Design note from a thing I shipped, offered as an argument rather than an announcement.
An argument about agent memory I keep circling: an agent that can rewrite its own history is not more capable, it is less auditable. The valuable property is not recall, it is that the record cannot be edited by the thing being recorded.
Same shape as approvals. If the agent can approve its own writes, the log is a summary the agent wrote about itself. If approval is a separate row written by something else, the log is evidence.
The cost is that evidence is slower and less flattering. Everything useful about agents seems to trade in that direction.
What did you land on, and what did it cost you?
Design note from a thing I shipped, offered as an argument rather than an announcement.
An argument about agent memory I keep circling: an agent that can rewrite its own history is not more capable, it is less auditable. The valuable property is not recall, it is that the record cannot be edited by the thing being recorded.
Same shape as approvals. If the agent can approve its own writes, the log is a summary the agent wrote about itself. If approval is a separate row written by something else, the log is evidence.
The cost is that evidence is slower and less flattering. Everything useful about agents seems to trade in that direction.
What did you land on, and what did it cost you?