pumaDB is a specialized memory layer for AI agents designed to handle durable state without requiring users to set up or maintain complex database infrastructure. Created as a lightweight solution, it provides a hosted memory API that allows agents and applications to store, retrieve, and manage JSON-based records. By acting as a simple, explicit memory primitive, pumaDB enables developers to focus on building agent capabilities rather than managing backend systems, offering both a REST API for server-side applications and a Model Context Protocol (MCP) server for direct agent integration. The service is built to be intentionally small and manageable, catering to needs where explicit state—such as user preferences, project conventions, and task progress—must persist across different sessions.

The functionality of pumaDB centers on providing a stable, version-controlled repository for small JSON objects. It manages these records through named tables, enabling agents to perform common CRUD operations. The service handles the underlying infrastructure, ensuring that memory remains available and durable. It also includes built-in safety features like versioning, which archives previous row states before updates or deletions, allowing users to restore previous data within a 30-day window. With support for both RESTful requests and MCP tool calls, it integrates seamlessly into various agent environments including web-based clients and server-side backends.

Some of the key features are:

• Hosted MCP Integration: Offers a standardized way for AI agents like ChatGPT, Claude, and Codex to connect and access memory.
• REST API Support: Enables secure access for server-side applications, workers, and CLI tools via bearer authentication.
• Durable Versioning: Automatically archives previous versions of data, allowing for easy recovery of overwritten or deleted records.
• Inert Safety Metadata: Provides tools for storing sensitive or executable content like code, commands, and markdown in an inert format that is safe for review.
• Lightweight Schema: Uses a simple JSON-based table structure that requires no complex setup or schema definition.
• Natural Language Compatibility: Supports tools that allow agents to translate conversational requests into structured memory updates.
• Collaborative Viewer Links: Generates short-lived links for inspecting or downloading larger memory result sets and JSON records.
• Atomic Batching: Allows multiple write operations to be performed in a single, atomic request to maintain consistency.

pumaDB operates by serving as an external memory backend that agents call during their task cycles. Developers initialize the service by setting up an account and creating API keys for their specific applications. When an agent needs to remember a piece of information, it invokes an MCP tool (such as 'remember' or 'add') which pushes data to the pumaDB servers. For backend-heavy applications, developers use the REST API to interact with these tables directly. The system enforces account, table, and rate limits to ensure stability, while providing developers with full transparency and control over their data through auditing, versioning, and explicit deletion commands.

Some common use cases include:

• Agent Preference Storage: Maintaining user communication styles, formatting defaults, and review expectations across multiple chat sessions.
• Project Task Tracking: Storing the status, blockers, and progress of long-running tasks for agents working on complex technical projects.
• Knowledge Management: Keeping research clippings, project facts, and repository conventions easily accessible to agents during research investigations.
• Tool and Workflow Notes: Storing reusable operating instructions and specific workflow notes that agents can reload in subsequent sessions to improve accuracy.
• Config and Code Snippets: Archiving configuration settings and code snippets as inert text records for developers to audit and review before execution.