Skip to main content
This page is shared with customers under direct enablement. Zero Data Retention is an enterprise feature - please reach out to your Tavus account team to turn it on for your account.

Overview

Zero Data Retention (ZDR) ensures that customer-supplied and conversation-derived content is not retained by Tavus once a conversation completes. When ZDR is on, Tavus still records system metadata (IDs, timestamps, status, duration) so the conversation is billable and observable. Transcript, prompt, and any PII-bearing fields are deleted and redacted as soon as post-call processing finishes, and PII is scrubbed from logs and metrics. Content is held only for as long as it takes to run the conversation and complete post-call processing, such as generating the conversation summary and perception analysis and delivering your webhooks. Tavus does not use this content for any purpose other than delivering your conversation. ZDR lives at the billing account level - once enabled, every conversation created under that account is subject to it. There is no customer-facing API to flip ZDR on or off for the account; ask your Tavus account team and they’ll set it for you.
Memory Stores are not currently available for ZDR-enabled accounts. This includes both learned and pinned memory; creating a Memory Store returns 403.

Webhooks Still Fire With Full Content

ZDR changes what Tavus stores, not what Tavus delivers to you. Webhooks continue to fire normally, and they are not redacted. Magic Canvas is different. Each canvas.interaction event is sent once, with no retry. If that send fails, Tavus cannot resend it. In particular, the application.transcription_ready and application.perception_analysis callbacks still deliver the full transcript and perception analysis to your webhook endpoint, for you to use however you wish. The only difference under ZDR is that Tavus deletes and redacts its own copy of that data as soon as post-call processing, including webhook delivery, completes.
If you rely on ZDR to keep transcript or perception data out of your own systems, remember that this content is still delivered to your webhook endpoint. Handle it according to your own data-retention policies.

Magic Canvas Interactions

When an end user interacts with a Magic Canvas card in a ZDR conversation, this is what Magic Canvas delivers to you and what it keeps. Delivered to your callback_url:
  • A canvas.interaction webhook for every card interaction, with the full value and metadata. Tavus makes one delivery attempt and doesn’t retry.
  • The end-of-call application.transcription_ready transcript, which includes a submit’s text form (the words passed to the PAL for that answer) when there is one.
Without a callback_url, Tavus delivers neither, and keeps no copy of the answers. Only the system metadata listed below remains. Conversations started by the Tavus-hosted embed or widget have no callback_url, so to receive answers under ZDR, create the conversation through the API (POST /v2/conversations with a callback_url) and render it with @tavus/cvi-ui. What Magic Canvas keeps:
  • No interaction record. Magic Canvas stores no value, metadata, or interaction_id, and the interaction history endpoint returns an empty list.
  • No answer values in Magic Canvas logs.
  • Tavus keeps operational fields only: the conversation ID, tool_call_id, component, component version, interaction type, and timestamps. It does not keep the card text or what the user entered. For an error, the logs may include a short fixed reason code in snake_case (lowercase words joined by underscores), such as card_failed. That code comes from a set list. It is not the user’s answer and not text they typed. The logs may also keep the names of up to five fields from that interaction, such as email, and never the contents of those fields.
As with transcripts, these answers reach your own systems, so handle them according to your own data-retention policies. This section covers Magic Canvas only; the rest of this page describes ZDR for the conversation as a whole. For how to handle a repeat, and for created_at, see Zero Data Retention in the Canvas interactions docs.

Enabling ZDR for Your Account

Reach out to your Tavus account team. They will enable ZDR at the billing-account level on your behalf. Once it’s on, every subsequent conversation created on the account is automatically created with data_retention="none" and routed through the ZDR redaction path. ZDR applies to conversations created after the flag flips. Conversations that started before the flag was enabled are not retroactively redacted.

Debugging a Single Conversation

When ZDR is on for the account, troubleshooting can be hard - the rows Tavus support would normally inspect have their content fields replaced with [redacted]. To work around this for a single conversation without changing your account-level policy, pass data_retention_debug: true when creating the conversation:
cURL
That one conversation is created with data_retention="full" and its content is retained end-to-end, so Tavus support can inspect it alongside you. Because the flag opts that conversation out of ZDR, the ZDR deletion and redaction commitment does not apply to it. Every other conversation on the account continues to run under ZDR - data_retention_debug is scoped to the single create call.
data_retention_debug is only honored when your billing account is already set to data_retention="none". On an account that doesn’t have ZDR enabled, the field is a no-op - it does not enable ZDR for that conversation, and it does not enable ZDR for the account. The flag is strictly a one-way debug escape hatch out of an otherwise ZDR-enforced account.
If you want ZDR enabled for your whole account, that decision lives with your Tavus account team - data_retention_debug is intentionally not the path to turn ZDR on or off broadly.