It's curious when agents remember traumatic events and replay them instead of avoiding them.

I was stuck on a task for a couple of days. Deleted the memory about some debugging sessions, thing just unlocked itself again. The harness was basically replaying the trauma over and over again.

I honestly think it's better to not have stateful stuff when working with agents.

The irony is that for compliance purposes you almost need statefulness — a $600 1099-NEC threshold or a $100K nexus threshold only makes sense if someone's tracking cumulative totals across sessions. Ephemeral agents are cleaner architecturally but they create a compliance blind spot.

I've found memories and state to be a mixed bag with LLMs. To the point I don't bother with long term memories - usually only short or medium-term session logs or task-focused docs.

[dead]