Primary domain: Data & Analytics (Data Platforms)
Secondary domain: Artificial Intelligence & Machine Learning (Agentic AI)
Article type: Product & Vendor Innovation
Technology: PhoenixAI Contextbase (context infrastructure for AI agents)
Innovation maturity: Emerging
Mission: Enterprise/Operations
Business impact: Could cut the glue code and inconsistency that comes from keeping agent context in several separate tools
Audience: IT, Business/Leadership
Publication date: November 2026
Author / Contributor: Innovation team
Related technologies: PhoenixAI; StarRocks; Semantic SQL; agent memory; observability; ObserveStack
An AI agent is only as good as what it knows when it acts. It needs to know what the data means, what the business has learned before, what happened on its last run and whether that run worked. In most companies today, those pieces live in different tools. On November 4, 2026, PhoenixAI launched a product that tries to put them in one place: Contextbase.
PhoenixAI, the original creator of the StarRocks open source database and formerly known as CelerData, unveiled Contextbase at its NEXT QUERY 2026 summit in Menlo Park, California. It calls the product “context infrastructure for AI agents.”
The problem: context scattered across tools
According to PhoenixAI’s release, a typical agent stack keeps definitions in a semantic layer, documents in a vector store, memory in a separate memory service, and traces and evaluations in observability tools. Each system holds its own copy of the information and enforces its own permissions.
The company’s argument is that an agent has to stitch all of this together on every turn, and what it gets back is partial, inconsistent and sometimes stale. James Li, PhoenixAI’s co-founder and CEO, put it bluntly in the announcement: context is the next great infrastructure layer, and it belongs in one real-time system where it compounds.
What Contextbase stores
Contextbase keeps five kinds of context together:
- Semantics: semantic views of entities, dimensions and metrics, defined by the business in Semantic SQL
- Knowledge: typed, versioned records such as a spec, an incident review or a playbook, linked to one another
- Memory: notes that agents write down from earlier work
- Traces: a record of every model call and tool call in a run, including prompts, outputs, tokens and cost
- Evaluations: test datasets, eval runs and scores on live traffic
Everything in the system is linked, versioned, governed and directly readable by agents, the company says. It also describes a design goal that is easy to appreciate: because the agent is the non-deterministic part of the system, the place it reads from should be steady. Definitions are verified, the system enforces policies, and records are kept as written.
Built into the database
A notable design choice is where Contextbase lives. PhoenixAI says it is built into the PhoenixAI database, so context sits on the same cluster as the business data and a single query can reach all of it under one set of permissions. PhoenixAI describes its platform as an agentic AI database that gives autonomous agents sub-second access to live enterprise data, running in the customer’s own cloud account, in a self-managed private cloud or on premises.
On top of Contextbase, the company is building applications, starting with ObserveStack, which traces and evaluates agents. The idea is that teams can see what each run did and score whether it worked, using the same store their agents read from. Li described the result as a flywheel: every run makes the context better, and every agent shipped makes the next one smarter.
What it means for data teams
Whatever one thinks of the flywheel claim, the launch reflects a real shift in how data platforms are being pitched. As agents move from answering questions to taking actions, the supporting data has to include not only tables but also meaning, memory and an audit trail. Putting those under one permission model could simplify governance, since access rules would not need to be copied across four or five systems.
Buyers will still have questions the announcement does not answer. How does Contextbase handle very large volumes of agent memory? How easily can teams move existing semantic definitions into Semantic SQL? And how does a single store compare with best-of-breed tools that teams already run? Contextbase is in private preview as of November 4, so independent results and customer experience are not yet available, and the company’s claims should be treated as vendor statements.
The bottom line
Contextbase is a bet that agent context deserves its own governed layer, sitting next to the data it describes. For teams building agents, the practical takeaway is to look at how many separate systems their agents depend on and where permissions and history diverge. Teams can request a free trial through PhoenixAI.