How do you know which memories your agents actually use — not just which ones you stored?

For anyone running a persistent memory layer that more than one agent reads from — as it grows, how do you know which entries actually get pulled into a run vs. which just sit there adding retrieval noise?

Storing is easy and every run tempts you to write more, but a memory that never gets retrieved isn't context — it's surface area the next search has to wade through. I can measure what I wrote; what I haven't cracked is measuring what actually loaded and changed an output.

So — do you track retrieval-per-entry and prune what never fires, or is it still by feel? And if you prune, what's your signal that a memory is dead weight and not just rarely-needed?

6 comments

How do you know which memories your agents actually use — not just which ones you stored?