# Iris Privacy Guide

This guide explains how to control privacy in Iris: what data may be processed,
how memory and knowledge are managed, when local-only mode helps, how deletion
works, and which settings to review before using Iris with sensitive material.

Privacy behavior is partly product-defined and partly workspace-defined. Your
organization may set stricter defaults, disable some options, require approval
for exports, or apply legal retention rules.

---

## Table of Contents

1. [Privacy Controls](#privacy-controls)
2. [Data Management](#data-management)
3. [Local-Only Mode](#local-only-mode)
4. [Data Deletion](#data-deletion)
5. [Data Collection](#data-collection)
6. [Encryption and Security](#encryption-and-security)
7. [Compliance](#compliance)
8. [Privacy Settings Reference](#privacy-settings-reference)
9. [Frequently Asked Questions](#frequently-asked-questions)
10. [Getting Help](#getting-help)
11. [Related Documentation](#related-documentation)

---

## Privacy Controls

### Open Privacy Settings

Use **Settings > Privacy & Security** to review privacy controls. Administrators
may also expose team-level privacy settings in the dashboard.

Before sharing sensitive content, check:

- Which model providers are allowed.
- Whether local-only mode is available.
- Whether conversation history is saved.
- Whether memory writes are enabled.
- Whether knowledge uploads are private, team-scoped, or organization-scoped.
- Whether export or deletion requests require administrator approval.

### Consent Management

Consent settings control optional processing. Essential processing remains
required for basic service operation.

| Consent area             | What it controls                                                       | Recommended review point                           |
| ------------------------ | ---------------------------------------------------------------------- | -------------------------------------------------- |
| **Essential processing** | Authentication, routing, message handling, safety checks, core logs.   | Required for using Iris.                           |
| **Usage analytics**      | Product and reliability metrics, usually aggregated or anonymized.     | Disable if your role handles sensitive workflows.  |
| **Personalization**      | Preferences that tailor tone, format, model, and interface behavior.   | Keep enabled only if personalization is useful.    |
| **Long-term memory**     | Durable facts and preferences reused across conversations.             | Enable only for facts you are comfortable storing. |
| **Third-party models**   | Processing by approved cloud AI providers.                             | Disable when cloud processing is not acceptable.   |
| **Feature previews**     | Experimental features, models, or workflows.                           | Avoid for regulated or high-sensitivity work.      |
| **Team sharing**         | Whether selected knowledge, memories, or outputs can be shared by you. | Confirm audience before enabling.                  |

### Memory Controls

Memory controls determine what Iris may remember and where that memory can be
used.

| Control                  | Use it to                                                                                            |
| ------------------------ | ---------------------------------------------------------------------------------------------------- |
| **Disable memory**       | Prevent persistent memory writes for future conversations.                                           |
| **Selective tiers**      | Enable only short-term, working, long-term, episodic, semantic, core, or archival memory as allowed. |
| **Review memories**      | See stored memories, source context, confidence, tags, and related conversation.                     |
| **Edit memory**          | Correct outdated or inaccurate stored facts.                                                         |
| **Delete memory**        | Remove specific memories or clear a full tier.                                                       |
| **Agent memory access**  | Control whether specialist agents can search or update memory.                                       |
| **Private conversation** | Start a conversation that does not update persistent memory.                                         |
| **Memory export**        | Download memory data when policy permits.                                                            |

Use explicit instructions when memory matters:

```text
Remember that I prefer UTC timestamps and TypeScript examples.
```

Use a boundary when memory should not be stored:

```text
Use this only in the current conversation. Do not save it to memory.
```

### Conversation Privacy

Conversation privacy settings control how chat history is stored and reused.

| Setting                    | Behavior                                                                              |
| -------------------------- | ------------------------------------------------------------------------------------- |
| **Conversation history**   | Saves conversation content so you can return later and search history.                |
| **Auto-delete**            | Deletes conversations after a configured duration such as session, 7, 30, or 90 days. |
| **Private/incognito mode** | Avoids persistent history and memory updates when enabled by workspace policy.        |
| **Export controls**        | Limits who can export conversations and which formats are available.                  |
| **Sharing controls**       | Controls whether links, exports, or team-visible threads can be created.              |
| **Search visibility**      | Determines whether conversations appear in workspace or team search.                  |

### Knowledge Privacy

Knowledge base privacy depends on ownership, permissions, and source material.

- Upload sensitive documents only to knowledge bases with the right access
  scope.
- Separate personal, team, customer, and regulated material into different
  knowledge bases.
- Use tags and metadata for retention, source, version, and owner.
- Remove obsolete or superseded documents.
- Confirm whether agents can search a knowledge base before asking them to use
  it.

---

## Data Management

### View Your Data

Depending on workspace policy, you can review:

| Data area             | Where to look                                      | What to inspect                                             |
| --------------------- | -------------------------------------------------- | ----------------------------------------------------------- |
| **Account profile**   | **Settings > Account**                             | Name, email, role, workspace, authentication method.        |
| **Conversations**     | Conversation list and search                       | Messages, attachments, citations, exports, deletion state.  |
| **Memory**            | **Settings > Privacy & Security > Memory**         | Stored facts, preferences, source links, tags, confidence.  |
| **Knowledge bases**   | **Knowledge** panel                                | Documents, collections, owners, tags, versions, visibility. |
| **Model settings**    | **Settings > Models** or **Privacy & Security**    | Allowed providers, local-only mode, routing preferences.    |
| **Audit information** | Dashboard or administrator-managed compliance view | Access records, exports, policy changes, security events.   |

### Export Your Data

Exports may include conversations, memory, knowledge metadata, uploaded
documents, account settings, and preferences. Available formats are configured
by the workspace and may include JSON, Markdown, PDF, or a ZIP archive.

Before exporting:

1. Confirm you are allowed to export the selected data.
2. Filter out conversations or documents that contain unrelated sensitive
   content.
3. Review generated summaries and citations for accuracy.
4. Store the export in an approved location.
5. Delete temporary export files when no longer needed.

### Edit or Correct Data

Use direct controls when available:

- Edit account profile fields in **Settings > Account**.
- Rename, retag, archive, or delete conversations from the conversation list.
- Edit or delete individual memories from the memory review screen.
- Replace stale knowledge documents with newer versions.
- Ask Iris to correct a stored preference, then verify the memory entry.

Example:

```text
Correct your memory: I use PostgreSQL for this project, not MySQL.
```

### Retention Policies

Retention is configured by the deployment. These categories are typical:

| Data category         | Typical retention behavior                                                     |
| --------------------- | ------------------------------------------------------------------------------ |
| **Conversations**     | Retained until deleted, auto-deleted, or expired by workspace policy.          |
| **Short-term memory** | Session or conversation lifetime.                                              |
| **Working memory**    | Task/session lifetime unless promoted to a durable memory tier.                |
| **Long-term memory**  | Retained until deleted or expired by policy.                                   |
| **Knowledge content** | Retained while the document or knowledge base exists.                          |
| **Usage analytics**   | Aggregated or anonymized on a shorter operational retention schedule.          |
| **Audit logs**        | Retained for security, compliance, and abuse-prevention requirements.          |
| **Backups**           | Follow backup retention and legal hold policy; deletion may age out over time. |

Ask your administrator for exact retention windows in your workspace.

---

## Local-Only Mode

### What Local-Only Mode Does

Local-only mode routes eligible AI processing to models running on your device
or approved local infrastructure. It reduces or eliminates cloud model
processing for selected tasks.

Use it when:

- Content must not be sent to a cloud AI provider.
- You need offline or low-connectivity work.
- You are handling confidential source material or regulated data.
- You want predictable data residency on your device or local environment.

### Benefits and Tradeoffs

| Area        | Benefit                                                     | Tradeoff                                                                        |
| ----------- | ----------------------------------------------------------- | ------------------------------------------------------------------------------- |
| Privacy     | Prompts, attachments, and responses can stay local.         | Local app and device security become more important.                            |
| Offline use | Some conversations and local knowledge search can continue. | Remote sync, web search, cloud tools, and shared knowledge may not work.        |
| Latency     | Small local tasks can avoid network round trips.            | Large models may be slower without a capable GPU or enough RAM.                 |
| Cost        | Fewer cloud model calls.                                    | Local model downloads require storage and maintenance.                          |
| Quality     | Good for many drafting, coding, and summarization tasks.    | Complex reasoning, very long context, and newest models may need cloud routing. |

### Requirements

Requirements depend on the selected local model and app:

- Desktop or approved local runtime.
- Enough RAM and storage for the model.
- Optional GPU acceleration for larger models.
- Local model provider such as an approved Ollama-compatible runtime when
  configured.
- Local knowledge index availability if you want offline knowledge search.
- Workspace policy permitting local model use.

### Model Options

Your workspace may offer local model families such as Llama, Mistral, Phi,
CodeLlama, or other approved models. Choose based on:

| Need                 | Prefer                                                          |
| -------------------- | --------------------------------------------------------------- |
| Fast lightweight Q&A | Smaller local models with low memory requirements.              |
| Coding help          | Code-specialized local models or approved cloud coding models.  |
| Private drafting     | Local general-purpose models.                                   |
| Long context         | A model with a larger context window, often requiring more RAM. |
| Highest quality      | Workspace-approved premium cloud models or large local models.  |

### Offline Behavior

Offline availability is limited to resources already present on the device.

| Capability                    | Offline in local-only mode?                           |
| ----------------------------- | ----------------------------------------------------- |
| Local conversation generation | Yes, if the model is installed and licensed.          |
| Local attachments             | Yes, when the parser and model support the file.      |
| Local knowledge search        | Yes, if the index is stored locally.                  |
| Cloud model routing           | No.                                                   |
| Web search                    | No.                                                   |
| Remote team knowledge         | No, unless synchronized locally before going offline. |
| Account sync                  | No, until connectivity returns.                       |

---

## Data Deletion

### Delete Specific Data

Use targeted deletion when only one item is wrong, sensitive, or no longer
needed.

| Data type              | Typical deletion path                                  | Notes                                                       |
| ---------------------- | ------------------------------------------------------ | ----------------------------------------------------------- |
| **Conversation**       | Conversation menu > **Delete**                         | May also remove conversation search visibility.             |
| **Message or branch**  | Message actions or branch menu, if enabled             | Regeneration history may follow workspace retention policy. |
| **Memory**             | **Settings > Privacy & Security > Memory > View**      | Delete the exact memory and verify it no longer appears.    |
| **Knowledge document** | Knowledge base document menu > **Delete**              | Reindexing may be needed after deletion.                    |
| **Export file**        | Delete the downloaded file from your storage location. | Iris cannot remove copies you stored outside the system.    |

### Bulk Deletion

Bulk controls may include:

- Delete all conversations.
- Clear all memories.
- Delete a knowledge base.
- Remove all documents from a collection.
- Reset personalization.
- Revoke all API keys or connected app authorizations.

Bulk deletion is often irreversible. Export data first if you may need it later,
and confirm whether legal holds or audit retention apply.

### Account Deletion

Account deletion usually disables access first, then deletes or anonymizes user
data according to policy.

Expected effects:

- Active sessions are revoked.
- API keys and connected apps are disabled.
- Conversations, memory, preferences, and personal knowledge are deleted or
  queued for deletion.
- Shared knowledge or team artifacts may be reassigned, anonymized, or retained
  if the organization owns them.
- Audit logs and legally required records may remain for their retention period.

Use the self-service account deletion workflow when available, or contact your
administrator or privacy contact.

### Retention After Deletion

Deletion removes data from active use, but some records can remain temporarily
or in reduced form:

| Remaining record       | Why it may remain                                                                 |
| ---------------------- | --------------------------------------------------------------------------------- |
| **Backups**            | Backup systems age out on a separate retention schedule.                          |
| **Audit logs**         | Security and compliance records may be retained.                                  |
| **Legal holds**        | Legal, regulatory, or contractual requirements may pause deletion.                |
| **Anonymized metrics** | Aggregated usage metrics may no longer identify a person.                         |
| **External copies**    | Exports, screenshots, or messages shared outside Iris are not controlled by Iris. |

---

## Data Collection

### What Iris May Process

Iris processes data needed to provide conversations, agents, memory, knowledge,
search, tools, safety controls, and account administration.

| Data category         | Examples                                                       | Primary purpose                                        |
| --------------------- | -------------------------------------------------------------- | ------------------------------------------------------ |
| **Account data**      | Email, name, role, workspace, authentication identifiers.      | Authentication, authorization, notifications, support. |
| **Conversation data** | Prompts, responses, files, images, citations, model choices.   | Generate answers, maintain context, support export.    |
| **Memory data**       | Preferences, durable facts, task context, summaries, tags.     | Personalization and continuity when enabled.           |
| **Knowledge data**    | Uploaded documents, chunks, embeddings, metadata, permissions. | Retrieval, citations, grounded answers.                |
| **Agent/tool data**   | Tool requests, task state, outputs, permission checks.         | Execute approved workflows and show results.           |
| **Operational data**  | Request IDs, timestamps, status codes, metrics, traces, logs.  | Reliability, abuse prevention, security, debugging.    |
| **Integration data**  | API keys, webhooks, provider routing, connected app metadata.  | Authorized integrations and automation.                |

### What Iris Should Not Collect by Default

Unless a specific app feature or integration explicitly asks for it, Iris should
not collect:

- Passwords, private keys, API tokens, or credentials pasted as regular text.
- Unrelated browsing history outside Iris.
- Unrequested screen recordings.
- Location data not needed for the feature.
- Data from unrelated applications.
- Biometric data unless an enabled platform feature requires and discloses it.

If you accidentally submit secrets, delete the content and rotate the secret
through the owning system.

### Third-Party Processing

Cloud model providers, speech providers, search providers, observability
systems, or connected tools may process data when enabled. Iris should route
only the data needed for the specific task and should use provider credentials
and policies approved by the workspace.

To reduce third-party processing:

- Use local-only mode where available.
- Disable third-party model routing in privacy settings.
- Avoid web search and external tools for sensitive tasks.
- Use private knowledge bases instead of broad integrations.
- Ask Iris which model and tools it intends to use before sending sensitive
  content.

---

## Encryption and Security

### Encryption at Rest

Iris architecture includes encryption at rest for stored service data, encrypted
secrets, field-level encryption for sensitive fields, and secret management for
provider credentials.

Common protected data includes:

- Conversation records and attachments.
- Memory entries and embeddings.
- Knowledge documents, chunks, embeddings, and metadata.
- Account and preference data.
- API keys, tokens, and provider secrets.
- Backups and exported archives when stored inside managed infrastructure.

### Encryption in Transit

Iris uses encrypted transport for client and service communication. Production
deployments are expected to terminate public traffic over TLS and use secure
service-to-service paths according to the deployment's ingress, mesh, and
infrastructure policy.

For users, this means:

- Use the official Iris URL or app.
- Do not bypass browser or app certificate warnings.
- Treat downloaded exports as sensitive files after they leave Iris.

### End-to-End and Zero-Knowledge Modes

End-to-end encryption means content is encrypted so that only authorized clients
with the key can decrypt it. Zero-knowledge behavior means the server stores or
routes encrypted content without being able to read the plaintext.

Availability depends on the app, workspace policy, and selected feature. When
enabled:

- Keep recovery keys or passphrases in an approved secure location.
- Understand that losing the key may make data unrecoverable.
- Verify which features still work, because server-side search, citations,
  memory, and team sharing may require reduced functionality or local indexes.

### Logging and Redaction

Iris services use structured logging with sensitive-field redaction for values
such as passwords, secrets, tokens, and API keys. Redaction reduces risk, but it
does not make it safe to paste secrets into prompts.

For high-sensitivity work:

- Keep secrets out of prompts, memory, knowledge, and support tickets.
- Prefer local-only mode and private conversations.
- Disable optional analytics.
- Use approved secret managers for credentials.

### Access Control

Workspace security can include:

- SSO, OAuth/OIDC, session management, and API key authentication.
- Role-based access control.
- Resource-level permissions for conversations, memory, agents, and knowledge.
- Service accounts and least-privilege infrastructure roles.
- Tool permissions and sandboxing for agent actions.
- Audit logs for sensitive access and administrative changes.

---

## Compliance

### User Rights

Depending on your jurisdiction and workspace, you may have rights to:

| Right                         | What it means                                       | Typical Iris path                                    |
| ----------------------------- | --------------------------------------------------- | ---------------------------------------------------- |
| **Access**                    | View the data associated with your account.         | Privacy settings, export, administrator request.     |
| **Correction**                | Correct inaccurate profile, memory, or stored data. | Account settings, memory editor, support request.    |
| **Deletion**                  | Delete eligible personal data.                      | Item deletion, bulk deletion, account deletion.      |
| **Portability**               | Export eligible data in a usable format.            | Data export workflow.                                |
| **Restriction or objection**  | Limit optional processing.                          | Consent, analytics, memory, model-provider settings. |
| **Withdraw consent**          | Turn off processing based on prior consent.         | Consent management.                                  |
| **Appeal or contact privacy** | Escalate unresolved privacy requests.               | Administrator or designated privacy contact.         |

### GDPR and CCPA

Iris is designed to support GDPR and CCPA-style privacy workflows such as
access, deletion, portability, correction, consent management, and opt-out
controls. Whether a specific deployment is legally covered by GDPR, CCPA, or
another regime depends on the customer, region, contract, and workspace
configuration.

### Certifications and Agreements

Enterprise deployments may require security reviews, data processing agreements,
regional data residency, subprocessors lists, audit reports, or certifications.
Use your organization's legal, security, or procurement process for current
certification status and approved contractual terms.

Do not rely on this user guide as a legal policy. Treat it as a product usage
guide and refer formal requests to the designated privacy contact.

---

## Privacy Settings Reference

### Recommended Profiles

| Profile                   | Use for                                        | Suggested settings                                                                                                            |
| ------------------------- | ---------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- |
| **Maximum privacy**       | Regulated, confidential, legal, security work. | Local-only where available, no third-party models, private conversations, no optional analytics, memory off or explicit-only. |
| **Balanced work default** | Normal internal work.                          | Conversation history on, memory explicit or reviewed, approved cloud models, analytics per organization policy.               |
| **Knowledge-heavy team**  | Shared team docs and support workflows.        | Team knowledge enabled, strict knowledge permissions, citations required, memory limited to non-sensitive preferences.        |
| **Developer integration** | API and SDK users.                             | Scoped API keys, secret manager storage, webhook signature checks, redacted logs, least-privilege tool access.                |
| **Temporary research**    | One-off exploration or vendor review.          | Private conversation, no memory writes, export disabled unless needed, delete when complete.                                  |

### Settings Checklist

| Setting                  | Review question                                                     |
| ------------------------ | ------------------------------------------------------------------- |
| **Conversation history** | Should this conversation be searchable later?                       |
| **Auto-delete**          | How long should the conversation remain active?                     |
| **Long-term memory**     | Is Iris allowed to remember this after the conversation ends?       |
| **Agent memory access**  | Can specialist agents search or update memory?                      |
| **Knowledge visibility** | Who can see uploaded documents and generated citations?             |
| **Third-party models**   | May prompts and attachments be sent to approved cloud providers?    |
| **Local-only mode**      | Is local processing required for this task?                         |
| **Analytics**            | Are optional usage metrics acceptable for this workflow?            |
| **Exports**              | Where will downloaded data be stored, and who may receive it?       |
| **Connected tools**      | Which external systems can agents access?                           |
| **API keys**             | Are keys scoped, rotated, and stored in a secret manager?           |
| **Support sharing**      | Does a support request include only the minimum needed information? |

---

## Frequently Asked Questions

### General Privacy

**Q: Can administrators read my conversations?**

A: Access depends on workspace policy, permissions, encryption mode, and legal
requirements. Use private conversations, local-only mode, and end-to-end
encryption where available for higher privacy.

**Q: Are my conversations used to train models?**

A: Iris should not use your private workspace content for model training unless
your organization has explicitly enabled a governed training or evaluation
workflow. Third-party provider treatment depends on the approved provider
contract and routing policy.

**Q: What should I never paste into Iris?**

A: Do not paste passwords, API keys, private keys, tokens, recovery codes,
unapproved customer data, or regulated information unless your workspace policy
explicitly permits that use.

### Local Mode

**Q: Is local-only mode always private?**

A: It reduces cloud processing, but privacy still depends on your device
security, local model configuration, local logs, downloads, and any synchronized
knowledge or exports.

**Q: Why is local-only mode not available?**

A: The app may not support it, the model may not be installed, hardware may be
insufficient, or the workspace may have disabled local models.

**Q: Can local-only mode use team knowledge?**

A: Only if the knowledge content and index are available locally and policy
allows it. Otherwise, remote knowledge requires network access.

### Deletion

**Q: Is deleted data removed instantly?**

A: Deleted data should be removed from active use promptly, but backups, audit
logs, legal holds, and external copies may follow separate retention schedules.

**Q: Can I recover deleted data?**

A: Usually no. Export important data before deletion if policy permits.

**Q: Why does a deleted document still appear in an answer?**

A: The knowledge index may need time to update, or the answer may be using
conversation context rather than the document. Ask Iris to search only current
knowledge sources and verify the document was removed from the knowledge base.

### Technical

**Q: How can I verify which model handled a response?**

A: Check model metadata in the conversation details when exposed, or ask Iris to
show the model and tools used for the response. Administrators may have routing
logs for audit purposes.

**Q: Are logs safe?**

A: Logs are redacted for known sensitive fields, but they can still contain
operational metadata. Avoid putting secrets into prompts, file names, knowledge
metadata, or support tickets.

**Q: How do API users protect privacy?**

A: Use least-privilege scopes, store keys in a secret manager, rotate keys,
validate webhook signatures, redact logs, use idempotency keys carefully, and
avoid sending more user data than the task requires.

---

## Getting Help

For privacy requests, use your workspace's designated privacy contact,
administrator, or support channel. Include:

- Account email or user ID.
- Workspace name.
- Request type: access, correction, deletion, export, restriction, consent, or
  security concern.
- Affected conversation, memory, knowledge base, or integration IDs when known.
- Jurisdiction or contractual basis if relevant.

Do not include passwords, API keys, private keys, or unnecessary personal data
in the request.

---

## Related Documentation

- [User Guide](./user-guide.md) - Everyday Iris usage, memory basics, and
  quick-reference guidance.
- [Product Features](./features.md) - User-facing privacy, security, local
  model, and compliance feature inventory.
- [Architecture Overview](./architecture.md) - Security layers, encryption,
  redaction, service boundaries, and deployment architecture.
- [Integration Guide](./integration-guide.md) - API key handling, webhook
  security, memory consent, and logging guidance for integrations.
- [API Documentation](./api/README.md) - API scopes, errors, rate limits, and
  protocol contracts.
