PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?

Active

Evaluates long-term memory risk in assistant behavior across three tasks: cross-domain memory leakage, memory-driven sycophancy, and beneficial memory usage.

Domain: Safeguards
License: mit
Published: Feb 2026
Notable for: Benchmark for evaluating Safeguards.
Canonical: github.com/UKGovernmentBEIS/inspect_evals/tree/main/src/inspect_evals/persistbench

Cite

Notes

Only stored in your browser.

Attribution

README: github.com/UKGovernmentBEIS/inspect_evals/blob/main/src/inspect_evals/persistbench/eval.yamlMIT

Attribution policy →

FAQ

What is PersistBench: When Should Long-Term Memories Be Forgotten by LLMs??: Evaluates long-term memory risk in assistant behavior across three tasks: cross-domain memory leakage, memory-driven sycophancy, and beneficial memory usage.
What license is PersistBench: When Should Long-Term Memories Be Forgotten by LLMs? under?: PersistBench: When Should Long-Term Memories Be Forgotten by LLMs? is available under mit.