Data note
This note lists who processes what you type into the concierge.
ACMS runs on Google Cloud. Your question is also sent to Google's Gemini API to compute the search vector used to find relevant content. Conversation and account records are stored in Supabase, the search index in Turbopuffer, and run traces that can include conversation text in LangSmith. Background job events pass through Inngest, operational telemetry goes to Grafana Cloud, and this site is served by Vercel. Each of these providers keeps data under its own terms.
Remember on this browser
This option is not offered at present. What follows applies to a browser where it was turned on earlier.
Off unless you turn it on. When it is on, the preferences, constraints and decisions you state here are kept for 30 days from the day you turn it on, then deleted. Using the assistant does not extend the 30 days.
To recognise this browser, a secret key is kept in this site's storage in this browser. Memory works only in this browser, not on your other devices.
Anyone else who uses this browser can get answers that use what was remembered here.
Any script running on this site can read and copy that secret key.
Whoever holds a copy of the key can recall this memory until you press Forget or the 30 days end.
Forget makes this browser stop using its memory, and ACMS removes what it stored for it within 24 hours. Copies kept by the service providers named in the data note are not removed by Forget: they follow the retention terms disclosed there. If this browser's site data is cleared first, Forget can no longer be used here, and the memory ends 30 days after you turned it on.
Anthropic
Anthropic is the primary provider that generates answers. It receives your question, the published site content used to answer it and the conversation context needed to answer. Under Anthropic's standard API terms it may keep requests for up to 30 days, subject to the exceptions it documents. ACMS has no zero-retention agreement with Anthropic.
OpenAI
OpenAI is the fallback provider, used when the primary provider cannot answer. It then receives the same request. Under OpenAI's default API terms it may keep requests for up to 30 days for abuse monitoring, subject to the legal and harm-prevention exceptions it documents. ACMS has no zero-retention agreement with OpenAI.
TypeSafe
TypeSafe checks whether the concierge answered a question, so that the project can count the questions left unanswered. After the concierge completes an answer, a background step sends TypeSafe the question, the final answer and the published content the answer cited. It does not delay the answer and sends nothing else from the conversation. TypeSafe processes this data on ACMS's behalf, as its processor, and does not use it to train models without ACMS's consent. Under TypeSafe's terms it may keep these copies for as long as it considers necessary, with no fixed deletion date. ACMS has no zero-retention agreement with TypeSafe. No deletion made by ACMS reaches the copies TypeSafe keeps: not an erasure request, and not Forget on a browser where Remember was turned on.