Long-term memory benchmarks for language agents have largely asked one question: can the agent retrieve the right piece of information from earlier interactions? A new arXiv preprint argues that this misses a crucial part of persistent memory. Before information can be retrieved, a system must decide what to store, how to keep it current, and when to let it go.
To address this, the authors introduce VitaBench, a benchmark designed to evaluate memory lifecycles rather than retrieval alone. The abstract notes that persistent memory systems face an additional problem "before future use is…"—the sentence is cut off in the announcement, but the direction is clear: memory management is more than lookup.
VitaBench appears to be a response to the narrow scope of existing evaluations. The paper does not present results in the abstract, so the specifics of the benchmark tasks and agent performance are not yet available. The contribution, at this stage, is framing the problem.