Legal

• Identity and Contact Data: We collect your name and email address when you create a CATO account. We may also collect or generate indirect identifiers (e.g., internal user IDs).
• Payment Information: If you purchase access to paid features, we collect payment information through our payment processor. CATO does not store full payment credentials on its servers.
• Inputs and Outputs:When you interact with CATO, your messages ("Inputs") generate AI responses and analysis results ("Outputs"). If you include personal data in your Inputs, that information will be processed as part of the conversation and may appear in Outputs.
• Uploaded Datasets: Files you upload for analysis, including their contents, filenames, and structural metadata (column names, row counts, data types).
• Literature Search Data: Search queries you submit through Literature Intelligence, which are transmitted to external academic databases (PubMed, PMC, Europe PMC, Unpaywall, CORE, Crossref, OpenAlex, DataCite, Semantic Scholar, arXiv, bioRxiv, medRxiv, ChemRxiv, SSRN, Research Square, OSF, Zenodo, HAL, ClinicalTrials.gov, and patent databases). Also includes literature collections you create, annotations you add to papers, saved watch alert queries, and papers you send to Chat via the literature-to-chat bridge.
• Memories: A small, user-controlled index of durable context, such as preferences, standing instructions, project definitions, constraints, and decisions. Automatically created memories must be grounded in a user message. Generated scientific findings are not automatically promoted to memory. Each memory records its scope and source so you can review, edit, pin, or delete it. When automatic memory is enabled, CATO reviews changed conversations in a daily batch. The batch contains human-readable conversation excerpts and existing memory text, but excludes uploaded files, figures, binary artifacts, recognized credentials, and sandbox storage access.
• Feedback and Communications: If you submit feedback or communicate with us, we collect the contents of those messages.
• Service and Agent Trajectories: We record prompts, responses, tool calls and results, generated code, citations, reviewer results, execution metadata, safety signals, provider-returned reasoning content where available, and User Content surfaced in those records during and after the beta. Uploading a file does not, by itself, place the full raw file in a trajectory. File excerpts, selected rows, summaries, tool results, generated outputs, and other content surfaced to the agent may be recorded.

When you use the Services, we receive certain technical data automatically, including device type, browser information, IP address, and usage information such as dates and times of access.

Google Cloud provides CATO's authoritative object storage for user files, project artifacts, operational trajectories, and training-approved trajectories. Railway hosts the application, PostgreSQL account, conversation, and project metadata, Redis cache and stream-replay data, Temporal workflow state, service logs, and a transient local trajectory delivery spool. The Railway spool is not the authoritative trajectory store. Storage may also include encrypted backups and security or error-monitoring systems as described in this Policy.

A trusted Google Cloud controller stages only the input files needed for an analysis into a temporary sandbox workspace and retrieves validated outputs afterward. Sandbox pods do not receive cloud-storage credentials, application secrets, or mounted object storage. Staged copies are removed when the sandbox pod or session is cleaned up. Uploading a file does not automatically duplicate its full raw contents into an agent trajectory, although content surfaced through file inspection, tools, analysis, or generated outputs may be recorded there.

Operational trajectories and training-approved trajectories are stored separately. The production writer can create operational records but cannot list, read, overwrite, or delete them. A restricted cleaning and approval pipeline may read operational records and create approved copies in the training-approved store. Training systems can read only the training-approved store and cannot access the operational store directly. This separation prevents an operational record from becoming training-readable without the intervening cleaning and approval step.

Primary account and project data is retained while your account is active or as needed to provide the Services. You can delete conversations, memories, datasets, or projects in the application. Those controls remove the primary application copy but may not immediately remove security, diagnostic, trajectory, or backup copies.

Identifiable trajectories are retained as reasonably necessary to operate and restore workflows, debug and evaluate CATO, prevent misuse, investigate security events, resolve disputes, enforce our policies, and improve our offerings. Records connected to abuse, security, fraud, disputes, or legal obligations may be retained longer when reasonably necessary for those purposes.

Contact contact@nyrus.ai to request deletion of identifiable trajectory records or to request that identifiable records no longer be selected for future model-improvement datasets. A request applies prospectively once processed and does not undo completed processing. We may retain deidentified or aggregated evaluation and training data where permitted by law. Data already included in a training run that is underway, and improvements already incorporated into trained model weights, cannot generally be traced back and selectively reversed.

CATO can save durable context that you state in a conversation and recall only a relevant subset in later conversations. Global memories can apply across projects; project memories remain within their project. Memory is not a replacement for conversation history and is not treated as scientific evidence. Current instructions override older memories. Automatically saving new memories, recalling saved memories, and cross-project recall have separate controls. Memories are stored in CATO's application database and can be viewed, added, edited, pinned, or permanently deleted from Settings > Memory.

AI model infrastructure is upgraded periodically; your data may be processed by updated model versions.

Literature full-text content is retrieved on demand. A paper may be cached for an active workflow and is stored with the project when you or CATO fetches or saves it to the workspace. Project copies remain until you delete the paper or project, subject to backup and trajectory retention described above.

The hosted Service uses the providers below. We limit each provider to the data needed for its function. We do not authorize AI providers to train their models on your User Content. Retention exceptions may apply for security, abuse prevention, legal obligations, and data that you explicitly ask a provider-backed feature to store.

We implement appropriate technical and organizational measures to protect your data, including encrypted credential storage, sandboxed code execution, controls that block sandbox access to internal cloud networks and metadata services, credential-free temporary sandbox file staging, separate user-file, operational-trajectory, and training-approved storage identities, and authentication-gated API access. Sandboxes may access public HTTP and HTTPS destinations when an analysis, package installation, or tool requires it. Data sent by code to a public destination is subject to that destination's terms. No security system is impenetrable, and we cannot guarantee absolute security.