AfterDeleteBench Reveals Information Persistence in LLM Agents Post-Deletion
October 23, 2026
AfterDeleteBench demonstrates that deleting original memory records in LLM agents fails to prevent information leakage through derived states or summaries. In tests with Qwen2.5-7B-Instruct, a single query recovered deleted secrets in roughly 66% of cases, even when verbatim strings were no longer present.
HOW THIS AFFECTS YOU
●
builderYou cannot rely on simple record deletion to ensure privacy in agentic systems with long-term memory.
●
policyThis highlights a critical technical gap in meeting data deletion and 'right to be forgotten' regulatory requirements.