xzf-thu/VoiceMem
VoiceMem is an open-source memory layer for real-time voice agents, built around separate factual and emotional memory paths.
The repo describes a “dual-brain” architecture: a left brain for schema/entity-based facts and a right brain for personality, emotion, and relationships. It claims streaming retrieval can begin while the user is still speaking, with low added latency and small memory context injection. The project ships utilities, model-family adapters, ChatMem-400K, finetuning code, and reproducible evaluation scripts. The latest listed v0.0.2 update fixes event-date handling, removes a redundant right-brain memory class, and opens the speech synthesis layer. GitHub · LLM repos' note.
The repo describes a “dual-brain” architecture: a left brain for schema/entity-based facts and a right brain for personality, emotion, and relationships. It claims streaming retrieval can begin while the user is still speaking, with low added latency and small memory context injection. The project ships utilities, model-family adapters, ChatMem-400K, finetuning code, and reproducible evaluation scripts. The latest listed v0.0.2 update fixes event-date handling, removes a redundant right-brain memory class, and opens the speech synthesis layer. GitHub · LLM repos' note.
score 4