Build & validate
Case JSON, fixed participant voices, speech synthesis, transcription checks, and audio manifests.
02 / REPRODUCIBILITY
From audio to
long-term memory.
The research implementation spans dataset construction, speaker-aware ingestion, and evaluation. This site currently releases a one-case audio preview; the full research code is not published yet.
Identify & write
Online acoustic speaker identity, name grounding, addressee inference, and structured memory updates.
Answer & score
Question manifests, evidence-grounded answering, baselines, and speaker ablations.
The project repository contains this preview website and compressed example audio. Dataset-generation, memory-system, and evaluation code will be linked here when their release packages are ready.