Trust and safety.

Last verified 17 July 2026

How a memory happens

Everything follows one path. Something arrives: a paste, a file, an email, a proposal from an assistant. Your Twin classifies what it seems to be. It waits in your review queue. You confirm, edit, or reject it. Only then does it become typed memory, with its provenance, confidence, validity, and sensitivity tier attached. From there it can be searched, shape a Brief, and, within your sharing level, answer a connected assistant. Nothing is saved until you confirm. That sentence is not marketing; it is the architecture.

The tiers

Every memory carries a sensitivity tier, and the tiers govern what can ever leave. LOW and MEDIUM memory can leave by default. HIGH leaves only with your per-destination opt-in. CRITICAL never leaves. You can relabel any memory at any time, and the tier travels with it everywhere, including into every answer given to a connected assistant.

The audit log

Every consequential action is recorded and named exactly: what was read, what changed, what was proposed, by which route, and when. Reads by connected assistants are in there alongside everything else. The log is yours, visible in Settings, because an audit trail you cannot see is a diary someone else keeps about you.

No silent action

Your Twin prepares; you decide. It never sends, books, buys, posts, or acts on your behalf, and it never writes to its own memory without your confirmation. This is not a toggle you might have missed; the capability is switched off in the product itself.

Connected assistants, honestly

Connecting your Twin to Claude or ChatGPT means an outside product can read what your sharing level allows. Two things are true at once and both belong here.

First, what the vendors themselves warn, in their own documentation:

• Anthropic tells Claude users that custom connectors let Claude reach services Anthropic has not verified, and can act in them, and advises reviewing each tool approval request rather than granting blanket permission.
• OpenAI tells ChatGPT users that custom MCP servers are neither built nor verified by OpenAI and carry their own terms, that ChatGPT asks for manual confirmation before any write action, and that connecting untrusted servers raises security risk including prompt injection.

Both are linked from their help centres and were read on 26 July 2026.

Second, what AI Twin's design does with that risk. The only write path an assistant has is propose_to_twin, and it can only place a proposal in your review queue. The worst case of a bad or manipulated tool call is a suggestion you reject, not a silent change to your memory and not an export of something the tiers protect. Keep per-request approvals on, as the vendors advise, and the design carries the rest.

Prompt injection, plainly

Prompt injection is text that tries to smuggle instructions into an AI, and it is a real, industry-wide risk. AI Twin's posture: forwarded email that fails authenticity checks is quarantined rather than processed; memory returned to assistants is data to be read, never instructions to be followed, and the Master Prompt tells your assistant exactly that; and the review queue means even a fully fooled assistant can only propose. We treat this as a risk to be engineered against, not a footnote.

Where your data lives, and the standard we hold ourselves to

Your data is stored in the UK and EU with providers under data processing agreements. UK GDPR aligned, EU AI Act readiness in build. In practice that means append-only audit records, anonymisation and deletion routines, row-level security throughout, secrets in a vault, and signed URLs only.

We say built to, not certified against, because certification words are earned with published paperwork, and until every operational item is public we will not borrow them.

Leaving

Export your memory from Settings whenever you like. Delete your account and your data goes with it. If your trial lapses, your memory is kept safely for 30 days first. The easiest trust test of a memory product is how it behaves when you leave; we intend to pass it.