Memory your team can govern

Agents got sharper at using what they already know. Your team also got a way to see, correct, and erase what they believe, so memory stays an asset instead of a liability.

Govern what agents remember

Experience only helps if someone can own it. Tenants can now inspect the facts agents have come to believe, correct a wrong one, or erase it entirely before the next run compounds the mistake. The credential that runs agents is no longer the same one that can rewrite their memory, so security and platform leads get a real separation of duties instead of an all-or-nothing key. What agents know stays visible, editable, and under your control.

Put earned facts to work

Agents should not rediscover what they already earned. Facts proven on earlier runs now show up in the plan itself, and agents can answer from that store instead of burning cycles re-asking, re-checking, or refusing because nothing was looked up again. The org keeps the benefit of prior work: faster runs, less wasted spend, and fewer "we already knew that" loops.

Drop rules the world outgrew

A rule that was right yesterday can be wrong this morning. Until now, the longest track record was also the hardest to overturn, so the most established guidance kept steering work even after a vendor API changed or a pipeline was reorganised. Agents now weigh recent outcomes more heavily than ancient wins. A single noisy miss does not kill a good rule, but a sustained shift in what the world returns does. When that happens, the old support no longer props the rule up. Trust has to be re-earned on what happens next, so stale playbooks lose their grip instead of outvoting reality.

We also wrote about agents that start with experience and catching false success before it reaches you if you want the longer versions of those threads.