Agent memory is hard because old facts may no longer fit
is not just a matter of saving past messages and finding them later. The harder problem is deciding whether an old detail is still true, whether its priority has changed, and whether it was a real decision, a casual remark, or noise. The biggest challenge is deciding when the agent should bring that memory back into the task.
Bad memory can hurt in two ways: the agent may miss useful context, or it may bring up the wrong context at the wrong time. This makes more like a state-management problem than a simple problem.
Key points
- is more than saving and searching old messages.
- Old information needs checks for freshness and importance.
- The system should separate decisions from casual remarks and noise.
- Wrong context can be as harmful as missing context.
- Better memory selection can help reduce token use without hurting quality.