WitMem
Evidence-witnessed governance for reliable long-term agent memory.
Overview
WitMem studies how erroneous verifier judgments can acquire persistent influence when they control admission to and removal from an agent’s long-term memory.
The framework separates transient judgments from persistent memory authority. It uses task-grounded evidence to govern lifecycle transitions, including candidate promotion, retrieval, and retirement, while limiting repeated influence from related records.
The project focuses on reliable memory governance under noisy witnesses, contaminated evidence, and long-horizon agent operation.