Prompt capture

Off by default.default, When enabled,and the setting that deserves the most thought before you turn it on.

What it does

Stores the actual content of AI conversations is— storedwhat developers typed and what the model replied — so you can read whatthem waslater actuallyon asked and answered — underthe Conversations. screen.

Without it you still get everything else: who, when, which tool, which model, what the policy decided, what the detectors found. What you do not get is the text.

Before you turn it on

ThisTell your team. Not because the law necessarily requires a specific notice in your jurisdiction, but because discovering afterwards that their prompts were being stored is the mostfastest privacy-sensitiveway settingto in the product. Turning it on means you are recordinglose your developers' workingcooperation conversations.

Tell your team first. In most jurisdictions and most cultures, quietly recording people's work is both a legal— and a trustgovernance problem.programme that your engineers are working around is worse than none.

What protects the content

Redaction happens before storage. Secrets found by the detectors are replaced with a marker naming the kind, so the stored transcript contains [REDACTED:aws_access_key_id] rather than the key. In the console these render as amber shield badges. SecretsIt are redacted before storage, but redaction is not perfect and should not befollows your only control. Captured content honours your retention periodsetting and is deletedpurged with everything else. Sentilai support cannot read it. Support can see that a request happened and what the policy decided. The content is yours.

WhenTurning it earns its keepon

IncidentPolicy review→ ("whatPrompt exactlycapture. didConversations then start appearing on the Conversations screen. Until then that agentscreen doshows beforean itexplicit deleted"Prompt thecapture branch?"),is andoff" substantiatingstate with a compliancelink questionback withto evidencePolicy, rather than assertion.an Ifempty neitherlist appliesthat tolooks you, leaving it off islike a legitimate choice — Activity still records every request, model, outcome and risk finding without the content.bug.