Where the scores moved
A new voice-memory system scored higher than comparison systems on tests of factual, persona and audio memory, according to an arXiv preprint. On information-memory benchmarks, VoiceMem led seven of 11 sub-categories and averaged 76.39. It was 24.12 points ahead of Mem0 and 15.90 points ahead of full-context processing.
The persona-memory test produced a score of 74.16 when VoiceMem used GPT-4o-mini and 76.56 with its fine-tuned response model. Both were reported against MemOS, with VoiceMem 1.89 points ahead. Because the two VoiceMem figures use different response-model conditions, they should not be treated as one directly interchangeable score.
On long, multi-turn audio, VoiceMem led in 11 of 14 categories. In three acoustic categories, it scored between 45.16 and 53.84, compared with 3.23 to 26.92 for text systems. The test extended the evaluation beyond written facts to information carried by sound.
A memory system built around streaming
VoiceMem separates two kinds of memory into parallel structures: an informational store and an affective store for emotional and persona memory. Its retrieval is divided into four stages: listening, speech tail, anticipation and searching. The training pipeline constructs memory worlds, uses small-language-model-verified online distillation, and applies human curation to create memory-dependent conversations; a human-curated subset is used for benchmark evaluation.
For the low-latency comparison, the retrieval-budget setting called K was 5, the study's default. At that setting, VoiceMem reached 91.2 with 430 memory tokens, and the reported retrieval time was 134 milliseconds. EverMemOS scored 83.13 with 1,899 tokens, leaving VoiceMem 8.1 points ahead while using 4.4 times fewer tokens. The 134-millisecond figure measures retrieval, not a fully specified end-to-end conversational latency; hardware and concurrency conditions were not reported.
On LoCoMo, VoiceMem was highest at every tested retrieval budget. Beyond K=5, increasing K to 10 added 1.3 points, while K=100 added 2.3 points for eight times the tokens. The reported curve flattened after K=5.
What the comparisons show
All five one-at-a-time ablations reduced accuracy on all four datasets. Removing the upper-layer index produced losses of 9.9 points on LoCoMo, 5.3 on ES-MemEval, 6.7 on ChatMem-Bench and 4.4 on Memora. When the index was transferred across backends on LoCoMo, the reported gains were 29.52 points for Mem0, 15.76 for LangMem and 22.92 for Zep, with a mean gain of 22.73.
The wider comparison used ten systems in four baseline groups. Baselines ran with GPT-4o-mini and text-embedding-3-small at temperature 0, while VoiceMem was tested at K values of 1, 3, 5, 10, 30 and 100. The reported left memory store held 510 items across six preset slots; two emergent slots covered 254 items, or 49.8% of the store. In ES-MemEval P1, the emergence strategy scored 74.40, versus 72.60 for static, 72.40 for a random split and 71.61 for a size threshold, in 32 sessions with 252 questions.
The evidence still has boundaries
ChatMem-Bench contains 316 questions across 53 hours of dialogue, divided into four ability groups and 14 sub-categories. The training pipeline uses a human-curated subset for benchmark evaluation. The reported outcomes are mainly LLM-judge scores and engineering measures; confidence intervals, variance estimates and formal significance tests were not reported.
The evidence does not show improved user satisfaction, empathy or real-world conversational outcomes. The supplied analysis describes no recruited participant cohort, prospective user study or quantitative live-user outcomes. It also leaves open whether the gains generalize to independently collected users and audio. The work is an arXiv preprint, version 1 dated 26 August 2026.
Paper data and sources
Original title: VoiceMem: Streaming Dual-Brain Memory for Real-Time Interaction
Authors: Zhifei Xie, Jiaqi Lang, Ze An et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text