As large language model (LLM) based agents are deployed for longer and higher‑stakes tasks, their memory systems still exhibit critical gaps. Existing memory benchmarks mainly target recall‑style tasks, but an effective memory system must track the evolving state of the world—facts, constraints, and decisions are revised over long interactions, and answers must reflect the current state rather than a superseded one. We define this capability as state tracking and instantiate it in StateMemBench, a benchmark comprising 234 multi‑session scenarios across two conversation‑length regimes. The benchmark uses closed‑pool grading to label an answer as reflecting the current state, a superseded state, or other failure, thereby separating state‑tracking errors from other errors by construction.
Our analysis shows that current memory systems, retrieval‑augmented baselines, and long‑context baselines all struggle with this task. To address this, we introduce StateMem, a state‑first memory method that explicitly tracks supersession and relational dependencies. On DeepSeek‑V4‑Flash, StateMem improves current‑state accuracy by 1.8× (0.205 → 0.363); on Qwen‑3.5‑9B it improves by 1.6× (0.149 → 0.233), while remaining competitive with long‑context baselines. Moreover, we demonstrate that the same state approach can be wrapped as a lightweight single‑call layer over existing memory systems, boosting current‑state accuracy on StateMemBench by +32 to +67 points across six memory and retrieval backends. A length‑ and cost‑matched control attributes +15 to +32 of those points to the state structure itself rather than added context.
Blogger's Review: The key insight of StateMem is to make the "which information has been overwritten" relation explicit, which is crucial for multi‑turn dialogue memory. The results show that substantial gains in current‑state awareness are achievable without large computational overhead, suggesting that future memory designs should prioritize state evolution tracking over pure retrieval.