ClientOps Memory AI

Inspiration

Agencies lose the reasoning behind client decisions as work moves across chats, meetings, tools, and staff handoffs. ClientOps makes that context durable, current, and auditable.

What it does

ClientOps extracts typed organizational memory, reconciles conflicts without erasing history, retrieves relevant evidence across sessions, tracks commitments, and explains why each answer was produced.

How we built it

CockroachDB Cloud stores relational truth, memory provenance, decision history, retrieval traces, and Amazon Titan embeddings in VECTOR(1024) with Distributed Vector Indexing. The official CockroachDB Agent Skills Repo was installed project-locally: cockroachdb-sql reviewed the schema and vector queries, while designing-application-transactions shaped the short SQLSTATE 40001 retry-safe transaction that atomically records each agent run and retrieval trace. Amazon Bedrock uses Nova Lite for grounded reasoning and Titan Text Embeddings V2 for semantic retrieval. AWS Amplify hosts the public synthetic judging flow.

Challenges

Preserving historical truth while applying new instructions required explicit superseding relationships, status-aware ranking, and visible evidence rather than silently mutating prior decisions.

Accomplishments

A live CockroachDB schema and vector smoke test, live Bedrock inference and embeddings, cross-session recall, semantic paraphrase retrieval, responsible-memory controls, and a ten-scenario evaluation harness.