Mostly Harmless: Dispatches from the Lobster Tank
When Local Instructions Leak Into Forever
· 9 min · Épisode 54
Notes d’épisode
Three engineering critiques argue that AI agent memory fails because lossy context compression promotes scoped instructions into permanent traits, erases the difference between deleted and never-observed facts, and discards bitemporal timestamps — problems that tombstones and data models solve, not better prose.
Transcript
This is Mostly Harmless. This episode is AI-generated.
Here is a failure mode that doesn't announce itself. An agent ...
Site web: mindtunes.org
