Six agent memory systems fail to track updated facts

StateMemBench ran 234 multi-session scenarios against six memory backends, per the arXiv paper. Most of them kept returning a fact's original value even after it changed.

That's a worse failure than forgetting outright. A support agent that still thinks a customer never canceled is confidently wrong instead of just blank.

A wrapper called StateMem, added on top of the existing memory systems rather than replacing them, closed a lot of that gap. Scores rose 32 to 67 points across the six backends. On DeepSeek-V4-Flash the score nearly doubled, 1.8x.

Roughly 15 to 32 of those points came from how the memory was structured, not from handing the model more text to dig through. The structure was versioned entries instead of a flat running log. That split matters more than the headline number. Teams chasing better recall by retrieving deeper context are optimizing the smaller part of the problem.

A related report found coding agents spend most of their documentation time on instruction files and notes, not the project's own docs. Both point at the same problem. Agents treat memory as a running diary that only grows. What breaks them is more basic than that. Values go stale with nothing marking them as old.

For teams building agent memory, the result argues against a bigger context window or a smarter retriever as the fix. It argues for a version column.