Inspiration
We were inspired by a simple but important problem: modern AI agents are powerful, but most of them are still ephemeral. They can generate impressive responses in the moment, yet they forget context the second the session ends. In real-world systems, that is not enough. Agents need durable memory, not just a large context window.
We wanted to build something that shows what agentic memory actually looks like in production: an AI system that observes events, stores data reliably, learns patterns over time, and adapts based on what it has seen before. CockroachDB and AWS gave us the perfect foundation to make that idea real.
What it does
Cockroach Sentry is an agentic monitoring system that continuously observes its environment and builds a memory of what is happening over time.
At a high level, it:
- detects motion using a camera and sensor pipeline,
- captures an image,
- uses Amazon Bedrock’s vision model to identify whether an insect is present,
- stores each detection in CockroachDB with metadata like location, timestamp, temperature, and humidity,
- creates embeddings for recurring patterns,
- searches historical observations semantically,
- and generates hypotheses about behavior and trends over time.
Instead of acting like a one-off chatbot, the system behaves like an autonomous agent with memory. It can remember where activity occurs, when it happens, and what patterns are emerging.
How we built it
We built the project using CockroachDB as the persistent memory layer and AWS services for AI inference and storage.
The system architecture includes:
- AWS Bedrock for image analysis and reasoning,
- Amazon S3 for storing snapshots,
- CockroachDB for structured event storage, vector embeddings, and hypothesis tracking,
- Python scripts to orchestrate the full pipeline from detection to memory updates.
The data model is designed around multiple memory layers:
- raw detections for event history,
- summarized observations for recurring patterns,
- embeddings for semantic search,
- hypotheses for evolving agent beliefs.
This allows the agent to move from “I saw a bug” to “I understand this pattern and can reason about it over time.”
Challenges we ran into
One of the biggest challenges was designing the system as a genuine memory architecture rather than just a simple AI workflow. We had to think carefully about:
- how to persist event data reliably,
- how to structure embeddings and semantic retrieval,
- how to store evolving hypotheses without losing traceability,
- and how to make the agent useful beyond a single inference.
Another challenge was making the pipeline robust enough to run in a real environment with sensor input, image capture, model inference, and database writes. We had to handle time delays, data quality issues, and ensure the detection flow remained reliable and repeatable.
Finally, integrating AI reasoning with a transactional database required us to balance speed, reliability, and production-grade design.
Accomplishments that we’re proud of
We’re proud that we built a working agentic system that demonstrates persistent memory in a real use case.
Some of our biggest accomplishments include:
- successfully integrating AWS Bedrock with CockroachDB,
- storing detections, summaries, embeddings, and hypotheses in a single memory system,
- building a workflow from sensor-triggered image capture to AI-powered classification,
- and creating a system that can reason over historical data instead of only reacting to the latest event.
We also created a project that clearly maps to the hackathon theme: agentic memory as a first-class design principle.
What we learned
This project taught us that memory is not an afterthought in AI systems — it is the foundation of real autonomy.
We learned that:
- reliable storage matters just as much as model quality,
- persistent memory enables contextual reasoning over time,
- and AI agents become far more useful when they can recall past events and update beliefs as new evidence arrives.
We also learned the importance of designing for observability, traceability, and system reliability. An agent is only as trustworthy as the data it keeps and the logic it uses to update its understanding.
What’s next for Cockroach Sentry
The next step is to expand Cockroach Sentry from a smart monitoring prototype into a more general agentic memory platform.
We envision:
- multi-zone and multi-device monitoring,
- more advanced agent actions based on learned hypotheses,
- richer retrieval and semantic search over historical patterns,
- and broader integrations with other AWS services and agent workflows.
In the long term, we want to generalize this beyond home monitoring and show how CockroachDB can serve as the memory layer for autonomous agents in other domains such as operations, security, logistics, and robotics.
Our vision is simple: build agents that do not just think in the moment, but remember, learn, and act reliably over time.
Built With
- agent-memory
- agentic-ai
- ai
- amazon-web-services
- architecture
- autonomous-agents
- bedrock
- cloud
- cockroachdb
- computer-vision
- database
- event-detection
- image-recognition
- iot
- machine-learning
- monitoring
- postgresql
- python
- rag
- real-time-systems
- semantic-search
- sensors
- serverless
- smart-home
- vector-database
Log in or sign up for Devpost to join the conversation.