What this covers
Papez is a memory for AI. You, or an assistant acting for you, save facts, preferences, decisions, and context. Papez keeps them in one causal graph, ranks them so the useful ones surface, lets stale ones fade, and can explain why any memory is there.
This policy covers the hosted service at papez.ai and everything that connects to it: the web app, the Claude connector, the ChatGPT app, Claude Code and other MCP clients, the Notion integration, and the API.
It does not cover the open-source engine when you run it yourself. The library runs on your machine, stores state locally, and sends nothing to us. Its licence and behaviour are described in the developer documentation.
What we store
Memory content
The core category and the most sensitive one. For each memory we store the full text you or your assistant provided, a short summary, a category label, the assistant that saved it, the title of the conversation it came from, and the metadata the engine needs to rank it: a retention score, causal weight, access counts and timestamps, a stability value, and whether you pinned it. When enrichment is enabled, named entities extracted from the text are stored too.
Embeddings and causal links
To find relevant memories we convert their text into numerical vectors and store them beside the memory. Memories are connected by typed edges, such as caused_by or supersedes, each recording the two memories, the type and strength, how the link was made, and, when enrichment is enabled, a one-line reason. Because a reason can mention a name, our erasure process scrubs personal identifiers from it.
Identity and authentication
Sign-in is handled by Clerk. We store your Clerk user ID, which ties memories to you, and never your password. Your email, name, and avatar live at Clerk and are fetched live when the web app needs them. When you connect an assistant, we store a SHA-256 hash of its access token, the OAuth client record, the scopes granted, and an expiry. Plaintext tokens exist only in server memory and refresh tokens are never written to disk.
Connected apps and imports
If you connect Notion, we store the Notion access token, encrypted, and a map of the pages you granted so syncs can be incremental. Content from those pages becomes memories in the same way as anything else you save. If you import a ChatGPT or Claude export, a Notion export, or an Obsidian vault, your own messages and notes become memories with links between them. Imports happen only when you upload a file, and the archive itself is discarded after processing.
Organisations, billing, waitlist
If you use Papez in a team, we store the organisation record, memberships and roles, synchronised from Clerk, and memories carry a visibility flag so shared ones are distinguishable from private ones. Paid plans are billed by Stripe through Clerk Billing. We store your plan and subscription status. We never see or store card numbers. If you join the waitlist we store your email, encrypted.
Operational data
Request logs carry IP address, user agent, path, and status for a short period so we can operate and protect the service. Rate-limit counters live in server memory only. Administrative actions are written to an audit log. Error reports go to Sentry with personal data reporting switched off, so memory content is never included.
We do not collect advertising identifiers, biometrics, precise location, or data from any source other than what you and your assistants send us.
How memories are created
A memory is saved when you show intent to remember: you ask an assistant to remember, update, or forget something, you share information you want kept for later, or you confirm a durable fact or decision. Assistants are instructed not to collect personal facts on their own initiative and not to store details you have not indicated you want kept.
Papez only sees what passes through a conversation or a file you deliberately import. It does not read your email, browsing, files, or other accounts. Everything saved is visible to you in the web app and through the assistant, and anything saved by mistake can be deleted immediately.
Why we use it
- To provide the service: saving, ranking, recalling, and explaining your memories so a connected assistant has the right context.
- To build the graph: linking related memories so you and your assistants can follow cause and effect.
- To authenticate you and isolate your data: every query is scoped to your user, or to organisations you belong to for shared memories.
- To operate and protect the service: rate limiting, abuse prevention, reliability, and billing.
We do not sell your data, use it for advertising, or train models on your memory content. We do not show your memories to anyone who is not you, an assistant you connected, or a member of an organisation you shared them with.
Who processes it
Your data reaches only the providers needed to run Papez. The AI providers below receive memory text for a specific job and are not permitted to use it for anything else.
| Provider | Role | What they receive |
|---|---|---|
| Neon | Managed Postgres, AWS us-east-1 | All data at rest. Memory content is encrypted before it reaches the database. |
| Railway | Application hosting | Data in transit through the running service. |
| Cloudflare | TLS, CDN, and edge protection | Request metadata in transit. |
| Clerk | Identity and billing | Your sign-in, profile, organisation membership, and subscription state. |
| Stripe | Payments, via Clerk Billing | Card and billing details. We never see them. |
| OpenAI | Embeddings | Memory text, to produce the vector used for search. |
| Anthropic | Enrichment | Memory text and pairs of memories, to extract entities, summarise, infer causal links, and detect contradictions. |
| Notion | Integration, only if you connect it | Requests for the pages you granted, authorised by your token. |
| Sentry | Error monitoring | Stack traces and request metadata. Personal data reporting is off. |
We will update this table before adding a provider that handles your content.
Where it lives and how it is protected
Memories, embeddings, links, and connection records live in Postgres on Neon in AWS us-east-1. The service runs on Railway. Traffic is served over HTTPS behind Cloudflare.
Memory content is encrypted at rest with a key generated for you alone, which is itself wrapped by a master key that never touches the database. Someone holding only database access cannot read your memories. The running service decrypts in memory to search and score, which is what lets it answer you. The full design is on the security page.
How long we keep it
Your memories persist until you delete them. There is no blanket expiry after a number of days.
The engine continuously rescores memories. A low score makes a memory less prominent and eventually dormant, but dormant memories are still stored and still retrievable. Scoring is not deletion. A memory is deleted automatically only if, all at the same time, its score has fallen below the forgetting threshold, it has no causal connections to other memories, it is not pinned, it is not a core memory, and it is not shared with an organisation. A memory that is connected, pinned, core, or shared is never removed automatically.
- Assistant access tokens expire within 24 hours. Expired records are ignored and cleaned up.
- Uploaded import archives are discarded once processed.
- Request logs are kept briefly for operations and security, then rotated.
- When you erase your account, private data is deleted at once and backups age out within 30 days.
Your controls
Through any connected assistant you can search, recall, and list what is stored, ask why a memory is retained, correct a memory with a superseding one, pin or unpin, and delete any memory permanently. Deleting a memory also removes its links, immediately and irreversibly.
In the web app you can browse the whole graph, timeline, and core memories, and export everything. An authenticated request to GET /api/export returns your complete graph, every memory in every status and every link, as one JSON document you can take to any other tool. Your memory is not locked to Papez or to any one AI vendor.
Disconnecting Notion deletes the stored token and page map. Disconnecting an assistant revokes its token.
Full erasure
To erase your account, email [email protected] from the address on the account. We run a complete erasure: all private memories and links are deleted, your organisation memberships and preferences are removed, your Notion connection is deleted, and your encryption key is destroyed so any remaining ciphertext is unreadable. Memories you promoted to an organisation are kept for the team but anonymised, with your identifier replaced and personal identifiers scrubbed from link explanations. Tell us if you want those deleted too and we will do that instead. Self-service erasure in the app is on the way.
Your rights
Wherever you are, you can ask us what we hold about you, correct it, receive a copy, restrict or object to processing, or have it erased. Under the GDPR the export above is your right to portability and the erasure above is your right to be forgotten. Under the CCPA you have the same rights of access and deletion, and we do not sell or share personal information. Email [email protected] and we will respond within 30 days. You may also complain to your local data protection authority.
Papez is not directed at children under 13, or under the age of digital consent where you live, and we do not knowingly collect their data. If a child has used Papez, contact us and we will delete the data.
We operate from the United States and store data there. If you use Papez from elsewhere, your data is processed in the United States by us and by the providers listed above.
Changes and contact
When this policy changes materially we update the date at the top and, where it matters, tell you directly. Continued use after a change means you accept it. Earlier versions are available on request.
Astrix Labs Inc. · astrixlabs.ai · [email protected]
Questions about this document are welcome. We answer them ourselves.