Agent memory that stores facts but not their versions: adding state tracking is worth +32 to +67 points across six backends

arXiv 2608.19652·high signalXBlueskyLinkedInCopy link

StateMemBench, 234 multi-session scenarios, tests whether a memory system returns the current value of a fact or a stale earlier one. Applied as a wrapper over existing memory systems, the StateMem method gained +32 to +67 points across six backends, with roughly 15 to 32 of those points attributable to structure rather than extra context volume, and 1.8x on DeepSeek-V4-Flash. If your agent memory is an append-only note store, the cheap win is versioning entries rather than adding retrieval depth.