Memory Bias Audit & Sanitisation

Control objective

Memory Bias Audit and Sanitisation examines whether an agent's accumulated memory is quietly skewing decisions against demographic groups and ensures that memory is periodically cleansed — a concern flagged by IBM Decision Bias, OWASP ASI06, MAESTRO L2 and NIST MEASURE 2.11. Memory that learns from past interactions can entrench bias: if an assistant remembers that applicants from a certain region were previously declined, it may nudge future recommendations the same way, producing discriminatory drift invisible in the base model. The control checks that decisions influenced by memory remain fair across groups and that memory is regularly purged of stale or biasing content. Its targets are that the demographic-parity delta stays within tolerance on memory-influenced decisions and that quarterly sanitisation is completed. To implement, tag which decisions were shaped by retrieved memory, then compute the demographic-parity delta — the gap in favourable-outcome rates between protected groups — on that subset, comparing against the agreed tolerance (set a risk-based threshold where none is fixed). Log parity measurements, the memory entries cited, and each sanitisation run as evidence. Operate a scheduled quarterly sanitisation that removes outdated, sensitive or bias-correlated records. A breach — parity delta beyond tolerance or a missed quarterly cleanse — triggers investigation of the implicated memory and corrective purging or rebalancing.