MENLO PARK – November 04, 2026 -- PhoenixAI, the creator of the StarRocks open source project, launched Contextbase, a system that stores semantics, knowledge, memory, traces and evaluations in a single governed database for AI agents. The company unveiled the product at its NEXT QUERY 2026 Data + AI Infrastructure Summit. Private preview opens November 4.
Contextbase replaces four separate tool categories with one store
PhoenixAI says most agent stacks scatter context across a semantic layer for definitions, a vector store for documents, a memory service, and observability tools for traces and evaluations. Each keeps its own copy of the data and its own permissions. Contextbase consolidates them.
The system holds five types of context:
- Semantic views of entities, dimensions and metrics, defined in Semantic SQL
- Typed, versioned knowledge records such as specs, incident reviews and playbooks, linked to each other
- Memory written by agents from earlier work
- Traces of every model and tool call, including prompts, outputs, tokens and cost
- Evaluations, from test datasets and eval runs to scores on live traffic
Context shares a cluster and permission set with business data
Contextbase is built into the PhoenixAI database, so context sits on the same cluster as business data. A single query can reach all of it under one set of permissions, according to the company. Records are linked, versioned and governed, and agents can read them directly. Definitions are verified, the system enforces policies, and records are kept as written.
PhoenixAI describes the agent as the non-deterministic part of the stack, which is why the store it reads from is designed to stay steady.
CEO and co-founder James Li said, "Context is the next great infrastructure layer." He added that each run improves the store and each shipped agent makes the next one smarter.
ObserveStack becomes the first application built on Contextbase
PhoenixAI is building applications on Contextbase, starting with ObserveStack for tracing and evaluating agents. Teams can see what each run did and score whether it worked, using the same store their agents read from.
Teams can request a free trial of the private preview now. PhoenixAI, formerly CelerData, is headquartered in Menlo Park, California.