Karthik Ranganathan

Karthik Ranganathan is a co-founder of Yugabyte, where he has helped build a distributed database that brings PostgreSQL’s familiar programming model to applications running across machines and regions. His work spans Facebook’s database infrastructure, the creation of YugabyteDB, and an effort to make knowledge reusable across teams of AI agents.

From Facebook infrastructure to YugabyteDB

From 2007 to 2013, Ranganathan worked on Facebook’s infrastructure team alongside future co-founders Kannan Muthukkaruppan and Mikhail Bautin. They built and operated Apache Cassandra and Apache HBase as the company’s services and user base expanded. Scaling those systems exposed a persistent problem: developers still needed secondary indexes and transactions, even when their databases had been designed around other priorities.

Ranganathan subsequently worked at Nutanix before starting Yugabyte with Muthukkaruppan and Bautin in 2016. The three began the database project that February, following conversations about what cloud applications would require. They wanted to combine distributed operation with the querying and correctness guarantees developers expected from relational databases. Ranganathan’s account of YugabyteDB’s origins connects that ambition to the limitations they had encountered at Facebook.

Yugabyte emerged from stealth with its first public beta in November 2017. Ranganathan served as CTO and was co-CEO by his 2026 presentation on agent learning. As YugabyteDB developed, PostgreSQL compatibility became central to its approach: applications should gain resilience and room to grow while retaining familiar database behavior.

Database behavior and customer choice

• PostgreSQL runtime compatibility: Ranganathan treats compatibility as a question of application behavior, extending beyond whether a database accepts the same SQL syntax. YugabyteDB reuses PostgreSQL’s query layer over its own distributed storage engine. In a 2024 release account co-authored with Suda Srinivasan, he described improvements to transactional behavior, retries, query planning, and data placement. Colocating tables can reduce latency; distributing them can accommodate larger workloads. The engineering challenge is preserving the developer’s expectations as the underlying system changes.

• Open source as customer optionality: Ranganathan argues that an unrestricted database gives customers a practical alternative if their commercial relationship with a vendor deteriorates. Yugabyte moved previously closed enterprise database features under Apache 2.0 in 2019. His case for the open-source business model rests on reducing adoption friction and letting broad usage improve the software through feedback, integrations, and contributions. Openness is part of his commercial reasoning as well as his product philosophy.

From agent memory to shared learning

Ranganathan’s work on Meko extends his concern with shared state to AI agents. His 2026 argument distinguishes an agent’s ability to remember from a team’s ability to learn together. A handoff can preserve an output while losing the reasoning, failed approaches, and context behind it. The next agent then spends tokens reconstructing work already done. More stored information or a larger context window does not, by itself, solve that coordination problem.

Two ideas shape his approach:

• Reusable project knowledge: Meko preserves memory, shared knowledge, conversations, and decision traces. Its project-scoped datapacks collect information for reuse across workflows and sessions. Ranganathan emphasizes repeatable work: accumulated knowledge should help the next agent rather than remain confined to the agent that acquired it. Shared context also addresses divergent state, where one agent updates a fact while another continues using an older version.

• Learning that can be investigated and corrected: Ranganathan’s approach to agent memory acknowledges that stored information can be wrong. Meko separates local memory from shared knowledge, requires explicit promotion into organizational knowledge, and tracks where information originated. These mechanisms help teams locate an error, limit its spread, and remove it; they do not make stored information automatically true. The emphasis carries his database experience into AI: when many participants depend on shared state, consistency, permissions, and the ability to investigate mistakes become essential parts of the product.