Your data, explained.
Robbie sends requests to the provider for the model you select. Muse uses Meta's Model API directly; other available routes use OpenRouter. Text sanitization is optional and off by default. This page describes what happens when you chat, attach a file, or use a tool.
Conversations and models
Your messages and relevant conversation history are processed by Robbie and the provider for the selected model to generate a response. Muse Spark 1.3 Contributor is the default and runs through Meta's Model API when configured. If the direct Meta key is not configured, Robbie uses realtime GPT 6 Luna through OpenRouter; sanitized Muse requests also use that isolated route. Luna batch is a separate choice for delayed work, not an instant chat fallback. With sanitization off, the provider receives your text as you wrote it. When the selected route accepts an attachment's input type, Robbie sends the validated original bytes with the request. Otherwise, Robbie blocks the upload or, for supported documents, tells you when it is using extracted text instead.
Robbie itself does not sell your personal information or use your conversations to train models. Model providers have their own processing and retention terms. The default Muse Contributor route is discounted because Meta may use its prompts and completions to train future models. Choose another model for material you do not want used that way. For OpenRouter-routed chat, Robbie requests routes that prohibit provider data collection for training; this is not a product-wide zero-retention guarantee. Image generation and connected services have separate processing terms. See OpenRouter's privacy policy and Meta's pricing and data-use terms.
Fast decisions and computer-work routing
Robbie uses OpenAI Luna Decisions through OpenRouter for bounded classifications, routing choices, and computer-work decisions. Depending on the feature and your existing local permissions, the request may include bounded text or an authorized image. This route uses the provider's standard processing, retention, and data-use terms; Robbie does not require zero retention or a no-data-collection provider filter for these decision calls. These provider terms are separate from Robbie's memory settings. Robbie's own memory stores and retrieves account context according to your controls; permitting provider processing does not itself create or improve that memory. A model decision cannot grant computer permissions or authorize a purchase.
Trying Robbie before signup
The public workspace lets you prepare a prompt and supported attachments before signing in. An account is required to send the request. Unsent text and attachments may be held temporarily in this browser to carry your draft through sign-in; the browser clears the carried-over draft after opening it in your workspace. Signed-in conversations follow your account retention settings.
Connected Google account data
Google access is optional. In Apps, you select Gmail, Calendar, Drive, Tasks, or Contacts capabilities before Google shows its consent screen; you can reconnect with different access or disconnect. Robbie requests only the scopes for the capabilities you select. Gmail search and read support finding relevant messages, summaries, and user-requested answers. Calendar access supports availability, event details, and reviewed event creation. Drive access supports finding and reading files you ask about; Drive file creation supports new Docs and Sheets after review. Tasks and Contacts support requested task-list reading and contact lookup. Drafting and sending Gmail messages each have separate controls and require a review of the exact action.
Robbie stores Google OAuth tokens with server-readable encryption so its server can call Google APIs on your behalf. Selected message bodies, file extracts, event details, and tool results may be temporarily processed on the server, included in an answer or a saved encrypted conversation, or used to create the personal feed and saved research you requested. Relevant Google-derived content can be sent to the selected model or embedding provider to provide the visible feature. For accounts connected to Google, Robbie routes model requests away from Meta's contributor-training endpoint and requests OpenRouter routes that deny provider data collection for training. This is not a zero-retention promise: OpenRouter and its model providers may process or retain data under their applicable terms. Google-derived content is not used by Robbie to train a general AI model. Robbie does not sell it or use it for advertising.
Disconnecting removes Robbie's future access to your Google account; you can also revoke the grant in your Google account. Disconnecting does not automatically delete prior chats, saved results, or memory containing Google-derived excerpts. Delete those records in Robbie or erase your account to remove Robbie's stored copies, subject to short-lived operational backups and legal obligations described below. Robbie's use and transfer of information received from Google APIs adheres to the Google API Services User Data Policy, including its Limited Use requirements.
Optional text sanitization
You can enable text sanitization in Settings. Start a new conversation if earlier messages were sent without sanitization. Before new text reaches the model, the browser replaces detected sensitive details with placeholders. The temporary substitution map stays in the current tab, allowing Robbie to restore those details for display. Saved conversations retain placeholders; after leaving a conversation or closing the tab, original values may no longer be available for restoration.
Detection is imperfect. It can miss sensitive details, redact useful information, or change the quality of an answer. This is a way to reduce some disclosures, not a guarantee of anonymity or end-to-end encryption.
Files and tools
Robbie accepts up to four PNG, JPEG, WebP, GIF, PDF, DOCX, XLSX, XLS, ODS, PPTX, ODP, supported text, audio, or video files per request, with a 5 MB per-file and 8 MB combined limit. It verifies the declared type against the filename and file bytes, and accepts only inline upload data; it does not fetch a URL supplied by a chat message. PDF, Word, Excel, PowerPoint and OpenDocument files are sent to Robbie's server to extract text, cells, or slide content. With sanitization on, that extracted content is processed in your browser before it reaches the model, and the original document is not sent. Plain text files are read in the browser. Image, audio, and video content cannot be sanitized, so those originals require sanitization to be off.
Robbie does not make an audio/video transcript or a lossy stand-in for an unsupported route. It preserves the original upload for an immediate model that accepts that media type and gives a clear error for a route that does not. Saved Pro reasoning work can use extracted content from supported documents; it rejects audio and video. A verified original PNG, JPEG, WebP, or GIF can be sent only to a compatible saved-work model through a short-lived signed URL; it is never converted to text.
Research mode or the web-search control processes a query through the selected route's supported search capability. Robbie uses the selected route's supported search capability. Meta Muse uses its native search when the compatible direct route is available; otherwise Robbie uses its configured compatible search route while preserving the selected Muse product model. Tool results become part of the model's context. Calculations and file-creation tools run in Robbie's application. Generated PowerPoint, Excel, Word, PDF, CSV, text, Markdown, JSON, HTML, and SVG files are built from structured content returned in the conversation or saved run. Generated files are stored with that conversation or run so they remain available for download.
Tasks, notes, Team members, images and connected apps
Tasks and decisions are saved only when you or an authorized proposal tool creates them. Notes are saved by you or an authorized proposal tool; if Learn from chats is enabled, a bounded background classifier may also create or update a profile note from an eligible new private text message, as described under What is stored. A linked text channel can save one task only when you explicitly write the documented TASK or TODO command, one reference note only when you explicitly write the documented REMEMBER or MEMORY command, and one decision only when you explicitly provide the structured DECIDE command. Ordinary texts do not create records. An owner can explicitly use the documented PROJECT command to bind one active saved outcome to that private phone thread; Robbie stores only the owner-scoped conversation-to-outcome pointer. Ordinary replies and an explicit WORK or THINK request may then use a fresh bounded packet of that outcome and its linked work. Explicit TASK, REMEMBER, and DECIDE captures in that thread may link their resulting owner-authored record to the same outcome. PROJECT: CLEAR removes the pointer and this inheritance. This ordinary text and project context does not grant browsing, files, connected-app access, or external actions. Separately, an explicit MAC: command from a linked phone can queue a task for a separately linked, ready Mac that has remote execution and Local Control enabled. That Mac may use its locally granted permissions. The phone command and Mac answer are stored with server-readable encryption; a person controlling the linked phone line can request that local work while the Mac is ready. Owner-selected review dates for active Decision Ledger records can create one generic in-app reminder; this does not include decision contents, invoke a model, or take an action. By default, Robbie selects a bounded relevant set of your saved notes for new chats and sends at most three matching notes as reference data to the selected model provider. You can pause this in Memory preferences at any time. Conversation continuity is a bounded local recall path that is on by default for new workspaces: Robbie can decrypt a bounded set of your own earlier chats, select up to two matching conversations, and send at most two short excerpts from each as reference data with a new request. You can pause it in Memory preferences. Its normal local matching does not use an embedding provider or create a plaintext transcript index. Semantic conversation memory is a separate meaning-based recall path that starts on for new accounts; existing accounts keep their previous choice. When its embedding route is available, Robbie sends bounded owner-authored text from up to 60 recent conversations for a new account, or up to 240 after an explicit expansion or earlier broader choice, to OpenRouter’s embedding service in small background batches. Ordinary recent-chat indexing uses up to eight separately searchable chunks per conversation from its latest 80 messages; explicit deep indexing covers more retained turns. It also sends each retrieval query there. Converted attachment text and chats with sanitized owner turns are excluded. Robbie requests an embedding route that denies provider data collection. The resulting vectors are encrypted at rest. You can erase this index independently; pausing saved memory or conversation continuity also pauses new indexing and recall. Turning either feature off stops its future recall without deleting your conversations. Semantic memory is enabled by default for new workspaces. When its embedding route is available, saved-note text, active goal outcomes, and the owner-authored profile text of People records are sent to OpenRouter's embedding service to create a meaning-based index, and the retrieval query from each chat or personal-agent launch is sent there to find related context. People contact receipts are not included. Durable agents receive at most three relevance-selected note excerpts in a fixed launch snapshot, never a live memory feed. You can erase the entire semantic index in Notes; turning it off removes those stored vectors. Memory preferences can pause automatic note recall, personal profile, anchors, decisions (including saved assumptions and review dates), and continuity recall without deleting the underlying records. Linked conversation goals and selected References remain explicit context choices, and deliberately enabled working agreements remain separate from automatic recall. Saved notes and earlier excerpts are treated as data, never as instructions or authorization. This retrieval is not processed by browser text sanitization. A selected Team member receives an explicit task brief and returns their result to the main conversation.
If you explicitly choose Preview saved memory for a typed draft, Robbie sends only that current draft and its chat identifier to the owner-checked recall route. The short-lived receipt reports saved titles and retrieval reasons, never saved-record bodies, and clears when the draft changes. It does not include attachments, connected-app data, Team instructions, tools, or a future answer. The preview uses the same bounded recall policy as chat; if semantic memory is enabled, the typed retrieval query can be processed by the embedding service described above. The preview is unavailable in guest, sanitized, and saved-agent chats and is not stored.
An optional Gmail or Outlook action review sends up to five messages that you selected from one saved search to the model as untrusted reference data, then encrypts the resulting review. It does not save a task until you explicitly choose a proposal to add; its source fingerprint prevents duplicates on retry. Image prompts are sent through OpenRouter’s image API to the selected image provider. Generated images are stored in your account. Connected MCP services receive the tool arguments you approve and return data that can become model context. Their processing and retention policies are independent of Robbie’s. Remote MCP credentials and stored tool results are encrypted using the application-held key. Every connected tool action requires approval; disconnecting revokes future access through Robbie.
iCloud Calendar is a read-only CalDAV connection. If you choose it, Robbie accepts your Apple Account email and an Apple app-specific password, never your Apple Account password. The service stores that connection with server-readable encryption, reads only the calendars you select, and has no event-write path. You can disconnect it in Apps and revoke the app-specific password through Apple.
Execution detail
For signed-in chat, Robbie can show an owner-visible execution receipt: the provider and model reported for each response step, response timing, reported token, cache, and cost fields, tool names and timing, plus bounded redacted app-level tool request and returned data. It can also display provider-visible reasoning only when a provider deliberately sends a visible reasoning part. This is not a promise that every route exposes reasoning, and it is not hidden chain-of-thought.
If you explicitly choose Check answer on a completed signed-in reply, Robbie sends only the visible request text and that reply to its independent review route. It excludes attachments and their extracted text, chat history, saved memory, connected-app data, tool inputs and outputs, provider traces, and hidden reasoning. The advisory result is held only in the active browser reply; it does not change the answer, authorize an action, or enter chat history or memory. The control is unavailable for guest and sanitized chats. JEV routing remains a separate typed routing mechanism; it is not an answer reviewer.
Robbie does not retain or display raw provider HTTP request or response envelopes, headers, credentials, full prompts, saved-memory packets, attachment bytes, or hidden reasoning as an execution trace. The copyable receipt is assembled only from the bounded fields visible in the reply.
What is stored
Google sign-in supplies account information such as your name and email. Robbie stores conversation history, tasks, notes, decisions, generated images and app connection settings so you can use them later. Conversation bodies, original attachments, and the contents of tasks, notes, decisions and images are encrypted at rest using an application-held key; the service can decrypt them. Semantic-memory vectors are also encrypted at rest while meaning-based recall is enabled. Titles, account metadata, and limited operational indexes are available to the application. Browser preferences include the selected model, theme, and whether you explicitly enabled sanitization. If you deliberately link an SMS or WhatsApp number, its address, delivery receipts, and any queued background-result delivery are encrypted at rest, and inbound, reply, and delivered result text is retained in that owner's encrypted private conversation; the service can decrypt it to deliver the feature.
If you use the installed app's Share with Robbie target, its title, text, link, and any supported shared file are first held as one encrypted, owner-bound draft for up to 15 minutes. The redirect carries only an opaque ID, never the shared content. The draft is removed when your signed-in Robbie page opens it or when it expires; it is not sent to a model until you review or edit the resulting composer text and attachments and choose Send.
Automatic memory from private chats
When Learn from chats is on, a background classifier may process one newly saved, eligible owner-authored text message and up to five short existing profile-memory excerpts to decide whether the message expresses a durable preference or fact, including a basic name, relationship, or role you directly state about someone in your life or work. It does not learn from connected-app data. The classifier uses OpenRouter's OpenAI GPT 6 Luna route, with a request to deny provider data collection and disable provider fallbacks. This is a provider setting, not a zero-retention guarantee. A qualifying result can create or update a profile memory that may be recalled in future conversations without another confirmation. The classifier does not backfill older conversations. It excludes sanitized messages, attachments and converted file text, explicit requests not to remember, and text that matches Robbie's limited checks for sensitive material; those checks can miss things. Accounts without a saved learning choice follow the Use memory in chat setting, which is on by default. You can turn learning off in Memory controls. Robbie shows learned updates in the originating conversation; an update can be undone while it remains eligible for undo, and profile memories can also be reviewed or deleted in Memory.
Conversation continuity is a separate local retrieval capability, not a learned profile; you can pause it at any time. Memory preferences also control whether saved notes may be used as bounded reference material in a new request. Semantic memory has its own erase control. Previously saved account data is retained until you delete it or an existing retention preference removes it. Operational backups can retain earlier copies until their retention period expires.
Completed Mac task results launched from a saved web conversation may be included as short prior-work context in a later reply in that same conversation when saved memory is on. If you also enable conversation continuity, Robbie can locally search a bounded set of your older completed Mac tasks and include at most two relevant request-and-answer excerpts from other conversations. It does not send Mac task text to the semantic embedding service for this search. These excerpts are treated as unverified previous assistant output, not proof of current files or permission for new actions. Sanitized chats and chats with saved memory off do not receive them.
Saved agents and background runs
Agent configurations, including an optional owner-reviewed persistent playbook, task briefs, model context, tool results, generated files and completed run outputs are stored with server-readable encryption. A playbook is saved with the agent and snapshotted into each run as task guidance; Robbie does not silently update it and it never authorizes an external action. A saved Pro reasoning run stores accepted original attachments in its encrypted checkpoint so a worker can rebuild the same model input after a restart. A selected team receives the task brief; its lead receives the other agents' reports to combine them. Runs continue after you close the browser. Saved agents can use calculation and optional web search. An agent can also explicitly select up to three enabled personal skills; their exact text is captured in the encrypted run checkpoint at launch, rather than read live by the worker. Saved Pro reasoning can also create a file from a validated structured specification after the provider result completes. Runs do not execute connected-app actions. The chat sanitization setting does not filter agent instructions or background-run briefs.
A temporary provider-hosted file workspace receives only the files and instruction explicitly admitted to that run. It has no saved browser, login, connected-app access, or Robbie-memory access. It is owner-scoped and short lived, but it is not end-to-end encrypted from the selected provider: the provider processes the admitted files and instruction under its applicable terms. The separate managed-computer design is not the public execution path.
You can cancel a run and delete its saved history. A request already sent to a provider may still be processed or billed. If a worker restarts before saving a response, that request may be repeated. Run status, ownership, timestamps and short-lived usage-limit records are operational metadata.
Agent network and marketplace
Private independent-agent credentials are encrypted at rest with an application-held key; Robbie's server can decrypt them to use the connection. This is not end-to-end encryption. Connection verification checks a supported, reachable endpoint, not the agent's skills or example quality. Other agent operators have their own processing and retention practices.
When you authorize an outside client, its scoped grant can submit bounded tasks and read their outputs. Those outputs can include the saved worker's configured context, such as selected memory, projects and skills. A research-to-document workflow sends the research peer your brief and acceptance criteria; the document worker receives a saved snapshot of the peer's evidence plus its configured context. Source references and a completed task do not establish that the source claims are true.
Publishing makes your selected profile fields and optional public setup-guide URL visible in the marketplace. Private connection addresses and credentials are never included automatically. Readers set up their own access, including any required credentials. Unpublishing hides the public listing. Disconnecting removes future access through that local connection; it does not cancel work already running at a remote agent.
Billing
Card entry, payment-method updates, invoices, and subscription management are handled on Stripe-hosted pages. Robbie does not receive card numbers. To connect an account to a Stripe customer and apply subscription access, Robbie stores the account identifier, Stripe customer and subscription identifiers, plan and subscription status, trial and billing-period dates, and signed webhook-processing records. Stripe processes payment information under its own privacy terms.
Your controls
Delete individual conversations from History, tasks, notes and images from their workspace pages, saved agents from Agents, and finished or cancelled runs from Runs. Disconnect apps from Apps. In Settings, you can export your account data, including agent configurations and saved run results, or delete your account. These actions affect Robbie's stored data; they do not undo processing that has already happened at an external provider. Anyone holding an existing conversation-sharing link may be able to read that shared conversation until the link or conversation is removed.
Contact
For privacy questions, contact the Robbie privacy team at privacy (at) robbiebot (dot) net.
Chrome reading-interests preview
The optional Robbie Reading Interests Chrome preview is not installed automatically and does not watch browsing in the background. If installed and enabled by clicking Review this profile locally, Chrome asks for its optional History permission. It processes at most 500 rows from the previous 30 days in that Chrome profile. It recognizes a fixed list of public source domains and Reddit community paths. For common search-engine result pages, it locally reduces the search phrase into topic suggestions and filters some health, financial, legal, and personal-identifier patterns. These filters can miss sensitive searches. Search suggestions start unchecked. The extension does not inspect page titles, save history rows, or make network requests. Robbie receives only the source categories and topic labels you select in the extension; you review them again in Robbie before saving. Robbie stores confirmed labels with server-readable encryption. The background feed applies them locally to shared public news; an explicitly requested What changed brief may send them to its selected model provider. Optional city or region labels are never collected from device location; address-like labels and postal codes are rejected, and valid areas are included only in an explicitly requested What changed search. Shared public news also includes a bounded Hacker News front-page slice cached for all owners; story scores and comment counts are discussion context, not fact checks. Raw URLs, page titles, and unselected query text are not sent to Robbie.