Can Saucepan AI Read Your Chats? Privacy and Retention

Saucepan stores your chats, according to its privacy policy, but the official material reviewed does not settle whether character creators can read private transcripts or how routine staff access works.[1][3][12] Its training program is opt-in and disabled by default.[1] Neither statement answers every meaning of “can Saucepan AI read chats.”

This review checks official text as delivered on October 26, 2026. Direct legal-page extraction returned an age gate, so the policy was read from the site’s publicly linked JavaScript asset without interacting with the gate.[8] The policy is dated July 3, 2026.[1] These are vendor statements, not findings from testing account access, network requests, or deletion.

Saucepan stores chats even when training is switched off

The short version: Storage, model training, and human access are separate questions.

The policy says: “Chat messages and conversation history are stored until deleted by the user.” It includes message content and timestamps, associated context, and images shared during conversations.[1] Turning off training does not mean those records stop existing.

For training, Saucepan states: “This program is disabled by default and requires your explicit consent before any data is used for training.”[1] Its two participation levels have different scopes:

• Rated messages only: The rating, the specific rated message, and the preceding message for context.[1]
• Rated message + history: The rated message and full conversation context; only conversations with ratings are marked for training.[1]

The policy also says nothing is used for training if an opted-in user never rates a message.[1] That is a consent promise about training, not proof that nobody can access stored content. It does not establish end-to-end encryption. The security section explicitly declines to guarantee absolute security.[1]

Character creators and platform staff have different roles

The short version: Do not claim either side has transcript access without evidence.

A character’s creator supplies its definition. Saucepan’s FAQ explains that hidden definitions and anti-scraping protections protect that creative work.[3] A hidden definition is not the same thing as a hidden conversation: it concerns what users or external providers can learn about the character’s instructions, not what the creator can read about you.

No explicit creator-transcript permission rule was located in the inspected policy, FAQ, terms, and wiki navigation. That gap does not establish a creator inbox, and it does not justify a promise that creators can never see transcripts.

Platform staff are a separate question. The recovered policy describes stored chat content but does not set out detailed routine staff-access controls.[1] This review therefore cannot tell you which staff roles may open content or whether such access is logged. Missing documentation is not evidence that employees habitually read messages.

Ask about creator permissions and staff procedures separately. The broader guide to human access in companion apps provides background; Saucepan’s own documentation must answer the Saucepan-specific questions.

External model providers receive prompts independently of creator privacy

The short version: A private character definition and a private user conversation have different data boundaries.

Saucepan’s wiki distinguishes its in-house models from vetted and custom external providers.[12] It defines an external provider as an LLM endpoint not run by Saucepan, connected through an OpenAI-compatible public API.[12]

The documentation is explicit that “the prompt sent to an LLM contains the companion’s full definition” and that a custom provider sees that definition.[12] This establishes a recipient outside Saucepan for the prompt submitted for inference. It does not establish the exact message-history fields included in every request or each provider’s retention rules.

A creator’s permission to use an endpoint should not be mistaken for a privacy promise from its operator. Nor does supplying your own API key prove that requests bypass Saucepan’s servers. This review did not trace requests. Nothing inspected establishes that a self-hosted endpoint removes the history Saucepan says it stores.[1]

For the wider distinction between stored records and material sent for inference, see the anatomy of AI companion app data.

Vetted providers are not a universal zero-retention guarantee

The short version: Saucepan says vetting primarily protects creators against scraping.

The wiki describes its checks this way: “We do basic checks to ensure their data handling and privacy policies meet a minimum standard, but our main focus is on scraping prevention.”[12] That statement is narrower than a guarantee that every vetted endpoint discards every prompt immediately.

Vetted providers are allowed by default, but creators can restrict a companion to in-house models.[12] Custom access depends on ownership or the creator’s permission; the wiki says custom endpoints are unreviewed and that Saucepan does not know what they do with received prompts.[12] It documents configuration using an endpoint URL and API key.[12]

Before using an external route, identify the actual endpoint and read the policy that applies to its API service. Check retention and training separately. Do not substitute an unrelated consumer-chat policy, or assume the word “vetted” answers those questions. This is a check to perform, not a claim that a particular provider misuses data.

Deleting an account includes a stated backup window

The short version: Deletion is not described as instantaneous erasure of every copy.

Saucepan’s policy says chat history persists until user deletion and promises permanent deletion of messages and chat data when an account is deleted.[1] It then qualifies the timing: deleted data may remain in regular system backups for up to 30 days.[1]

Other retention categories are listed separately. Usage analytics have a 90-day period, error logs have 30 days, and message ratings and training data are retained until message deletion or account deletion.[1] The error-log period should not be mistaken for a rule that all conversations expire after 30 days.

These are Saucepan’s stated periods. They do not establish deletion from an external provider, removal of copies you exported or shared, or reversal of training already incorporated into model parameters. No actual deletion was tested.

The policy also permits specified disclosures for valid legal demands, protection of legal rights, fraud or security issues, and emergencies involving imminent physical harm.[1] Those exceptions do not document routine staff reading, but they matter when evaluating a claim that content can never be disclosed.

Check the active route and consent before sharing sensitive details

The short version: Base the decision on named controls, not a broad “private” label.

Use this checklist before entering information you would not want another party to hold. An unanswered question is a reason to request clarification, not evidence of wrongdoing.

The policy lists privacy@grubs.ai for privacy questions.[1] Ask about the particular companion and endpoint you use rather than requesting a general assurance that everything is private.

If you are comparing Saucepan with Local Waifu, apply the same questions to the model route you select. Keep identifying details out until the answers meet your needs. Saucepan documents storage and optional training; creator transcript permissions and detailed staff access remain questions this review cannot settle.

Sources

[1] https://saucepan.ai/privacy-policy - privacy [3] https://saucepan.ai/faq - faq [8] https://saucepan.ai/assets/index-Cvqzn0DF.js - main-bundle [12] https://saucepan.ai/api/v1/wiki/page?slug=external-providers - providers-api-local

Questions people ask

Back to the blog