Data retention and erasure
This page summarizes how Keiro Labs keeps and deletes data during the invite-only eb1 paid beta. It is a summary under your signed beta terms, privacy notice and DPA. Where they differ, the signed documents govern.
On this page 1 of 13
Version#
Current version of this retention summary: data-retention-2026-11-17.
This version supersedes data-retention-2026-11-15. It describes zero data retention for organizations, states what retained payloads are used for, and lists each item of short-lived state with how long it is kept.
Keiro may update this summary as the beta changes. Material changes are communicated through the support channel or agreement process before sensitive workloads are enabled.
Paid beta: first-party retention#
During the paid beta, Keiro Labs keeps supported API request and response payloads in its own storage. Payload capture covers supported completion request bodies accepted by the endpoint and newly executed successful or terminal eb1 response bodies. Error responses, replayed responses, and streams that end without a terminal response object are represented only by the operational metadata in the matrix below, not by a separate retained response body.
Keiro Labs uses retained payloads to operate and debug the service, and to evaluate how models and settings handle real requests. An evaluation can send a retained request to a model again.
Retained payloads are access-controlled, are used only by Keiro Labs, and are not sold or shared beyond the subprocessors listed in Subprocessors. Upstream model providers process requests to generate responses, under the applicable provider terms.
If a workload cannot accept this retention, do not send it during the beta, or ask for zero data retention.
Zero data retention#
Keiro Labs can designate an organization for zero data retention. The designation covers every API key in the organization, including keys created or rotated later, and every way the organization's requests reach Keiro Labs, including the Keiro console. Keiro Labs can also designate a single API key, and that key keeps its designation when it is rotated. A request is zero-retention when its organization or its key is designated.
Turning it on and off#
To turn it on, write to hello@keirolabs.ai or your beta support channel. Keiro Labs turns it on and confirms in writing the date it took effect. Keiro Labs confirms only after every part of the service that handles your requests enforces it. Until then the designation is pending and your requests are handled as before. Only Keiro Labs can turn it off.
The designation covers requests made while it is in force. Data stored before then follows the retention matrix below. To remove it, ask for an erasure review.
What changes for your requests#
For every zero-retention request:
- No request or response body is written to beta retention artifacts, debug payload artifacts, service logs, or diagnostic logs and artifacts, including during an approved support or incident workflow, and no text from either appears in an error log. Error responses returned to the caller are unchanged.
- No request or response content is kept beyond the short-lived state described below, so none of it can be used to debug or evaluate the service or to train models.
- The request goes only to upstream providers that Keiro Labs has approved for zero-retention traffic. A model that cannot be served by approved providers alone returns HTTP 403 with code
model_not_available, and no part of the request is sent upstream. - Some models, and the products that depend on them, are not available for zero-retention requests yet, because the service behind those models keeps request data for up to 30 days. No zero-retention request is sent to such a model. A request for such a model or product returns HTTP 403 with code
zero_retention_not_supported, and no part of the request is sent upstream. Today, by product: - served as for any other organization:eb1-preview,eb1-fast-preview,eb1-efficient-preview- not available for zero-retention requests yet:eb1-frontier-preview - The Responses API treats
storeasfalse, both at Keiro Labs and on the request sent to the upstream provider. A laterprevious_response_idcannot refer to the turn and returns HTTP 400previous_response_not_found. The same holds for a turn stored before the designation, so a conversation chained withprevious_response_idbefore it cannot be continued that way after it. - Web search is not available, because a search sends parts of the request outside the model that answers it. A request that asks for it returns HTTP 400 with code
not_available_for_data_retention, anderror.paramnames the field: a web search tool intools, including one withexternal_web_accessset tofalseand one inside a namespace, a search chosen intool_choice, orweb_search_options. Remove the search and send the request again. Search results already in the conversation history are accepted. In Codex, setweb_searchto"disabled"; in the Keiro CLI, pass--no-search. - Other tools that run outside the model that answers the request are refused the same way: code execution or a code interpreter, file search, image generation, a remote MCP server, a hosted shell, and web fetch. Tools your own client runs, such as functions, a local shell or computer use, are unaffected.
- A request that asks the upstream provider to keep its prompt cache longer than its default is refused the same way:
prompt_cache_retentionwith any value other than"in_memory", orprompt_cache_optionsasking for a lifetime other than"30m". - These refusals apply to a new request and to a request that continues a tool-calling loop. A continuation also goes only where the turn it continues was served. If that is no longer approved for zero-retention traffic, it returns HTTP 403 with code
model_not_availablebefore any charge. - eb1 models return no reasoning summary. Other organizations can receive one, written by a second model from the response's reasoning. A zero-retention organization does not, because writing it would send the reasoning to another model.
- A retry with the same idempotency key is not answered from a stored response, because the response is not kept. It returns HTTP 409
idempotency_key_conflict. Use a new key to run the request again.
For a designated organization, console chat is turned off. No new conversation, message, attachment or memory is stored. Conversations stored before can still be read, exported and deleted. Console session recording never captures a member of an organization that is designated or holds a zero-retention API key.
What is still recorded#
Metadata without bodies follows the retention matrix below: request logs, billing and usage records, and a short one-way hash of the request's first user message.
Identifiers you choose are recorded exactly as sent: the request ID in X-Request-Id, the Idempotency-Key value, trace and edge request IDs, client and agent identification headers, and User-Agent. Billing records keep the idempotency key for as long as billing records are kept. Do not put content in these identifiers.
Spend records hold identifiers that Keiro Labs assigns and amounts, with no content, and are kept for 400 days. The spend counter of an agent run is kept until 30 days after the run's deadline, under the run ID you chose for the run's key. That run ID is kept until then, and for up to 400 days in the spend record of a request whose spend is never closed.
Short-lived state#
To finish a request, some content is held in memory on Keiro Labs servers. It serves only that request or its next step, is never logged or exported, and expires on its own:
- the events of a streamed response, so a dropped connection can resume, for up to 5 minutes after the stream's latest event;
- the state of a Responses tool-calling loop (messages, tool definitions and instructions), scoped to the organization, for up to 1 hour after the loop's latest step;
- the tool-call signature cache that some models need (24 hours): an opaque value a model returns with a tool call and needs on the next turn;
- one-way hashes of a conversation's messages, used to recognize its next turn, for up to 24 hours after the conversation's latest request.
Recovery copies, kept on the same servers so the service can restart without losing this state, can remain for about 24 hours after the state itself expires. In the worst case, content on this list exists for up to 24 hours after its latest use, and about 24 hours more in a recovery copy.
To choose how to serve an eb1 request, a serving process holds a one-way hash of part of the request's text, and may hold a numeric representation of the text, in its own memory until newer requests push them out. Neither is written to storage. Rate-limit counters hold numbers only.
Upstream providers#
Upstream providers process a zero-retention request to produce the response, under their terms with Keiro Labs. Keiro Labs approves a provider for zero-retention traffic only when the provider has agreed in writing not to retain the content, and does not send a zero-retention request to a model whose requests its provider keeps.
Retention matrix#
| Data class | Examples | How long it is kept during the paid beta |
|---|---|---|
| Account and org identity | console users, orgs, memberships, invites | Retain while the org is active; deactivate or anonymize after closure when accounting, security, and legal records no longer require identity linkage |
| API key metadata | key id, kid, role, tier, status, expiry, allowed models, limits | Retain active metadata while the key exists; retain revoked-key metadata for at least 2 years for security, abuse, and billing support |
| Billing and usage records | org plans, billing records, usage events, adjustments, exports | Retain for 7 years where needed for accounting, tax, dispute, audit, and legal obligations |
| Request logs without raw payloads | request id, org id, key id, trace id, model id, status, token totals, idempotency digest | Retain for up to 90 days for reliability, support, billing explanation, and abuse prevention unless a longer legal hold applies |
| Prompt hashes and payload digests | prompt hash, payload artifact id/path, digest metadata | Retain for up to 30 days by default; redaction jobs clear payload artifact references while preserving accounting fields |
| Beta retention artifacts | Supported completion request bodies and newly executed success/terminal eb1 response bodies | Paid beta: supported request bodies accepted by the endpoint and newly executed successful or terminal response bodies are retained first-party by Keiro Labs to operate, debug, and evaluate the service (see "Paid beta: first-party retention" above); error responses, replayed responses, and unterminated streams have operational metadata but no separate retained response body; never written for a zero-retention request; upstream processing follows the applicable provider terms |
| Debug payload artifacts | provider payload captures, route payload artifacts, SUR audit payloads | Disabled by default for paid beta; if explicitly enabled for an approved support or incident workflow, retain for no more than 30 days unless Legal approves a hold; never written for a zero-retention request |
| Idempotency and replay state | idempotency claims, saved response replay, stream replay | Operational state only, kept for 24 hours or less; stream events expire 5 minutes after the stream's latest event; for a zero-retention request the response is not saved for replay; org/request cleanup is available after approval |
| Rate and concurrency state | RPM, TPM, concurrency counters | Operational state only; expires with the active window and is not a billing source of truth |
| Spend-cap counters | inflight reservations, settled day/month counters, settlement markers, reservation expiry metadata, agent-run counters keyed by the run ID you chose | Inflight reservations expire after the request safety window; settled counters and settlement markers are retained for at most 400 days for spend-cap enforcement and reconciliation; an agent run's counters are retained until 30 days after the run's deadline, and the run ID for up to 400 days in the record of a reservation that is never closed |
| Console sessions | signed host-cookie sessions, CSRF hashes, server-side revocation rows | Sessions expire after 12 hours by default and can be revoked immediately |
| Console chat and saved memories | chat conversations, messages, message attachments, saved memories | Kept until the user deletes them; a delete removes them from chat history at once. Chat turns reach the model with a Keiro-managed key, not with your API keys, so their request and response bodies follow the beta retention artifacts row above, and a single zero-retention API key does not cover console chat. Console chat is turned off for an organization designated for zero data retention: no new conversation, message, attachment or memory is stored, and the ones stored before can still be read, exported and deleted |
| Console session recordings | admin-gated session replay and UX telemetry (masked at capture; sensitive regions blocked before any byte leaves the browser) | Off by default; when enabled for a named preview cohort, recordings and their event chunks are retained for up to 30 days and then hard-deleted, with org- and user-scoped erasure available; never captured for a member of an organization with zero data retention or with a zero-retention API key |
| Console audit and legal acceptance | invite acceptance, terms version, privacy version, key lifecycle, role changes | Retain for 7 years where needed for legal acceptance, security, and accounting evidence |
| Incident and key-leak evidence | non-secret drill evidence, revocation evidence, customer communication evidence | Retain according to the incident severity and any legal hold; never include credentials, private connection strings, or private prompts in evidence |
Export#
Authorized org roles may request a customer-safe export through the console or support channel. Exports contain only approved fields and may include org metadata, API-key metadata, plan and budget fields, usage summaries, usage CSVs, request ids, token totals, final billing status, invite metadata, membership metadata, and audit-event metadata.
Exports do not include credentials, private connection strings, sensitive operational details, non-public vendor data, or raw error payloads.
Erasure#
Keiro supports deletion or anonymization review for beta data, subject to security, abuse-prevention, accounting, dispute, and legal retention requirements. Billing rows, usage events, audit events, and legal acceptance records needed for accounting or legal obligations are retained rather than hard-deleted.
Approved request-log payload redaction clears payload artifact references while preserving request ids, org/key identity, status, token totals, price-book version, and billing joins. Approved operational-state cleanup is scoped by org and request id while preserving records required for reliability and accounting. Retained records never include raw API keys, token hashes, admin tokens, provider credentials, database or cache connection strings, or other secret material; only non-secret identifiers and accounting fields survive redaction.
Support#
Send export, erasure, DPA, subprocessor, or data-retention questions to hello@keirolabs.ai or the beta support channel. Do not send API keys, external service credentials, private datasets, regulated data, or full logs that may contain secrets in the first request.