Store Memory

What it does

An operations copilot for online stores that keeps the store's real context — catalog, recurring stockouts, supplier terms, pricing rules — in a persistent memory layer, so every session starts with full context instead of from zero. You teach it facts; later, a brand-new agent session with no chat history recalls exactly the facts it needs and reasons over them.

The thesis

The intelligence is the persisted memory, not the model. A fresh agent instance — after a restart or deploy — behaves as if it already knows the store, because the memory outlives the process. In the demo, session 2 is a new agent that only knows the store because the memory persisted in CockroachDB.

How we built it

• Memory: Amazon Bedrock Titan embeddings stored in a CockroachDB VECTOR(1024) column, recalled by nearest-neighbor with CockroachDB Distributed Vector Indexing (CREATE VECTOR INDEX + <->).
• Reasoning: Claude Sonnet 4.5 on Amazon Bedrock, over the recalled memory.
• Schema tuned with the CockroachDB Agent Skills Repo (skill cockroachdb-sql): its mandatory EXPLAIN step surfaced a covering-index optimization we applied (docs/agent-skills.md).
• Stateless agent (recall -> reason -> respond -> remember), so persistence is provably the database. Pluggable backends run offline on mocks and switch to the live stack via .env.

CockroachDB tools used (2 of 4)

Distributed Vector Indexing + Agent Skills Repo.

AWS

Amazon Bedrock — Titan Text Embeddings v2 + Claude Sonnet 4.5.

Built With

• amazon-bedrock
• claude
• cockroachdb
• flask
• python
• titan-embeddings
• vector-search