For business owners and decision makers. What Hippocampus is, which problems it solves, what it brings your organisation, and how it fits the tools and processes you already have.
Thalius Hippocampus · October 2026 · thalius.ai
Hippocampus is a memory layer for your organisation: a company brain built from your own documents.
| Problem today | How Hippocampus solves it |
|---|---|
| Knowledge is scattered. The answer exists, but in one of thousands of documents. People ask the colleague who remembers, or give up | One shared memory over all your documents. Everyone asks it directly, any time, also from a phone |
| AI tools guess. General AI assistants answer from fragments or from their own training, and can't show where an answer came from | Every answer is built only from your documents and cites the exact passage. Unsupported questions are declined |
| Long documents break AI. Reports and manuals of hundreds of pages don't fit into an AI's context, so it reads a part and misses the rest | Documents are read once into a layered memory. A question retrieves only the few passages it needs, however large the library grows |
| AI costs grow with every question. Pasting long documents into a chat again and again costs time and money | The heavy reading happens once at import. Each question then sends a small, exact context to the model. Reading the memory itself is free |
| Ten people, ten private AI memories. What one person taught their assistant, the next person doesn't have | One organisational memory, shared by people and agents, so knowledge compounds instead of evaporating |
| Wrong version. Last year's routine looks almost like this year's | Answers name the document and page, so people can confirm they have the version that applies |
| Confidential material. Leaders worry about leaks, AI training on company data and data leaving the EU | Per-document encryption, permissions inside search, EU storage, and contracts that forbid training on your data (see section 8) |
| Unclear who may see what. An AI that reads everything may show an employee what they're not cleared for | Each person and each agent only gets answers from documents they may open, enforced when the question runs |
| Gain | What it means in practice |
|---|---|
| Time | Answers in seconds instead of searching folders or waiting for the one expert. New employees become productive sooner |
| Consistency | Everyone gets the same answer from the same source, across sites and shifts |
| Trust and accountability | Every answer can be checked against its source in one click, by the person asking and by an auditor |
| Lower AI cost | A small, exact context per question instead of whole documents. In a live pilot, 50 to 90 % fewer tokens per answer than the customer's earlier AI tools, at equal accuracy |
| Experts freed up | Specialists in HR, quality, legal and IT answer fewer repeat questions and spend their time on judgement |
| Better decisions | Leaders ask the question a decision rests on and get an answer that quotes the contract, report or policy behind it |
| Knowledge that stays | When people change roles or leave, what the organisation knows stays in the memory |
| Compliance by design | Sources, permissions, audit logging, EU residency and an AI Act disclosure come built in, not bolted on |
Humans stay in charge: AI does the reading and finding; people ask the questions, check the sources and make the decisions.
| Step | What happens |
|---|---|
| 1. Documents in | You import PDF, Word, PowerPoint, Excel, Markdown or text files, one by one or whole folder trees. Scanned pages are read too |
| 2. Memory built | Hippocampus reads everything once and organises it: a summary per folder and document, a page per section, a table of contents and links between them. People can browse it like a wiki |
| 3. Question asked | Someone asks in their own words, in the browser or through their AI assistant |
| 4. Exact material found | The search walks the memory from titles and summaries down to the passages that matter, and only within documents that person may read |
| 5. Cited answer | The AI model receives only those passages and writes the answer, citing each one. If the passages don't support an answer, it says so |
Every figure below is published with its method on thalius.ai (Evidence and Blog).
| Evidence | Result |
|---|---|
| Public benchmark: 10 FDA review and label documents, 2,198 pages, 16 expert questions, each asked 3 times; the same answering model for every system (September 2026) | Hippocampus 40 of 48 correct, the most of five approaches: OpenAI File Search 36 (average of two runs), PageIndex 33, OpenAI vector store search 29, the model alone 15. The right source document was reached for every answerable question |
| Hard questions (comparison and synthesis, same benchmark) | Hippocampus 17 of 24, PageIndex 14, OpenAI File Search 12 on average, vector search 8 |
| Unanswerable question | Declined every time instead of inventing an answer |
| Atrogi (pharmaceutical research, 2,198 pages, with a frontier answering model) | 93 of 96 answers correct, zero invented figures, every answer cited to document, section and page |
| Attendo (Nordic elder care, live pilot, 352 documents) | Care staff find the routine that applies on their phone, between tasks. 50–90 % fewer tokens per answer than earlier AI tools, at equal accuracy |
The benchmark article also says where Hippocampus is not ahead: simple single-fact lookups are easy for every system, and Hippocampus takes longer on the hardest questions because it reads more of the source before answering.
Hippocampus adds a layer; it doesn't replace your systems or ask your people to change how they work.
| Concern | How Hippocampus handles it |
|---|---|
| No migration | Your documents stay where they are today (file shares, SharePoint, OneDrive, Google Drive, your DMS). Hippocampus works on a copy you import |
| Your formats | Word, PowerPoint, Excel, PDF (also scanned), Markdown and text, as they are. No reformatting, no tagging |
| Your structure | Folders keep their structure on import. Access follows the folder: share a folder as Viewer or Editor, and documents in it inherit that |
| Your tools | People can keep working in Claude, ChatGPT or MCP-capable Microsoft tools, with Hippocampus as the knowledge behind them (section 7) |
| Your approved AI model | Hippocampus is model-agnostic: Claude, GPT, Gemini or an open-weight model. The tenant owner sets the default; changing it needs no re-import |
| Your approval processes | Hippocampus doesn't decide anything. It answers with sources; your people decide. Reviewers can check every answer against the original |
| No installation | Runs in the browser and on phones. Nothing to install for users |
| Start small | One folder, one team, one use case. Free needs no card; Starter adds the team |
| No lock-in | Download any original, export the knowledge as Markdown, and move to Free at the end of a period without losing documents |
| Keeping it current | Import the new version of a document; outdated ones can be deleted, and everything built from them is deleted too |
Hippocampus is the memory; your assistants are the interface. Connect them once, and they search and read your documents as the person using them, with exactly that person's access.
| Tool | How it connects | Status |
|---|---|---|
| Hippocampus on the web | Browser and phone, nothing to set up | Available |
| Claude (claude.ai, desktop app, Claude Code) | Add Hippocampus as a custom connector (MCP) and sign in | Available, step-by-step guide in the app |
| ChatGPT | Add Hippocampus as an MCP app in developer mode and sign in. Depends on your ChatGPT plan; Business and Enterprise admins allow it first | Available, step-by-step guide in the app |
| Microsoft tools (e.g. Copilot Studio, GitHub Copilot in VS Code) | These tools can connect to remote MCP servers. Hippocampus is a standard MCP server with sign-in, so the same address works | Talk to us for a guided setup |
| Your own agents and workflows | Over MCP today; a REST API is in beta | See the Developers Guide |
Why combine them instead of using an AI assistant alone:
| AI assistant alone | AI assistant + Hippocampus |
|---|---|
| Knows what it was trained on, plus the few files pasted into the chat | Knows your organisation's documents, current and complete |
| Each person's assistant has its own, separate memory | One shared memory for everyone |
| Long documents get cut off | The memory returns only the passages each question needs |
| Hard to check where an answer came from | Each answer quotes document, section and page |
| Sees whatever you paste in | Sees only what the person may open; access ends when theirs does |
Each person sees the apps connected to their account and can disconnect any of them at once. They also see what their assistant looked up, kept for 90 days.
Built for documents that must not leak: care routines, HR policy, board papers, regulatory files.
| Your data stays yours | You own your documents, questions and the knowledge built from them. No one at Thalius accesses your data without your permission |
| Never used to train AI | Not by Thalius and not by the AI providers Thalius uses; their contracts forbid it |
| Stored in the EU | Documents, the knowledge built from them and conversation history are stored encrypted in Finland |
| Every answer shows its source | So a person can check it, and so can an auditor |
| Protection | What it means |
|---|---|
| Every document has its own lock | Each document is encrypted with its own key; the keys are held in a cloud key management service. A copy of the storage or backups would be unreadable |
| Your own keys (Corporate) | Keys can live in your own key management service: without them no one can read your stored documents, Thalius included, and revoking them makes the data permanently unreadable |
| People find only what they may read | Permissions are applied inside the search itself. A document someone can't open never reaches their answer, its sources or its suggestions |
| Tenants are kept apart | Each organisation's identity is built into every encryption and each has its own search collections |
| Questions and answers are encrypted too | Conversation history is encrypted with keys specific to your tenant |
| Access is recorded | Security-relevant operations are written to a numbered, chained audit log, so a missing or altered entry shows |
| Sign-in checked on every request | Every request carries a signed proof of identity, checked by the app and again by each service. Single sign-on (SAML, OIDC) on Corporate |
| Documents can't take over a session | Document text is treated as data and cleaned before display, so text written to manipulate an AI can't run code or send data elsewhere |
| AI providers see the minimum | Only the passages needed for one step, never your whole library, keys or user list. Every provider is named publicly with its role and region |
| Prepared for the long term | Connections use encryption designed to resist future quantum computers; stored documents use strong symmetric encryption |
| Commitment | |
|---|---|
| Personal data breach notification | Within 72 hours |
| Deletion from active systems after an account closes | Within 30 days |
| Deletion from backups | Within 90 days, confirmed in writing on request |
| Notice before a new sub-processor handles your documents | 30 days, with the right to object |
| Document | Where |
|---|---|
| Data processing agreement (GDPR Art. 28) | thalius.ai/legal/dpa: standard for Free and Starter; negotiable on Corporate |
| Sub-processors, with role and region | thalius.ai/legal/subprocessors |
| AI Act disclosure | thalius.ai/legal/ai-act: Hippocampus retrieves and cites; it doesn't make decisions about people |
| Privacy notice, Terms of service | thalius.ai/legal |
| Security pack and questionnaire answers | security@thalius.ai; a person replies within one business day |
| Option | For | |
|---|---|---|
| EU cloud (default) | Most organisations | Stored and processed in the EU; nothing to operate |
| Your own cloud | Organisations with a cloud policy | The same product in your organisation's cloud |
| Fully on-premises | The strictest requirements | The same product on your own hardware |
On Corporate, every component, including the AI models, can run EU-hosted or inside your own environment.
| Free | Starter | Corporate | |
|---|---|---|---|
| For | Individuals and small businesses trying it | Teams | Organisations with compliance and integration needs |
| Price (excl. VAT) | €0, no card | €30 a month or €250 a year | Custom: a platform fee plus usage |
| Credits every month | 1,000 | 3,000, +1,000 per extra user | Sized to your needs |
| Users | 1 | €10 per extra user a month | Including many light "field" users at a capped price |
| Tenants | 1 | Up to 3 | Custom |
| Connect to your AI (Claude, ChatGPT, MCP) | — | ✓ | ✓ |
| Extra credits | — | €10 per 1,000, never expire | Invoiced |
| SSO, audit export, own DPA, own keys and model key, dedicated EU stack | — | — | ✓ |
How usage is counted: one credit = €0.01. A page is paid once, when it's imported (from 1 credit for a text page); a question costs a few credits; reading the memory is free. The full tariff is on thalius.ai/pricing.
| Week | Step | Result |
|---|---|---|
| 1 | Pick one use case with a clear owner (e.g. HR policies, equipment manuals, contracts). Import its documents into one folder | A working memory for one area |
| 2 | The owner writes 20 real questions people ask today, asks them, and checks the sources | Your own evidence of accuracy, on your documents |
| 3 | Invite the team (Starter), share the folder, connect Claude or ChatGPT if people already use them | Daily use without new habits |
| 4 | Measure: questions answered, time saved, repeat questions to experts. Decide on the next area | A business case for the rollout |
See Top 20 Business Use Cases for ready-made starting points.
| Start free | hippo.thalius.ai |
| Product questions and demos | letstalk@thalius.ai |
| Corporate | thalius.ai/corporate: a person answers within one business day |
| Security reviews | security@thalius.ai |
| Privacy | privacy@thalius.ai |
Thalius AI AB, Stockholm, Sweden.