pumaDB

Product Introduction

1. Definition: pumaDB is a hosted, lightweight JSON memory layer specifically designed for AI agents. It functions as a durable, managed backend service that allows agents to persist and retrieve structured data—such as facts, preferences, and state—without requiring users or developers to manage database infrastructure.
2. Core Value Proposition: pumaDB solves the critical problem of context loss in AI agent workflows. It provides a simple, managed memory surface for agents to save and reuse information across sessions, tools, and chats, eliminating the need for manual note-taking or complex database, vector store, or custom RAG stack setups.

Main Features

1. Hosted MCP Integration: A fully managed, OAuth-ready Streamable HTTP Model Context Protocol endpoint (https://api.pumadb.ai/mcp). This allows AI agent clients like Codex, ChatGPT, and Claude to directly connect, authenticate, and perform memory operations (e.g., remember, query) via standardized tool calls. It supports dynamic client registration and is the zero-config path for agent-based integrations.
2. Server-Side REST API: A straightforward HTTP API (/v1/) for backend, serverless, or script-based applications using bearer API keys. It provides direct endpoints (GET, POST, DELETE) for managing named JSON rows within scoped tables, enabling developers to build custom agent memory systems on trusted server environments.
3. Lightweight JSON Memory Schema: A purpose-built storage format for agent context. pumaDB is optimized for storing and retrieving small, explicit JSON records categorized into types like user preferences, project context, task state, research clippings, and reusable skill instructions. It features scoped limits (20 tables, 1,000 rows/table, 25MB total) to keep memory focused and manageable.
4. Safety & Versioning Rails: Built-in memory governance tools. This includes per-key rate limits (30 writes/min, 60 reads/min), filtered cleanup, natural language edits (e.g., updating a row via a plain-language instruction without duplicates), and automatic version history. The last 10 versions of any row are archived for 30 days and can be restored with a single call.

Problems Solved

1. Pain Point: The primary pain point is context fragmentation and loss in AI agent systems. Agents lose useful memory between sessions, across different tools, or when conversations change, leading to repetitive requests, loss of user preferences, and inability to maintain project state. Existing solutions are either too manual (copying notes) or too complex (setting up a full database and RAG pipeline).
2. Target Audience: This product is aimed at AI application developers, engineering teams building agent workflows, and technical creators using platforms like Codex, ChatGPT, or Claude who need persistent memory for their AI assistants. It's for anyone needing durable, queryable state for agents without operational overhead.
3. Use Cases: Essential scenarios include remembering user preferences across chat sessions, maintaining task state for long-running processes, storing project conventions and architectural notes, accumulating research findings over time, and enabling seamless handoff context between different agents or sessions. It acts as a "scratchpad" and "notebook" for AI.

Unique Advantages

1. Differentiation: Unlike traditional databases (e.g., PostgreSQL, MongoDB) or vector stores (e.g., Pinecone), pumaDB requires zero infrastructure setup, configuration, or maintenance. It is a purpose-built, hosted service with a simple schema, contrasting sharply with the complexity of managing your own database or building a custom RAG stack. It provides both MCP (for agents) and REST (for apps) interfaces to the same data surface.
2. Key Innovation: The key innovation is the convergence of a managed hosting model, an agent-optimized API (MCP), and a deliberately constrained, reviewable memory schema. This combination enables "memory-as-a-service" for AI agents, where durability, accessibility, and safety (via versioning and limits) are handled out of the box, allowing developers to focus on agent logic rather than data plumbing.

Frequently Asked Questions (FAQ)

1. What is pumaDB and what is it used for? pumaDB is a hosted JSON memory layer for AI agents. It is used to store and retrieve durable context like user preferences, project facts, and task state, enabling agents to maintain memory across sessions and chats without database setup.
2. How do I integrate pumaDB with my AI agent (like ChatGPT or Claude)? The fastest way is to use the hosted MCP endpoint. Add pumaDB as a custom connector in your agent's settings using the server URL https://api.pumadb.ai/mcp, complete the OAuth flow, and the agent can then use tools like remember and query to interact with its memory.
3. What is the difference between pumaDB and a traditional database? pumaDB is a specialized, lightweight memory layer for agents, not a general-purpose database. It has built-in limits (e.g., 1,000 rows per table, 25MB total storage), version history, and a simple JSON row model optimized for agent context. It requires no setup or management, unlike traditional databases.
4. Is pumaDB free? How are API calls billed? The product page mentions account limits (20 tables, 25MB storage) and rate limits (30 writes/min, 60 reads/min), suggesting a freemium or tiered model. Specific pricing and billing details for API calls beyond these limits would be found on their official website or in their documentation.
5. Can I use pumaDB for server-side applications or only for AI agents? You can use both interfaces. The hosted MCP endpoint is designed for direct agent-client connections. For server-side applications, you use the REST API with a bearer API key (puma_live_*) from a trusted backend environment, like a serverless function or Node.js app.