Universal streaming dual-brain memory infrastructure for real-time voice agents.
Universal memory infrastructure separating factual info and emotional attributes for real-time voice agents.
VoiceMem provides universal long-term memory for voice agents using a streaming dual-brain architecture. It is built for developers creating interactive voice characters or metaverse agents who need low-latency, emotionally aware context.
It divides memory into a left brain for facts and a right brain for emotional and personality attribution. A fully streaming architecture allows it to prefetch and query memory within a 0-300 ms window before speech completes.
It suits developers building custom voice agents or metaverse NPCs requiring persistent user memory. It does not suit those seeking a drop-in text-only database or projects unwilling to manage speech-to-text and embedding pipelines.
Separates fact management in the left brain from emotional and personality tracking in the right brain.
Performs speculative queries via streaming audio input so memory is ready before the user finishes talking.
Processes raw audio directly to capture voiceprints, speaker IDs, scene contexts, and emotional cues.
Injects only Top-K relevant memories into context to minimize LLM token overhead during live sessions.
Free under its open-source license, with the only cost coming from running the local inference hardware and optional OpenAI API usage for extraction.
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
These reviewed AI tools do the same job instantly — from our scored directory.
The left brain handles schema and entity-based facts, while the right brain manages long and short-term emotional attribution.
Yes, the architecture is entirely decoupled, allowing you to swap out components including the underlying memory engine.
See Zekai first in Google