Ownership
Your data stays yours
No one at Thalius accesses your data without your permission. We never sell or share it, and only the sub-processors we list publicly can ever touch it.
Security and trust
Built for documents that must not leak.
Teams put their most sensitive material into Hippocampus: care routines, HR policy, board papers, regulatory files. This page explains how we protect it, in plain terms first and with the technical detail underneath.
Training
Never used to train AI
Not by us, and not by the AI providers we use. Their contracts with us forbid it.
Residency
Stored in the EU
Documents, the knowledge built from them and your conversation history are stored in Finland, encrypted.
Accountability
Every answer shows its source
Each answer points to the exact passage it came from, so a person can check it, and so can an auditor.
Encryption keys
Who holds the keys is your choice
Self-service · Free and Starter
Thalius manages the keys in Google Cloud's key management service. Each document has its own key, and the master keys never leave that service.
Corporate
The keys live in your own key management service. Without them, no one can read your stored documents, Thalius included. Every use shows up in your own audit log, and revoking the key makes your data permanently unreadable, backups included.
What happens to a document
From upload to deletion
From the moment you upload a file to the moment you delete it, these are the steps it goes through.
Upload
The file is encrypted with a key of its own before it is stored. Storage only ever holds the encrypted version.
Understand
Hippocampus reads the document, maps its structure and writes summaries. Readable text exists only in memory while this runs.
Ask
When someone asks a question, the search only looks through documents that person is allowed to read.
Answer
The AI model receives only the passages needed for that one answer. The answer cites each passage it used.
Delete
Deleting a document also deletes everything built from it: summaries, search index entries and links.
Protections
The protections, one by one
Each one is described in plain language. The line underneath gives the technical detail a security reviewer will ask for.
Every document has its own lock
Each document is encrypted with its own key. The keys themselves are locked by a cloud key management service, and the master keys never leave it. Someone who copied our storage or backups would get unreadable data.
Technical
XChaCha20-Poly1305 with a 256-bit key per document, wrapped via Google Cloud KMS. Stored object names are opaque keyed hashes, so file names are not visible in storage either.
People find only what they may read
Permissions are applied inside the search itself, before results are ranked. A document you cannot open never reaches your answer, its sources or its suggestions. If permissions cannot be checked, the search returns nothing.
Technical
Folder and document grants, groups and inheritance, enforced as a filter in the search engine. A summary built from several documents is visible only to people who can read all of them.
Workspaces are kept apart
Each organisation has its own workspace. The workspace's identity is built into every encryption, so data from one workspace cannot be decrypted as if it belonged to another, even by mistake.
Technical
Workspace, document and version are bound into each encryption as authenticated data. Each workspace has its own search collections.
Questions and answers are encrypted too
Your conversation history is encrypted before it is stored, with keys specific to your workspace. Our operational databases hold structure and timings, not readable content.
Technical
Conversation records sealed with AES-256-GCM under per-workspace keys derived with HKDF-SHA256, each bound to its own record so it cannot be moved or swapped.
Access is recorded
Every security-relevant operation on the document store is written to an audit log. Each entry is numbered and linked to the one before it, so a missing or altered entry shows up.
Technical
Sequence-numbered, hash-chained audit log with gap detection. Because each answer cites a precise passage, the log can show which content was used, not only which file was opened.
Sign-in is checked on every request
Every request carries a signed proof of who you are. The web app checks it, and so does each service behind it, independently. Larger organisations can sign in with their own identity provider.
Technical
Google Identity Platform with a separate tenant per workspace. Signed tokens verified at the application and again by each backend. Administrator rights require a verified email address. SAML and OIDC single sign-on on Corporate.
A document cannot take over your session
AI systems read whatever is in your files, including text written to manipulate them. Hippocampus treats document text as data. Anything a document or an answer contains is cleaned before it is shown, so it cannot run code in your browser or send your data elsewhere.
Technical
All rendered HTML passes an allowlist sanitiser. A Content Security Policy and frame protection are set on every page.
Prepared for the long term
Data stored today should stay safe as computing advances. Our connections already use encryption designed to resist future quantum computers, and stored documents use strong symmetric encryption that is considered quantum-resistant.
Technical
Public endpoints negotiate hybrid X25519 + ML-KEM-768 key exchange. Documents use 256-bit symmetric keys. The one public-key step in our key wrapping is scheduled to move to a post-quantum scheme.
AI providers
What AI providers see
Reading documents and writing answers uses large language models from specialist providers. We keep what they receive to the minimum each step needs.
What is sent
Only the text needed for one step: a page to read, a section to summarise, or the handful of passages that answer one question. Each request stands alone. Providers never receive your whole library, your keys or your user list.
What providers may do with it
Nothing beyond answering that request. They may not train on it or use it for anything else. Every provider is named on our sub-processor list with its role and region. By default some run in the US. On Corporate, every model can run EU-hosted or inside your own environment.
More choice of models
We are adding models from Mistral and Cohere alongside the ones we use today. You will be able to choose which provider reads your documents and writes your answers, including providers based in Europe.
Status
Planned. Text extraction already runs on Mistral in the EU.
Commitments
Our commitments in writing
These are part of our terms and data processing agreement.
72 h
Maximum time before we tell you about a personal data breach that affects your data.
30 days
Your data is deleted from active systems within 30 days of closing an account.
90 days
It is deleted from backups within 90 days. We confirm deletion in writing on request.
30 days
Notice before any new sub-processor handles your documents.
Running a security review?
We will send our data processing agreement, the sub-processor list and answers to your questionnaire. A person replies within one business day.