Full-duplex speech agents face a tricky problem: they need to update persistent memory while simultaneously listening and speaking. But it is not clear which events should be stored, revised, or excluded. A new arXiv preprint, DuplexLedger, proposes a benchmark for this task.
The authors argue that existing benchmarks isolate interaction timing from memory outcomes, meaning they test either how well an agent handles real-time conversation or how well it remembers, but not both together. DuplexLedger appears designed to couple these dimensions, though the abstract gives no experimental results yet.
As a preprint, the work is not peer-reviewed. Still, it highlights a real gap in evaluating conversational AI systems that must maintain state while engaging in fluid dialogue.