AI agents are moving fast from demos into production. They’re writing code, running pipelines, diagnosing incidents. But there's a catch: most of them forget everything the moment a session ends. So this summer, CockroachDB challenged developers worldwide to fix that. The CockroachDB × AWS Hackathon: Build with Agentic Memory, hosted on Devpost, asked builders to create agentic applications using CockroachDB as a persistent memory layer - storing conversation history, task state, embeddings, or transactional data - deployed on AWS.
Entrants had to put at least two CockroachDB tools to work (the Cloud Managed MCP Server, Distributed Vector Indexing, the agent-ready ccloud CLI, or the open-source Agent Skills Repo) alongside at least one AWS service like Bedrock, Lambda, or S3.
The response blew past expectations: roughly 3,700 developers registered and more than 600 projects were submitted. Our judging panel: myself, Harsh Shah, David Joy, Kikia Carter, Jim Hatcher, and Virag Tripathi. We narrowed that field down to 78 finalists before selecting the top three, scored on:
Agentic memory design
Technical implementation
Real-world impact, and production readiness.
1st Place: Anchor - Viraj Chogle 
Anchor is an on-call agent for CockroachDB Cloud clusters. It diagnoses incidents, fixes them through the ccloud CLI, and actually remembers what it did. So the third time the same incident shows up, it gets resolved faster than the first, which is more than most of us can say after a 3am page.
Under the hood, it stores that memory in CockroachDB's distributed vector index and searches it by similarity, recency, and salience, with AWS Bedrock handling embeddings. Memory writes and remediation actions land in the same transaction, so Anchor never ends up remembering a fix it didn't actually apply, or applying one it forgot to write down.
2nd Place: Unsay - Jonathan Andrei
Unsay is a medication-safety agent that answers questions from live FDA drug data, then does something most chatbots never do: takes it back. When the data changes, so does the answer. Built on a bitemporal memory schema in CockroachDB and running on AWS Lambda, Unsay can spot when a new FDA recall invalidates something it said last week, track down everyone it told, and draft the corrections. Turns out "I was wrong, here's the update" is a feature, not a bug.
3rd Place: Interlock - Ujwal Vanjare and Arpita Kalburgi
As more teams point multiple AI coding agents at the same codebase, a new failure mode has shown up: agents stepping on each other's changes. Interlock plays referee. Instead of blocking the merge or throwing out an entire branch when two agents collide, it figures out what the conflict actually broke, fixes only that, and leaves the rest alone. Surgical, not scorched earth.
It's a practical answer to a problem every team running agents at scale is about to hit, and a good reminder that memory and state tracking have uses beyond remembering what someone said in a chat.
Thank you!!
To everyone who registered, shipped a project, and stayed up late arguing with a vector index - thank you. And to our judges, who volunteered real time out of very busy schedules to review dozens of submissions each: we couldn't have pulled this off without you.
Congratulations to Anchor, Unsay, and Interlock. We can't wait to see what you build next!
