polarityinc/zenith

ZenithDB is an open-source database built specifically for storing and querying AI agent execution traces, the detailed logs of everything an AI agent does during a task. When an agent runs, it generates long sequences of events: model calls, tool invocations, intermediate results, errors, and metadata. Existing observability databases (designed for tracking web requests and server performance) are not a good fit for this data because AI agent traces are much longer, more loosely structured, and have very different query patterns. ZenithDB addresses this with a columnar storage engine (a database design that stores data column by column rather than row by row, which makes analytical queries much faster) optimized for several specific properties of agent trace data: the traces are long and sparse, they contain large text fields, they often have late-arriving annotations added after the fact, and they arrive in bursts. Key design decisions include grouping all spans (individual events) from one trace into the same storage chunk for efficient retrieval, never loading unused columns during queries, and embedding a full-text search capability directly inside the storage segments. The database accepts data in standard observability formats and can be queried using SQL. It runs as an HTTP and gRPC server, starts in seconds with no external dependencies for development, and supports production deployments backed by Postgres and cloud object storage. A web dashboard is also included. The project is currently in alpha, the engine is feature-complete but the storage format may still change before version 1.0.