The paper argues an effective agent memory system must track the evolving state of the world: as facts, constraints, and decisions are revised over a long interaction, answers must reflect the current state, not a superseded one. It defines this as state tracking and releases StateMemBench: 234 multi-session scenarios across two conversation-length regimes, with closed-pool grading that scores whether an answer reflects the current state, a superseded state, or fails otherwise — separating state-tracking failures from other errors by construction. Existing memory systems, retrieval-augmented baselines, and long-context baselines all struggle. The proposed StateMem, a state-first memory that explicitly tracks supersession and relational dependencies, improves current-state accuracy 1.8x over the strongest same-backbone baseline (0.205 -> 0.363 on DeepSeek-V4-Flash) and 1.6x over the strongest memory system (0.149 -> 0.233 on Qwen-3.5-9B). As a lightweight single-call wrapper over six existing memory/retrieval backends, it lifts current-state accuracy by +32 to +67 points, and a length- and cost-matched control attributes +15 to +32 of those points to state structure rather than added context.
The best absolute accuracy is still only 0.363, so evolving-state tracking remains an unsolved failure mode across every current memory architecture; closed-pool grading makes it measurable in production evals. For teams building agent memory today, the wrapper result (+32 to +67 points across six backends with a single call) is immediately actionable: track supersession explicitly instead of hoping recall-shaped benchmarks cover it. This is the third paper this week (after ev-20260820-02 and ev-20260821-02) converging on the same meta-finding: aggregate or recall-shaped evaluation misses what agent workloads actually depend on.
| Benchmark | StateMemBench: 234 multi-session scenarios, two conversation-length regimes, closed-pool grading (current state / superseded state / fail) |
|---|---|
| Baselines | existing memory systems, retrieval-augmented baselines, long-context baselines — all struggle |
| Statemem | state-first memory tracking supersession + relational dependencies |
| Results | 1.8x over strongest same-backbone baseline: 0.205 -> 0.363 current-state accuracy (DeepSeek-V4-Flash) · 1.6x over strongest memory system: 0.149 -> 0.233 (Qwen-3.5-9B) · single-call wrapper over 6 memory/retrieval backends: +32 to +67 points · length/cost-matched control: +15 to +32 points attributable to state structure itself |
| Relation To Kb | same overwritten-value recall line as ev-20260820-02 (arXiv 2608.18578); third paper in the week's evaluation-blind-spot cluster |