Multi-Account Autonomous Incident Investigation, Root Cause Analysis (RCA), and Vector-Based Memory System
Multi-Account Trust & Security
• Hub-and-Spoke OIDC Trust: Central Vegapay-AI account assumes roles in spoke accounts using temporary read-only credentials (1hr TTL).
• Zero Secrets Stored: No banking passwords or long-lived API keys are kept in the database.
• Read-Only Scoping: IAM policies limit the agent to logs (CloudWatch), metrics (Prometheus), and container states (K8s) — no write permissions.
Three-Tier Memory System
• Tier 1: Hot Memory (Postgres/pgvector): HNSW indexing on 1024-dimensional embeddings allows instant semantic search of past incidents.
• Tier 2: Visible Memory: Slack threads hold real-time investigation transcripts. FastAPI Dashboard allows full-text search and MTTR analytics.
• Tier 3: Cold Memory (S3): Complete raw tool execution traces are archived in S3 with SSE-KMS encryption for a 2-year regulatory retention period.
Feedback Loop & Learning
• Socket Mode Listener: Captures engineer reactions (👍/👎) and direct text corrections in Slack threads.
• Auto-Decay & Hygiene: Memory Worker updates confidence scores based on feedback. Low-confidence memories decay over time to prevent LLM context pollution.
• Auto-Evolving Runbooks: Successfully resolved incidents promote runbook chunks, making subsequent investigations faster and more precise.
Nika v0.2.2 • Designed for Vegapay Production Environments