Real-Time
Voice Agents.
Conversation dies in the pauses. Aetherfy runs your agent in the same region as its vector memory — replicated across regions from Performance up — so a voice agent can recall context mid-sentence without a cross-continent round trip.
The turn-taking latency budget.
Natural conversation tolerates only a few hundred milliseconds of silence. Every hop counts — so the memory lookup that grounds the agent's next reply cannot cross an ocean. Aetherfy keeps the vector replica in the same region as the agent, turning retrieval into a local read.
Retrieval served from the in-region replica — no cross-region fetch mid-conversation.
With keep_alive: true the machine never sleeps, so a call finds it running. Without it an idle agent is suspended and the first request adds about half a second. How a wake works.
Place the agent where your callers are — US East, EU Central, or Asia Pacific. How many regions a workspace can use depends on its plan.
Memory that keeps up with speech.
Stream the transcript, embed the turn, and retrieve relevant history — all inside the region hosting the call. The agent never waits on a distant database to remember what was said two sentences ago.
Always-on compute.
Set keep_alive: true and the agent stays running between calls. On a paid plan it is an add-on priced by memory; on Free the agent runs until 75% of the month is used, then pauses until the month ends; see pricing. Agents without it sleep when idle, which controls spend. The LiveKit and Pipecat frameworks run as a dockerfile service with more memory than Free allows, so they need a paid plan.
In-region memory.
Regional Qdrant replicas serve retrieval locally, so the lookup that shapes the next sentence is a same-region read.
Place it near callers.
Run in us-east-1, eu-central-1 or ap-southeast-1 — put the agent and its memory in the region closest to the people it talks to.