ClientOps Memory AI

ClientOps Memory AI is a persistent organisational-memory console for agencies and service businesses. Talal Khawaja and Aqeela Urooj built it for the CockroachDB × AWS "Build with Agentic Memory" hackathon on Devpost. It was a new project for that event, and all its demo data is synthetic.

Problem

In agency work, the real state of a client lives in fragments: a kickoff call, a routing rule someone mentioned in passing, a task promised in a thread, an exception agreed last week. A chat transcript can replay those fragments, but it can't say which instruction is current, which one was replaced, or why. That's what an operator needs when they come back to a client after a week away.

How it works

The operator talks to an agent in a Next.js console. The memory orchestrator classifies what matters into typed records:

• Episodic memory holds events, conversations and completed actions.
• Semantic memory holds stable facts, preferences and standing instructions.
• Decision memory holds approvals, rationale, provenance and explicit superseding links.
• Commitment memory holds owners, due dates, task state and completion evidence.

CockroachDB is the system of record, not a logging sidecar. It has 12 tables covering workspaces, clients, conversations, memories, embeddings, decisions, tasks, links, agent runs and retrieval events. Titan embeddings are stored as VECTOR(1024) with a vector index, and cosine distance ranks paraphrased queries. Nova Lite reasons over the ranked evidence. Each agent run is written with its retrieval trace in a short, retry-safe transaction, following Cockroach Labs' official Agent Skills guidance.

The demo shows the whole loop. The operator records a lead-routing rule for a synthetic roofing client, starts a fresh session, recalls it, adapts it with an exception, then asks why the behaviour changed. The agent cites both the current decision and the superseded one.

Key decisions and trade-offs

• Fact and inference are kept apart. Answers separate stored facts from AI inference and cite the source memory. Deleted or inaccurate memories drop out of retrieval.
• Honest degraded mode. The public Amplify demo is credential-free. It keeps synthetic memory in browser storage, and the UI and health endpoint say so. The live server path with Bedrock needs a runtime role, and it never fakes a Bedrock response.
• Scores kept separate. The ten-scenario harness is deterministic, and live infrastructure results are recorded separately rather than replaced with mocked scores.
• Stated limits. The live vector smoke test covered one embedded demo memory, and CockroachDB's managed MCP is described as optional and is not claimed.

Client work fragments across meetings, notes, handoffs, tasks and approvals. Chat history can repeat text, but it can't reliably answer what is current, what was superseded, who committed to what, and what evidence supports the answer.

Conversations are turned into four kinds of typed memory (episodic, semantic, decision and commitment) and stored in CockroachDB alongside vector embeddings. Amazon Bedrock (Nova Lite) reasons over retrieved evidence, and every answer shows its sources, confidence and timestamps. Superseded decisions stay queryable instead of being silently overwritten.

• The deterministic evaluation harness passes 10 of 10 long-term-memory scenarios, including paraphrased recall, superseding, new-session persistence and deleted-memory exclusion.