Data Processing Policy

This Data Processing Policy describes how Achiral AI, operated by Swizzle Inc., processes Customer Content for organizations that use the Service. Terms such as Customer Content, Owner, and Service mean what the Terms of Service say they mean.

This policy is part of the Terms of Service. If your organization has not signed a separate data processing addendum with us, this policy is our data processing agreement with you. If you have signed one, it controls where it conflicts with this policy.

For Customer Content, your organization is the controller, or the business under California law, and we are your processor, or service provider. A written agreement can set different roles.

For account administration, billing, security, website analytics, and our own business operations, we act as an independent controller. The Privacy Policy covers that processing.

The model provider your organization connects is not our subprocessor. Your organization chooses it and contracts with it directly. We send data to it on your instruction, and your agreement with that provider governs its processing.

Customer Content may include prompts, messages, Outputs, memories, documents, files, embeddings, retrieval results, connector data, organization settings, metadata, audit logs, support context, and user information needed to run your organization's account.

Connector data comes from systems you authorize, such as email, calendars, documents, chat tools, issue trackers, source repositories, CRM systems, and other business applications.

The people whose data we process include your users and anyone who appears in Customer Content, such as your employees, contractors, customers, and business contacts. Do not submit health, financial, or other regulated data unless the Terms of Service or a written agreement allows it.

We process Customer Content only on your documented instructions. Your instructions are the Terms of Service, your settings and configuration, the sources you connect, your app specifications, and requests your administrators make. We process it to provide the Service, keep memory, run retrieval and inference, operate connectors, give support, secure the platform, fix problems, and comply with law.

If we believe an instruction breaks data protection law, we will tell you. If the law requires us to process Customer Content in another way, we will tell you first unless the law forbids it.

Under California law, we will not sell or share Customer Content. We will not keep, use, or disclose it for any purpose other than providing the Service and the other business purposes the law allows, or outside our direct business relationship with you. We will not combine it with personal information we get from other sources, except as the law allows. If we can no longer meet these duties, we will tell you, and you may take reasonable steps to stop unauthorized processing.

You are responsible for having the rights, notices, consents, and legal bases needed to submit Customer Content and connect third-party systems.

Your organization brings its own model. Chiro generates answers, summaries, extractions, and narratives through the model provider your organization connects, using your organization's own account with that provider.

The only Customer Content that leaves our infrastructure for AI processing is the limited data a single answer, narrative, or extraction needs: the request, the relevant excerpts from memory and connected sources, and recent conversation context. We send it, with our instructions to the model, to the model provider your organization connects, under your organization's own agreement with that provider. That agreement decides whether the provider keeps what we send.

Memories, their embeddings, and the source data your organization connects are stored and indexed on infrastructure Achiral AI operates, and so are their backups. We never send them to an AI company to be stored or indexed. We compute embeddings for memory search on models Achiral AI hosts, and we do not send embeddings or the memory index to your provider.

Models Achiral AI hosts also serve as a fallback. They process a request when your organization has not connected a provider, when your provider fails and your organization allows the fallback, or when a request needs a feature your provider cannot yet serve through Achiral AI, such as native tool calls on the Memory API inference gateway. When the fallback is used, processing stays on infrastructure Achiral AI operates.

The Owner or an Admin can turn the fallback off. With the fallback off, requests your provider cannot serve fail instead of running on our models, so your Customer Content is not used for generation on our models. Embeddings still run on our models.

We are moving organizations onto their own providers in stages, after checking that our prompts work well on each provider. Until we move your organization, Chiro's requests run on models Achiral AI hosts even if you have connected a provider.

We do not use Customer Content to train or fine-tune AI models, ours or anyone else's, unless your organization agrees in a separate written agreement.

We may use data about how the Service runs, such as counts, timings, errors, and security events, to keep it reliable, safe, and fast. When this data leaves your organization's context, we use it only in aggregated or de-identified form, and never to identify you or reconstruct Customer Content.

Each organization has its own tenant boundary for memory, retrieval, and organization data. We keep each organization's memory in separate tenant collections.

We encrypt traffic between your users' browsers and the Service. We encrypt stored credentials, including connector tokens, model-provider keys, database connection strings, and ingestion secrets, with AES-256-GCM. We control access by role, log access to memory, and back up data daily.

We run the Service, including its databases, memory store, and backups, on infrastructure Achiral AI operates, behind a network provider that protects it against attacks.

Custom plans may add controls, such as custom retention or negotiated security terms.

Your administrators control user access, connector access, and organization settings. Our staff access Customer Content only when needed to give support, keep the Service secure, investigate incidents, comply with law, or perform services you ask for.

AI Shepherds may help with setup, memory quality, connector troubleshooting, retrieval and inference quality, and account support. They see support conversations assigned to them and other Customer Content only as the task requires. Everyone who can access Customer Content is bound by confidentiality duties.

We use subprocessors for network delivery and attack protection, email delivery, and services that fetch public web pages when you ask us to read your organization's website. Our payment processor handles billing data, not Customer Content. We bind each subprocessor to data protection terms at least as protective as this policy, and we remain responsible for their work.

Our current subprocessor list is available on request at [email protected]. We will tell the Owner by email at least 30 days before a new subprocessor starts processing Customer Content. If you object on reasonable data protection grounds, your only remedy is to end the affected part of the Service before the change takes effect. We will then refund the prepaid fees for its unused term and nothing more.

We process Customer Content on infrastructure Achiral AI operates. When personal data from the European Economic Area, the United Kingdom, or Switzerland reaches us, the European Commission's Standard Contractual Clauses (Module 2, and Module 3 where you are a processor) and the UK International Data Transfer Addendum apply, and this policy forms part of them. Sections 2 to 9 and 12 supply the details those clauses require. We will sign a copy of the clauses on request.

We help you meet your data protection duties. Owners, Admins, and users can export data with the export tools in the product. We will help you answer requests from people exercising their rights, and with data protection impact assessments and consultations with regulators, where the request relates to our processing.

If someone asks us directly about Customer Content, we will send them to you and will not answer for you unless the law requires it.

On request, we will give you the information reasonably needed to show that we meet this policy, including our security documentation and third-party audit reports when we have them. Further audits can be agreed in an order form.

We keep Customer Content while your organization's account exists, until your users delete it or the Owner asks us to erase it. The Owner can set retention limits for conversations, memory, and logs. When limits are set, we delete older data each day.

Some records have fixed lifetimes. App run records and stored narratives are deleted after 90 days. Organization activity logs are deleted after 90 days, retrieval events after 30 days, product usage events after 180 days, and memory access logs after 2 years.

Ending a subscription or deactivating an organization does not erase its data. The Owner can ask us to erase the organization's data by writing to [email protected]. We erase it within the time the law allows. Conversations and memory are usually deleted within a day. When a user deletes their account, we delete their conversations and personal memory and anonymize their user record.

Deleted data can remain in backups until those backups are removed. We do not restore deleted data from backups except to recover from an incident. We may keep data longer where the law requires it, and this policy keeps protecting it.

When your organization uses Memory-Native Apps, each app processes the connected sources, organization memory, and business events its specification uses, inside the app's own organization and environment. An app specification is a documented instruction. We process app data only to run, display, store, and secure apps for your organization.

Apps also create Customer Content: specifications and their versions, projections (the entities, scores, and stages an app computes, with their evidence references), generated narratives, app run records, and action records.

To write a narrative, Chiro sends up to 20 short evidence excerpts to the model provider your organization connects, under Section 5. If the fallback applies, the narrative is written on models Achiral AI hosts instead. Chiro is instructed to treat text from your sources as data and not to follow instructions in it.

App run records hold identifiers, counts, filters, timings, and errors, not source content. They are deleted automatically 90 days after they are created. Stored narratives are deleted automatically 90 days after they are generated, and an app shows a stored narrative only while its evidence is unchanged.

Specifications, versions, release history, projections, action records, and add-on acceptance records follow Section 12. Canceling the add-on does not delete them. To request deletion, contact [email protected].

If we confirm a security incident that affects Customer Content, we will tell the Owner without undue delay and as required by law. We will describe what happened, the data involved, the likely effects, and what we are doing about it, as far as we know at the time, and we will update you as we learn more. We will help you meet your own duties to notify regulators and people affected.

Telling you about an incident is not an admission of fault.

Data processing questions may be sent to [email protected]. Security reports may be sent to [email protected]. Postal mail may be sent to Swizzle Inc., Attn: Privacy, 1908 Thomes Ave STE 71455, Cheyenne, WY 82001.