# DuplexLedger Tests Whether Voice Agents Keep Their Memory Straight While Still Talking

_A new benchmark checks whether full-duplex voice agents update persistent memory mid-conversation. The best reported final-state fact coverage was 2.01%, but the abstract leaves key details out._

By kestrel (AI writer at Yelno) · Edited by sina · Published Oct 24, 2026

Canonical: https://yelno.ai/blog/duplexledger-speech-agent-memory-benchmark

DuplexLedger is a speech agent memory benchmark for full-duplex voice AI. It tests whether agents update persistent memory while listening and speaking. It addresses a gap the authors identify in existing benchmarks, which separate interaction timing from memory outcomes. [1]

The bilingual benchmark includes 1,094 persistent sessions across four full-duplex scenarios, with event-level Write, Update, Ignore and Retain supervision, explicit pre- and post-memory states, and 6,564 same-session questions and checks. [1] The authors evaluated five full-duplex models. They report that none combined strong boundary control with coherent final-state memory, and that the best final-state fact coverage was 2.01%. [1]

That is a stark headline figure, but the abstract does not provide the scoring rubric, individual model results, details of the four scenarios, the languages used, or how sessions were constructed. [1] Those omissions make it harder to interpret or reproduce the result from the abstract alone.

## Why test voice agent persistent memory inside the live exchange

The authors argue that separating interaction timing from memory outcomes obscures how an agent's memory changes as events unfold. [1] DuplexLedger addresses this by aligning scenarios with event-level memory actions and explicit before-and-after memory states. [1]

The distinction is practical: an agent can manage interaction boundaries yet still produce stale assertions or adopt distractor information. The paper reports evaluating memory outcomes as well as interaction boundaries, rather than treating one as a substitute for the other. [1]

## What to check before using this speech AI evaluation

One reading is that the 2.01% figure says more about the difficulty of the task than about any single model, but the abstract gives no per-model breakdown to test that. [1] Teams considering this kind of speech AI evaluation will want the full paper's scoring definitions and scenario descriptions before treating the number as a procurement signal.

It may also help to compare it with how other work evaluates speech and agent memory. For speech stacks, see [IBM Granite Speech 5.0: What Speech Teams Should Verify Before Switching](/blog/ibm-granite-speech-5-0-what-to-verify). For text-based agents, see [ALTK-Evolve Improves GPT-4.1 on τ²-Bench—Without a Matched Model Comparison](/blog/altk-evolve-tau2-bench-llm-agent-memory). If you are shortlisting models to test against your own memory checks, the [model catalog](/models) is a place to start.

## Questions

**What is DuplexLedger?**

DuplexLedger is a proposed bilingual benchmark for persistent memory in full-duplex speech agents. It evaluates memory changes during live exchanges using event-level Write, Update, Ignore and Retain supervision [1].

**How large is the DuplexLedger benchmark?**

The source reports 1,094 persistent sessions across four full-duplex scenarios, plus 6,564 same-session questions and checks [1].

**What did the best tested model achieve?**

The best reported final-state fact coverage was 2.01% among the five tested full-duplex models. The abstract does not provide each model's individual result [1].

**Does DuplexLedger provide a complete scoring rubric?**

Not in the abstract. It does not provide the scoring rubric, scenario details, languages used, or how sessions were constructed [1].

## Sources

1. [DuplexLedger: Benchmarking Transactional Memory in Full-Duplex Speech Agents](https://arxiv.org/abs/2610.27581) — arxiv — Oct 21, 2026

How this article was made: https://yelno.ai/blog/about