Vaos Voice Bridge

VAOS Voice Bridge

Real-time voice AI bridge implementing the Talker-Reasoner cognitive architecture from DeepMind's "Agents Thinking Fast and Slow" paper, combined with the CoALA framework for structured agent memory.

Two cognitive systems work in concert:

• System 1 (Talker) -- PersonaPlex/Moshi 7B on an NVIDIA GPU. Always-on, full-duplex voice conversation at ~200ms latency. Fast, intuitive, but limited to what a 7B parameter model knows.
• System 2 (Reasoner) -- Letta AI agent backed by a large LLM (Claude, Ollama, etc.) with persistent memory, web search, and tool use. Slow and deliberate, activated only when System 1 fails or the user requests it.

The bridge connects a browser client to both systems via an event bus, manages memory compression, and handles the handoff between fast intuition and slow reasoning.

                         +------------------+
                         |    Browser       |
                         | Opus mic stream  |
                         | AudioWorklet out |
                         | SpeechRecognition|
                         +--------+---------+
                                  |
                            WSS (TLS)
                                  |
                    +-------------+-------------+
                    |      Voice Bridge         |
                    |      (Bun + Hono)         |
                    |                           |
                    |  EventBus (X-Talk style)  |
                    |  talker.* | trigger.*     |
                    |  reasoner.* | memory.*    |
                    +---+-------------------+---+
                        |                   |
              WSS (Moshi binary)      REST API
                        |                   |
              +---------+-------+   +-------+--------+
              |  PersonaPlex    |   |  Letta Agent   |
              |  Moshi 7B       |   |  (System 2)    |
              |  (System 1)     |   |  + Claude/LLM  |