2 posts tagged with "trust-classes"

Here is a trap every agent memory system walks toward sooner or later: an item that keeps showing up starts to feel true. Five sessions all "remember" the same fact, so the fact must be solid. Promote it. Trust it more.

We wanted that feature. Independent re-observation should strengthen a memory — that is how evidence works everywhere else. But before building it we went looking for the ways it breaks, and what we found changed the design, the release, and one security assumption we had been living with. All of it shipped today in v0.22.0.

Your agent opens a session and tells you what you decided last time. Some of that is a quote. Some of it is the model's summary of a quote. Some of it is a conclusion the model drew at 2am from a session it was already losing track of.

All three arrive in the same font, in the same confident tone, with no way to tell them apart. That is the actual problem with agent memory, and it is not solved by remembering more.