MemoAir Voice Memory

Drop-in persistent memory for voice agents. Sub-10 ms local recall, cross-session continuity, per-user isolation, and an SDK-managed runtime pool — ships with a first-class LiveKit integration and a framework-neutral custom-agent pattern.

v0.3.2: memoair-voice now ships on npm for Node/TypeScript agents and PyPI for Python. Both use per-call user={id,name,metadata}routing through a bounded local runtime pool. LiveKit has drop-in adapters for Python and TypeScript; Pipecat, Vapi, and Retell use the custom-agent pattern until dedicated adapters land.

Quickstart API Reference

Start here

5-minute Quickstart

Install the SDK, grab keys from the dashboard, and call search_memory from any LLM. The local runtime is bundled — no Docker.

Get started

Concepts

Profile, working brief, permanent facts, and org context — how the four lanes work together.

Learn the model

Integrations

LiveKit is available now. Pipecat, Vapi, and Retell use the custom-agent pattern in v0.3.1.

Wire your stack

Multi-workspace SaaS

Run voice agents for multiple tenants with per-client and per-end-user memory isolation.

Deployment guide

SDK References

MemoAirVoiceClient, MemoAirLiveKitAgent, tools, and result types for Python and TypeScript.

Python SDK JS/TypeScript SDK

HTTP API Reference

Loopback runtime wire contract — health, session lifecycle, search_memory, turn/after.

Wire contract

What you get

•Local-first recall. Every search_memory call completes on-host. No network hop on the audio path.
•Cross-session continuity. Permanent facts extracted from prior calls land back on disk before the next call starts.
•Per-user isolation by construction. The SDK pool spawns and reuses one local runtime per (project, user) while keeping a hard cap on warm processes.
•LiveKit drop-in + custom framework path. Install memoair-livekit for the drop-in MemoAirLiveKitAgent. Non-LiveKit frameworks use MemoAirVoiceClientdirectly with per-call user={...}.