🧠 OpsMemory AI

"OpsMemory doesn't just retrieve documents. It remembers operational experience."

CockroachDB Distributed Vector Index CockroachDB Cloud MCP Server Amazon Bedrock Nova Pro AWS Lambda × 6 Amazon S3 CloudWatch SQS
Ready — click Start Demo
60s
1 incident
—
—
idle
1
🚨 New Incident Arrives — INC-1947
pending
Lambda: opsmemory-orchestrator | Route: POST /incidents
Service: payment-processor
Symptoms:
• Messages are repeatedly becoming visible
• Lambda processing appears to restart
• Queue depth is growing
• Some messages are processed more than once

⚠️ Note for judges: This wording is intentionally DIFFERENT from INC-1832. No mention of "timeout" or "backlog". A keyword search would never connect these. Only semantic vector search can.
What happens when you click Start:
  • HTTP POST to API Gateway → triggers opsmemory-orchestrator Lambda
  • Orchestrator validates input, generates symptoms embedding text
  • INSERT into CockroachDB incidents table
  • Returns incident with status: "open"
2
🔍 AWS Telemetry Investigation
pending
Lambda: opsmemory-telemetry | AWS: CloudWatch GetMetricData, FilterLogEvents
Agent fetches real-time metrics from CloudWatch for the affected resources.
What happens behind the scenes:
  • Orchestrator invokes opsmemory-telemetry Lambda
  • Telemetry Lambda calls CloudWatch GetMetricData for SQS + Lambda metrics
  • Fetches: ApproximateNumberOfMessagesVisible, Duration, Errors, Throttles
  • Calls FilterLogEvents for Lambda error logs
  • Parses anomalies: queue depth > 100, duration near timeout
  • Returns structured telemetry summary for the reasoning engine
3
🧠 CockroachDB Vector Memory Search
pending
Lambda: opsmemory-memory-search | CockroachDB: HNSW Vector Index + MCP Server
THIS IS THE CORE DIFFERENTIATOR. The agent searches its experiential memory using CockroachDB Distributed Vector Indexing via the MCP Server interface.
What happens (CockroachDB + Bedrock integration):
  • Orchestrator invokes opsmemory-memory-search Lambda
  • Lambda calls Amazon Bedrock Titan Embed Text v2 to generate 1024-dim embedding of current symptoms
  • Embedding vector sent to CockroachDB via MCP Client interface
  • SQL: SELECT ... FROM incident_embeddings e JOIN incidents i ... ORDER BY e.embedding <=> $1::vector
  • HNSW vector index returns nearest neighbors by cosine distance
  • Converts distance to similarity: similarity = 1 - cosine_distance
  • Filters by threshold (0.5) — returns matches with full historical context
  • INC-1832 found at 86.4% similarity despite zero keyword overlap!

🧠 OPERATIONAL MEMORY MATCH

—%
INC-1832
payment-processor • Resolved July 15, 2026
Previous symptoms: Lambda executions timing out; SQS queue backlog increasing; Processing latency increasing
Root cause found: SQS visibility timeout (60s) too short for Lambda processing time (~120s)
Remediation applied: Increased visibility timeout from 60s → 180s
Outcome: ✅ Successful — backlog cleared within 5 minutes
Lessons learned: Visibility timeout should be ≥6× average processing time
4
🤖 Amazon Nova Pro AI Diagnosis
pending
Lambda: opsmemory-reasoning | AWS: Bedrock InvokeModel (amazon.nova-pro-v1:0)
Nova Pro receives: current symptoms + telemetry + historical experience from INC-1832.
It reasons about the problem informed by what worked before.
What happens (Bedrock Nova Pro integration):
  • Orchestrator invokes opsmemory-reasoning Lambda
  • Lambda builds structured prompt with: incident data, CloudWatch telemetry, and memory matches
  • Calls bedrock:InvokeModel with model amazon.nova-pro-v1:0
  • System prompt instructs: "Cite historical incidents, provide confidence 0-1, be specific about AWS API calls"
  • Temperature: 0.1 (low = consistent, reliable reasoning)
  • Nova Pro outputs structured JSON: diagnosis, root cause, confidence, remediation with exact parameters
  • Response parsed and validated; fallback if JSON extraction fails
5
✋ Human Approval Gate
pending
Security: Human-in-the-loop | CockroachDB: approvals table
The agent NEVER modifies infrastructure autonomously.
Every destructive action requires explicit human approval. This is a core safety principle.
What happens:
  • Agent state transitions to "awaiting_approval" — stored in CockroachDB agent_runs table
  • Recommendation displayed to operator with full context
  • Approval recorded in CockroachDB approvals table with: who approved, when, what was proposed
  • Audit trail maintained — every action is traceable
  • If rejected: agent stops, incident marked as "blocked", no infrastructure change
6
⚡ Remediation Execution
pending
Lambda: opsmemory-remediation | AWS: SQS SetQueueAttributes
Agent executes the approved fix against real AWS infrastructure.
What happens (actual AWS change):
  • Orchestrator invokes opsmemory-remediation Lambda
  • Lambda validates action is in allowed registry (sqs_update_visibility_timeout)
  • Captures previous state via sqs:GetQueueAttributes (for rollback)
  • Executes sqs:SetQueueAttributes → VisibilityTimeout = 180
  • Returns: {success: true, previousState: {timeout: 60}, newState: {timeout: 180}}
  • IAM role scoped to ONLY the payment-queue ARN — least privilege
7
✅ Verified & Experience Stored
pending
Lambda: opsmemory-verification + opsmemory-memory-search | CockroachDB: incident_embeddings
Agent verifies the fix, then stores INC-1947 as new memory. Next time, it will remember both incidents.
What happens (memory growth):
  • Verification Lambda checks: SQS queue attributes match expected state
  • CloudWatch metrics confirm: queue depth stabilizing, no new errors
  • Memory-search Lambda generates new 1024-dim embedding for INC-1947 symptoms
  • INSERT into CockroachDB incident_embeddings: new vector stored
  • Incident marked "resolved" in CockroachDB
  • Agent run completed with full audit trail
  • Memory now contains 2 experiences — agent is smarter
  • Future incidents matching either INC-1832 or INC-1947 will be recalled

📐 Architecture

OpsMemory AI Architecture Open full size ↗

🏆 Judging Criteria

1. Agentic Memory Design

CockroachDB stores ALL memory: incidents, 1024-dim vectors (HNSW indexed), agent state, approvals, audit log. Vector search finds 86% similarity between differently-worded incidents.

Steps 3 & 7 demonstrate this

2. Technical Implementation

HNSW vector index, MCP client interface, 6 Lambda functions, Nova Pro structured reasoning, esbuild bundling, SAM deployment.

Steps 2-6 demonstrate this

3. Real-World Impact

SRE incident response — reduces MTTR by providing instant historical context. Agent actually changes AWS infrastructure.

Steps 5-6 demonstrate this

4. Production Readiness

Human approval gate, IAM least-privilege, Secrets Manager, structured logs, fail-safe principle, audit trail.

Step 5 demonstrates this

5. Creativity & Originality

"Experiential memory" — not RAG, not document retrieval. The agent remembers what it DID and what HAPPENED.

Steps 3 & 7 demonstrate this

📋 Select Scenario

Click a scenario below to load it. Each uses different wording to test semantic memory.

🔥 SQS Visibility Timeout

New: "messages reprocessing, queue growing"
Should match: INC-1832 (Lambda timeouts, backlog)

📈 Lambda Memory Exhaustion

New: "function crashing, exit code 137"
Should match: INC-1756 (OOM errors, retries)

🔄 DynamoDB Throttling

New: "API errors, database requests rejected"
Should match: INC-1698 (capacity exceeded)

🌐 ALB 502 Errors

New: "website down, health checks failing"
Should match: INC-1601 (security group removed)

✓ All backed by real 1024-dim embeddings in CockroachDB

🔑 Key Differentiator

Different words. Same experience.

INC-1832: "Lambda timeouts, SQS backlog"
INC-1947: "Messages reprocessing, queue growing"

Zero keyword overlap. But CockroachDB's vector index found 86.4% semantic similarity.

This is what makes experiential memory different from document search or runbook lookup.