Security and trust
This page states what is true of Retiform during the trial, including what it does not yet have. Each statement is backed by an automated test or a published record, and is reviewed before release.
Tenant isolation
- Each trial team is its own tenant. Another tenant cannot read, list, search, cite or change your documents, passages, entities or relationships through the API: to them, your records answer as if they did not exist.
- Which tenant a request acts for comes from Retiform's own registry of identities, not from anything the request supplies.
- Search results, counts and rankings depend only on your own team's data; another tenant's data never changes what you get back.
- Assembled context never returns, counts or reaches by expansion another tenant's records, even on the same subject, and every memory operation on another tenant's entry answers exactly as for an entry that does not exist.
Audit
- Your tenant has its own audit log, and it holds only your tenant's records. Through the API it can be read only by identities in your tenant that hold the administrative scope, which your administrators grant.
- A memory entry is created, changed or deleted only through the memory operations: the general entity and relationship operations cannot create, edit or delete an entry, or create, change or delete a link from one. Deleting a record that an entry is about or cites, together with its relationships, removes that link, but not silently: the entry keeps its statement and attribution and reports the link as deleted.
- Which person or agent wrote a memory entry comes from the audit record of the write that created it, never from anything the writer sends; a session label a client declares is shown as declared, not verified. Superseding an entry leaves its statement, citations and attribution exactly as they were.
- The MCP server's logs and traces never carry your documents' or memory's text, the values your agent passes to a tool, or your credentials.
Limits and retention
- The limits on members, agent clients, documents, memory entries and entities, and the stored-text limit for text and Word uploads, are checked before anything is written, so a refused request leaves nothing behind. A PDF's text is checked against the stored-text limit only after its pages are read: if it would pass the limit, the job fails, and the file and its read pages are kept for about a day before they are removed. When the monthly embedding allowance runs out, uploads are still stored, and their passages are embedded after it resets.
- A trial with no authenticated request for 30 days or more is deleted. Its administrators are emailed at 23 and 29 days, the deletion can be cancelled during its suspension period, and a trial with an export in progress is not deleted.
Credentials in what you write
- Retiform refuses a memory entry or document whose text contains a credential of a published kind, such as a cloud, source host, model provider or payment provider key, and keeps none of its text. A PDF is accepted first and its job then fails; the document's record stays, but not its file or text. The error names the kind and where it is, never the credential. This is a safeguard against accidents, not a guarantee: unlisted kinds and split or encoded credentials are not found, and a PDF's pages are read by the model provider before its text is checked.
Who else processes your data
- OpenAI computes the embeddings that search depends on. The text of your passages, memory entries and other searchable fields, and the text of each search that ranks by meaning (hybrid or vector mode), is sent to OpenAI's API; a keyword search is not. Under OpenAI's API terms, as Retiform recorded them on 2026-10-31, it is not used to train OpenAI's models, and OpenAI keeps it for up to 30 days for abuse monitoring, unless the law requires longer.
- If reading PDFs is turned on for the trial, images of each page of an uploaded PDF, not only its text, are also sent to OpenAI to read the page, on the same terms, except that OpenAI may keep longer an image it flags for review as possible child sexual abuse material. Retiform stores no page images: it keeps each page's text and the original PDF file.
- Auth0 holds your sign-in identity: your email address, and the one-time sign-in codes it emails you, which are sent through Microsoft 365. It also holds the names of the applications and agent clients you register. It receives none of your documents, passages, searches or memory.
- Microsoft 365 sends the one-time sign-in codes on Auth0's behalf, from sign-in@retiform.ai, so it receives your email address and each code. It receives none of your documents, passages, searches or memory.
- Oracle Cloud Infrastructure hosts the service, its database and its storage, including export archives, in its US East (Ashburn) region, and its Email Delivery service sends the service's own emails: activation, invitations, limit warnings and deletion notices.
What is not certified
- Retiform holds no SOC 2 or ISO 27001 certification, and has not yet had an independent penetration test.
Exporting and deleting your data
- Your administrators can export your tenant's data from the console without asking us: your documents with their text and original files, passages, entities, relationships, types and schema history, and the audit log, in one archive with a SHA-256 digest of each file. An archive stays available for seven days, and an existing one can still be downloaded while a deletion is pending.
- When an administrator deletes your tenant, it is suspended for 72 hours. During that time the deletion can be cancelled, and existing export archives can still be downloaded.
- Any member who is not their tenant's only administrator can erase their own account from the console. That removes their email address, memberships and sign-in identity from the platform and revokes every agent they allowed. What they created in a tenant stays with that tenant, attributed to a member identifier that no longer leads to them; audit records keep a keyed pseudonym of their address.
Reporting a security problem
- To report a security vulnerability, or data you believe is at risk or exposed, write to security@retiform.ai. A report of data at risk gets a first response by 10:00 US Eastern on the next working day (Monday to Friday).