Encrypted, isolated, and yours to take with you.
Your memory is some of the most sensitive data an agent touches. We protect it with layered controls — and we describe them honestly, tagging what's live versus what's still being built, because a security page that claims perfection is a red flag.
What protects your data today
Encrypted at rest
Every store — the structured database, object storage, the vector index, and the key/value cache — is AES-256 encrypted on disk. Raw uploads carry a second layer: application-layer envelope encryption, so a copy of the object store alone is useless without our key.
Encrypted in transit
TLS 1.3 with perfect forward secrecy on every connection. HTTP is redirected to HTTPS; there is no unencrypted path in or out.
Tenant isolation, tested
Every capsule is a private namespace. Isolation is triple-enforced — a namespace filter on every vector query, a scoped re-fetch, and a metadata guard that rejects any mismatch — and an automated isolation suite runs on every change, deriving its route list from the router itself so a new capsule-scoped route that skips the guard fails on the day it is written.
Keys you control
API keys are stored only as a hash, salted with a secret pepper held outside the database — so a copy of the database alone cannot verify or reproduce a key. Shown once, never again. Every key is scoped and revocable, and the control-plane API has no public route at all.
Your data is portable
Export the whole capsule — corpus, manifest, eval set, and write log — as an open, documented format and rebuild it anywhere. We never sell your data or train shared models on it.
Payments off-platform
Card details go straight to Stripe — the platform never sees or stores a card number. Billing webhooks are signature-verified and idempotent.
Privacy, and a hard line on abuse.
Strong privacy and a safe platform aren't in conflict — but honesty about the boundary is. Retenta is not end-to-end encrypted: retrieval and answering require reading your content, so we can technically access it to run the service. We hold that access to a minimum and never sell it or train shared models on it.
Because we can see content, we take responsibility for it. We run automated trust-and-safety scanning for a narrow set of illegal material — child sexual abuse material, and credible threats of mass harm. When something is confirmed, we preserve the evidence, remove the content, and cooperate with valid legal process and report to the appropriate authorities as the law requires. This is a tiny, targeted carve-out — not general monitoring of your memory.
What we do and don't allow is spelled out in the Acceptable Use Policy; how we handle data is in the Privacy Policy.
What we're hardening next
Beyond the platform's AES-256 at rest, we're rolling out application-layer envelope encryption for store-only sensitive artifacts (raw uploads, archived transcripts, audit and evidence blobs) under a pluggable key-custody model — so a database read alone can't surface that plaintext. Queryable content stays readable by the retrieval engine by design; that's the honest limit of any system that answers questions over your data. We'll mark this page's claims live only as each control ships.
Own the memory. Control the access.
Export your capsule any time, revoke a key in one click, and read exactly how the format works.