Plain language, accurate to the running code. If the code and this page ever disagree, tell us and we will fix whichever one is wrong.
The short version. Your knowledge is captured on your machine, secrets are redacted on your device before anything syncs, your memory store is yours to read and delete note by note, nothing is shared with another person unless you explicitly say yes, and we do not sell your data, and it is never used to train anyone else's model.
The local installer and client read AI session history that already exists on your computer (Claude Code, Codex, and Cowork session files) and distill it into short structured notes — decisions, findings, constraints. A redaction pass runs on your device, before anything leaves it, removing API keys, tokens, and credential-shaped strings. Depending on your setup, distillation runs on a local model, through an AI provider using an API key you control, or — if you choose the option the installer offers with a consent prompt — transiently through our servers at our cost. On that relay path the redacted text passes through us to the AI provider and we store nothing from it except the notes you approve; the raw material is not retained.
The connector (claude.ai, ChatGPT, and similar) stores only what your AI explicitly saves during a conversation — notes it writes with the remember and wrap-up tools, and shares you initiate. It does not transcribe your conversations.
The browser extension captures conversations only from AI chat tabs (claude.ai, chatgpt.com, gemini.google.com), only after you switch it on through its consent screen, with the same credential redaction running in the page before anything is sent. It has a visible on/off state and does nothing until you provide your own token.
The web app uses your private token to show you your own store. Server logs record ordinary operational events (requests, errors) — not conversation content.
Each person's store is a separate database on our servers (hosted on Fly.io in the United States), reachable only with that person's private token. You can read every note the system holds about your work in the app, and delete any note or any whole source at any time. Deleting is deleting — not hiding. We run those servers, so your notes are encrypted on disk but not end-to-end encrypted.
Sharing is consent-gated on both sides: you choose what slice to offer, the recipient must explicitly accept, and nothing crosses before that. Shared knowledge is read-only, scoped to the topic you named, labeled with your name, and revocable by you at any time. Share invitations by email include a one-click way to never receive them again.
We send email only for things you or a collaborator did: invitations, share offers, and occasional product updates to beta participants. Reply to any of them to reach a human.
Delete individual notes or whole sources directly in the app. For complete deletion of your store and our copies, email us — we confirm when it is done. You can export your store's contents; ask and we will help.
When our practices change, this page changes with them, and beta participants hear about material changes by email.
Oliver Zahn — oliver@coretx.ai. Sapience is built by CoreTx, Inc.
Last updated August 27, 2026 (added the opt-in server-relay extraction path).