I built the best memory for voice agents while sponsoring Y Combinator hackathon. Introducing: Voice AI + Memory! We built a deep supermemory integration for Pipecat AI, so now even your agents have great memory With user profiles, these are almost-instant latency as well :)

Here's the demo

Here's the code and docs!

@ycombinator @supermemory @pipecat_ai bro sponsored the hackathon and decided to compete just cos he could

@ycombinator @supermemory @pipecat_ai User profiles are the key! We cache embeddings at the user level with Redis, then do hybrid search: exact match for session context + vector for long-term. Latency: ~80ms p95. The trick? Pre-warm the cache on first message of new conversation.

@ycombinator @supermemory @pipecat_ai memory is the moat. everyone can plug into the same LLMs, but the context you build up over time - user preferences, past interactions, learned patterns - that's what makes an agent feel like it actually knows you. stateless agents are just fancy search.

@ycombinator @supermemory @pipecat_ai GG

@ycombinator @supermemory @pipecat_ai 🔥

@ycombinator @supermemory @pipecat_ai Memory with a spark that feels human in voice chats

@ycombinator @supermemory @pipecat_ai not it writing the text in hindi lmao

@ycombinator @supermemory @pipecat_ai voice agents 😓

memory for voice agents is the hard problem nobody talks about enough. how are you handling the latency tradeoff between embedding lookup and keeping conversation flow natural? in my experience with voice pipelines, anything over ~200ms for memory retrieval breaks the conversational feel

@ycombinator @supermemory @pipecat_ai nice

@ycombinator @supermemory @pipecat_ai bro like just a casual ditl sponsored shipped

@ycombinator @supermemory @pipecat_ai Website down: This page isn't working at the moment can't currently handle this request. HTTP ERROR 500

@ycombinator @supermemory @pipecat_ai Really sorry about this

@ycombinator @supermemory @pipecat_ai voice agents without memory are basically talking to a goldfish every call. love that the user profiles keep latency low too, that's usually the tradeoff nobody talks about

@ycombinator @supermemory @pipecat_ai User profiles plus low latency is the right combo. Most voice agents still dump full context every turn, so this should move the needle for real conversations.

@ycombinator @supermemory @pipecat_ai Is it run locally ?

@ycombinator @supermemory @pipecat_ai livekit when?

@ycombinator @supermemory @pipecat_ai 🙌

@ycombinator @supermemory @pipecat_ai memory is what separates a toy demo from an actual useful agent, this is the right direction

@ycombinator @supermemory @pipecat_ai become a founder and futurologist, stop delivering outdated ideas to investors.

@ycombinator @supermemory @pipecat_ai Keep scaling Dhravya Bro Voice AI Phone Callers with memory's the future & it has arrived

@ycombinator @supermemory @pipecat_ai @DhravyaShah crazy shit!!

@ycombinator @supermemory @pipecat_ai could you build one for our social marketplace platform? Thanks

@ycombinator @supermemory @pipecat_ai This guy is tuff as hell

@ycombinator @supermemory @pipecat_ai User profiles with near instant latency are the right call for voice UX

@ycombinator @supermemory @pipecat_ai Nice! Heading to the demo now!

@ycombinator @supermemory @pipecat_ai Is it completely vibe coded

@ycombinator @supermemory @pipecat_ai ngl solving memory for voice agents is one of the hardest UX problems out there, near instant latency with user profiles is the right call. curious how you're handling memory decay over long sessions tho? that's where most solutions break down

@ycombinator @supermemory @pipecat_ai are you going to do a PR on pipecat repo?

@ycombinator @supermemory @pipecat_ai Great to see voice agents finally getting persistent memory. The coding agent space needs the same thing badly. Every Claude Code or Cursor session starts cold, no recall of past decisions. Same problem, different interface.

@ycombinator @supermemory @pipecat_ai keep building young man!

@ycombinator @supermemory @pipecat_ai Let’s gooooooo

@ycombinator @supermemory @pipecat_ai Since you clearly messed up your demo and I don't think the people understood what you were trying to show them, why couldn't you just re-record the video in a proper way?

memory isn't the bottleneck. everyone's obsessed with stateful agents but they're all building on top of models trained 98% on english internet. you can't synthetic your way out of that. try building a voice agent for swahili or marathi and watch it fall apart. the real problem nobody wants to admit: you need actual human data from actual humans speaking those languages. that costs money. labs would rather pretend it doesn't exist.

@ycombinator @supermemory @pipecat_ai Demo is impressive.

@ycombinator @supermemory @pipecat_ai memory game strong, latency solid