Where agent memory actually breaks: 6 reads on read-time curation, stale updates, heavy tails and harness-side control
Most "agent memory" demos stop at the moment something gets saved. The interesting failures happen later: the right fact is in the window and the model still answers with the old one, or…