Keewano's $12M agent database asks for a second copy of your event stream

The problem the founders hit was in gaming, where a single user generates a history of interactions large and complicated enough that an agent asked why that user stopped paying has to walk the sequence itself ^[5]. Keewano's argument is that relational databases and warehouses were built for structured queries written by people, not for agents examining long sequences of behavior to work out why an outcome occurred ^[4]. The company, formally Sandstorm Ltd. and based in Tel Aviv ^[3], is selling the store it built for that job. "We realized that the current tabular architecture wasn't built for AI agents or machines, when it comes to analytics specifically, and reasoning in general," said co-founder and chief executive Mark Kardashov ^[6].

The deployment story is additive. KeewanoDB reads Parquet and Apache Iceberg and takes event streams through systems such as Kafka, and customers can send the same events to KeewanoDB for their agents and to Snowflake for conventional business intelligence ^[12]. "We can live alongside your existing database," Kardashov said ^[13]. The comparison a buyer runs is the cost of keeping two copies of one stream against the tokens an agent spends reading the warehouse directly ^[12][15].

Keewano puts each event at about four bytes, laid out on solid-state drives for reading ^[9], and says the system can query about 250 million events in less than half a second ^[11]. At four bytes an event, that is roughly one gigabyte of event data per query ^[1]. The performance figures have not been independently verified ^[11].

Take the 84% token saving at face value and the remaining spend is 16%, so a question that costs 100,000 tokens against the warehouse costs about 16,000 here ^[2]. Part of that comes from an in-database engine that runs Lua scripts the agent sends over the Model Context Protocol, filtering where the data sits so the model receives a smaller window of relevant context instead of the complete event history ^[16]. "Part of the reasoning happens inside the database and not by the engine," Kardashov said. "So the database becomes an extension of an agent" ^[17]. How much of that shows up in a given bill depends on how much of an agent's context today is events it reads and discards.

There are two products. KeewanoDB is the general-purpose store for agentic analytics and reasoning, and a separate analytics layer covers product analytics and real-time operational decisions ^[21]. That layer includes an agent called Signal that applies machine learning and statistical methods, tests its hypotheses and presents supporting event sequences as evidence ^[20]. A semantic engine builds a business layer from incoming events automatically, and customers can add mappings and business context by hand ^[18].

Two questions settle this for a team. The first is whether the agent's work depends on one entity's ordered history or on counts across many entities, since the warehouse already answers the second kind ^[4]. The second is what share of the tokens an agent spends today goes to events it reads and throws away, because that share is the ceiling on any saving from pushing the filtering into the store ^[16]. Where the answers are an ordered single-entity read and a large discard share, a second store at Keewano's per-active-entity pricing has a case to make ^[14].