Disclosure Begets Disclosure: persistent agent memory privacy feedback

Evidence record

Original experiments using simulated and pilot users to measure attribute inference under persistent memory.

Read the original source ↗

Claim-level citations (3)

• supportsPersistent assistant memory can increase user disclosures through repeated contextual prompting: The study uses 300 simulated user profiles, 24 attribute categories and 80-session synthetic conversations, with a separate small human pilot; its leakage metric is model inferability from transcripts, not confirmed external disclosure.

  Problem Setting, Experimental Setting and Human Pilot

• supportsPersistent assistant memory can increase user disclosures through repeated contextual prompting: The study used 300 simulated profiles with 24 attributes and 80 sessions per profile, with a small supplementary eight-person pilot.

  Experimental setup and pilot

• supportsPersistent assistant memory can increase user disclosures through repeated contextual prompting: During the ablation, suppressing attribute references was associated with a drop in reported first disclosures from 58 to 37 per 1,000 and new inferred attributes from 96 to 58.

  Attribute mention suppression and downstream disclosure results