In How to Keep Your CLAUDE.md Light, I pointed to memory as one of the places your trimmed-out rules should go. This guide is about that memory layer.
Before we start, a disclosure. I'm building Monet, a memory platform for AI agents. And this week we reversed its direction: we dropped generic memory, the store-anything-recall-anything layer. This post is the record of the failures that forced that call.
If you're explaining the same thing a third time
When a session ends, the agent forgets everything. The design direction you settled yesterday, the reason you rejected that library — gone by the next session. Catch yourself explaining the same context for the third time and you reach the conclusion everyone reaches: add memory.
We did. Then the next problem arrives.
