Overview
pumaDB is a hosted memory service for AI agents. It provides a simple shared place to save and reuse notes, facts, preferences, project context, transcripts, task state, and other useful memory — without setting up a database, vector store, or custom RAG stack. Agents connect via MCP (Streamable HTTP) or REST API, and pumaDB handles storage, versioning, and retrieval.
Hush is an open-source noise suppression model from weya AI. It removes background noise, competing voices, and audio interference from real-time calls, turning chaotic phone audio into clean, ASR-ready speech. Designed for voice AI agents, Hush runs on CPU with minimal latency (0.9 ms per 10 ms frame) and is small enough (8 MB) to deploy in any cloud or data center.
Feature Comparison
| Feature | pumaDB | Hush |
|---|---|---|
| Primary Function | Durable memory storage for AI agents | Real-time noise suppression for voice AI calls |
| Target Users | Developers building AI agents, chatbots, or workflows needing persistent memory | Voice AI teams, call centers, bot developers needing clean audio |
| Deployment Model | Hosted MCP server or REST API; no infrastructure to manage | Open-source model deployable on CPU in cloud or data center |
| Integration Method | MCP tool calls (remember) or REST endpoints (/v1/{table}) | Drop-in audio processing pipeline; processes 10 ms frames in under 1 ms on CPU |
| Data Handling | Stores JSON rows in tables; supports version history, natural language edits, viewer links | Processes real-time audio streams; outputs clean speech for downstream ASR or agents |
| Performance Metrics | 20 tables, 1,000 rows/table, 25 MB storage; 30 writes/min, 60 reads/min per key | Top-5 on Hugging Face Audio-to-Audio leaderboard; 0.9 ms per 10 ms frame on CPU |
| Safety & Recovery | Version history (last 10 versions, 30 days), rate limits, filtered cleanup, natural edits with opt-in | Not applicable (audio processing); focuses on real-time noise removal without data persistence |
| Open Source | No (hosted service) | Yes (open-source model on GitHub) |
Pricing
pumaDB is a hosted service with account limits: 20 tables, 1,000 rows per table, and 25 MB total storage. Rate limits are 30 writes per minute per key and 60 reads per minute per key. Pricing details are not explicitly listed on the website; it likely operates on a freemium or usage-based model.
Hush is open-source and free to deploy on your own infrastructure. weya AI also offers a 2-week pilot program where they implement Hush in one of your workflows, measure the lift, and then discuss full rollout. Pricing for the hosted version or pilot is not disclosed.
Pros and Cons
pumaDB Pros
- No database setup or infrastructure management required.
- Simple MCP and REST integration for agents and server-side apps.
- Built-in version history and natural language memory editing.
- Supports multiple memory types (preferences, task state, research clippings, etc.).
pumaDB Cons
- Limited storage (25 MB total) and row count (1,000 per table).
- Rate limits (30 writes/min per key) may constrain high-volume use.
- Not open-source; vendor lock-in for hosted service.
Hush Pros
- Open-source and lightweight (8 MB model) for easy self-deployment.
- Real-time performance on CPU (0.9 ms per 10 ms frame).
- Top-5 ranked on Hugging Face leaderboard for speech enhancement.
- Trained on 10,000+ hours of real-world noisy audio.
Hush Cons
- Focused solely on audio noise suppression; no memory or data storage features.
- Requires integration into existing voice AI pipeline; not a standalone service.
- Pilot program may have limited availability or require commitment.
Verdict
pumaDB and Hush serve entirely different purposes: pumaDB is a memory layer for AI agents, while Hush is an audio cleaner for voice AI. Choose pumaDB if your agents need persistent, easy-to-manage memory across sessions. Choose Hush if you need to improve call audio quality for ASR and agent performance. They can even complement each other — Hush cleans the audio, and pumaDB stores the resulting insights.
Hush
pumaDB