. Check that anything you add is appropriate for everyone in that scope.

Review, correct or delete the memory in the relevant scope. Keep case-specific information in project memory instead of publishing it to the team or firm. Memory is context for a model and can still be summarized or applied incorrectly. Check the facts and rules being used for important work.

Community Edition can be assessed for self-hosting, with storage and models running on the firm’s own servers. Configurable local-model integration exists, but tool use, vision and other capabilities need validation for the chosen model. Do not assume full parity with the default cloud features.

To keep a particular workflow’s material inside the firm, all relevant storage, models and processing services must run there. External models, cloud OCR or transcription, web search, mobile cloud relays, website shared memory, remote sync and other outbound connections must be disabled or replaced. Buying a server or switching only the model is not enough to guarantee a fully isolated application.

Contact the developer with your deployment size, model, required features and network restrictions to assess a suitable setup. This concerns Community Edition deployment, not the Enterprise Edition product.

The model needs context to perform the task. Your prompt, relevant conversation, file content you supply or a tool reads, and memory used in that request can be included. The request is not necessarily just a short instruction, nor does it automatically include your entire project.

The desktop platform channel uses an allocated key and allowance to call OpenRouter directly, which routes the request to a model provider. The website manages accounts and allowances; it does not proxy every desktop request. Other channels, add-ins and online tools follow their own configuration.

For ordinary desktop model calls, the website is not in the request-body forwarding path. However, the desktop app saves local conversations, and some logs may contain response fragments. Cloud Office add-in conversations, website Skill drafting, shared memory and mobile relays have their own storage behavior.

“No retention anywhere” would therefore be inaccurate. External model providers’ retention and training policies also need separate review; Community Edition cannot make one zero-retention promise for every supplier.

Separate local file work from AI or tool actions that connect to an online service. Check what the particular action sends, where it goes and which policies apply. Supply only what the task needs, removing unnecessary identifying information where appropriate.

A local file is not permission to send its contents to a cloud model. Follow the client’s agreement and your organization’s requirements when deciding what may leave the device. Local storage alone does not guarantee confidentiality or compliance across the entire workflow.

Team membership should not be understood as automatic sharing of your entire local work directory. Shared memory is available to the corresponding members, and team dashboards show usage information allowed by roles and settings. Sharing project names has a separate setting.

Keep invitations, management permissions and shared content limited to what is needed. Do not put one client’s confidential details into general memory that the whole team can read.

Verify them against the original source first. Check jurisdiction, publication and effective dates, the applicable version, case number, court and whether the quotation supports the conclusion. Models can invent sources or misapply real ones.

Use AI output as research and drafting material. The people handling the matter remain responsible for reviewing the facts, applicable law and final work product.

Check names, amounts, dates, negations, clause numbers, attachments and tables spanning pages. Recognition and summarization can omit details, and a model may not have the entire document in context.

Keep the original and compare edits. Manually review changes affecting rights or obligations, quotations and conclusions; a summary should not replace the underlying evidence.

Understand the files it reads, the services it contacts and the actions it can perform. Third-party tools introduce their own recipients and processing rules. Documents and web pages can also contain instructions that mislead a model.

Give trusted tools only the access they need. Check the target and result of important writes, deletions or external submissions. Do not post client material, keys or passwords to the public community for troubleshooting.

The mainland site, aiworkdeck.com, has a separate account and online-data system from the overseas sites. workdeck.ai and aiworkdeck.us share the overseas user-data system.

Connect the desktop app to the service region where you registered and topped up. Changing the website language does not move an account, balance or local project.