PersistBench: Evaluating Long-Term Memory Safety in AI Models

PersistBench explores a critical safety question in AI regarding when long-term memories should be forgotten. The benchmark evaluates whether models can use memory helpfully without falling into specific failure modes, such as cross-domain leakage, while maintaining utility. It was created by PersistBench and last updated on February 3, 2026.

Use Cases

• Benchmarking model safety against cross-domain leakage based on user context influence described in the summary.
• Evaluating the trade-off between memory utility and safety risks in AI systems.
• Testing for specific failure modes in models with persistent memory.

Strengths

• Focuses on a specific, critical AI safety question regarding memory persistence.
• Designed to evaluate multiple failure modes, including cross-domain leakage.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

PersistBench

Freshness

Last updated 2026-02-03 06:42:45; freshness should be verified.