hindsight-livekit: long-term memory for LiveKit voice agents

[namandidthis]

I built hindsight-livekit, which gives LiveKit Agents voice agents long-term memory across calls, using Hindsight.

pip install hindsight-livekit

Repo: https://github.com/namandidthis/hindsight-livekit (MIT)

What it does

• Works with both STT → LLM → TTS pipelines and realtime models (OpenAI Realtime, Gemini Live).
• When a call starts, it recalls what Hindsight knows about the caller and adds it to the agent's instructions.
• In pipelines, it also recalls memories relevant to each user message before the LLM replies.
• It saves each call to Hindsight as one document while the conversation happens.
• It includes hindsight_retain / hindsight_recall / hindsight_reflect tools, so realtime models can look things up mid-call.

memory = HindsightMemory(bank_id="user-123", document_id=ctx.room.name)
memory.attach(session)
ctx.add_shutdown_callback(memory.aclose)

await session.start(agent=HindsightAgent(memory=memory, instructions="..."), room=ctx.room)

Tested against a real Hindsight server and real LiveKit models. CI runs on livekit-agents 1.2 through the latest release, on Python 3.11–3.14.

Feedback and issues are welcome! Integrations Hub listing: #5809