mirror of
https://github.com/awslabs/amazon-bedrock-agentcore-samples.git
synced 2025-09-08 20:50:46 +00:00
f496048c13
* feat: integrate long-term memory system into SRE agent - Add AgentCore Memory integration with three memory strategies: * User preferences (escalation, notification, workflow preferences) * Infrastructure knowledge (dependencies, patterns, baselines) * Investigation summaries (timeline, actions, findings) - Implement memory tools for save/retrieve operations - Add automatic memory capture through hooks and pattern recognition - Extend agent state to support memory context - Integrate memory-aware planning in supervisor agent - Add comprehensive test coverage for memory functionality - Create detailed documentation with usage examples This transforms the SRE agent from stateless to learning assistant that becomes more valuable over time by remembering user preferences, infrastructure patterns, and investigation outcomes. Addresses issue #164 * feat: environment variable config, agent routing fixes, and project organization - Move USER_ID/SESSION_ID from metadata parsing to environment variables - Add .memory_id to .gitignore for local memory state - Update .gitignore to use .scratchpad/ folder instead of .scratchpad.md - Fix agent routing issues with supervisor prompt and graph node naming - Add conversation memory tracking for all agents and supervisor - Improve agent metadata system with centralized constants - Add comprehensive logging and debugging for agent tool access - Update deployment script to pass user_id/session_id in payload - Create .scratchpad/ folder structure for better project organization * feat: enhance SRE agent with automatic report archiving and error fixes - Add automatic archiving system for reports by date - Include user_id in report filenames for better organization - Fix Pydantic validation error with string-to-list conversion for investigation steps - Add content length truncation for memory storage to prevent validation errors - Remove status line from report output for cleaner formatting - Implement date-based folder organization (YYYY-MM-DD format) - Add memory content length limits configuration in constants Key improvements: - Reports now auto-archive old files when saving new ones - User-specific filenames: query_user_id_UserName_YYYYMMDD_HHMMSS.md - Robust error handling for memory content length limits - Backward compatibility with existing filename formats * feat: fix memory retrieval system for cross-session searches and user personalization Key fixes and improvements: - Fix case preservation in actor_id sanitization (Carol remains Carol, not carol) - Enable cross-session memory searches for infrastructure and investigation memories - Add XML parsing support for investigation summaries stored in XML format - Enhance user preference integration throughout the system - Add comprehensive debug logging for memory retrieval processes - Update prompts to support user-specific communication styles and preferences Memory system now properly: - Preserves user case in memory namespaces (/sre/users/Carol vs /sre/users/carol) - Searches across all sessions for planning context vs session-specific for current state - Parses both JSON and XML formatted investigation memories - Adapts investigation approach based on user preferences and historical patterns - Provides context-aware planning using infrastructure knowledge and past investigations * feat: enhance SRE agent with user-specific memory isolation and anti-hallucination measures Memory System Improvements: - Fix memory isolation to retrieve only user-specific memories (Alice doesn't see Carol's data) - Implement proper namespace handling for cross-session vs session-specific searches - Add detailed logging for memory retrieval debugging and verification - Remove verbose success logs, keep only error logs for cleaner output Anti-Hallucination Enhancements: - Add tool output validation requirements to agent prompts - Implement timestamp fabrication prevention (use 2024-* format from backend) - Require tool attribution for all metrics and findings in reports - Add backend data alignment patterns for consistent data references - Update supervisor aggregation prompts to flag unverified claims Code Organization: - Extract hardcoded prompts from supervisor.py to external prompt files - Add missing session_id parameters to SaveInfrastructureTool and SaveInvestigationTool - Improve memory client namespace documentation and cross-session search logic - Reduce debug logging noise while maintaining error tracking Verification Complete: - Memory isolation working correctly (only user-specific data retrieval) - Cross-session memory usage properly configured for planning and investigations - Memory integration confirmed in report generation pipeline - Anti-hallucination measures prevent fabricated metrics and timestamps * feat: organize utility scripts in dedicated scripts folder Script Organization: - Move manage_memories.py to scripts/ folder with updated import paths - Move configure_gateway.sh to scripts/ folder with corrected PROJECT_ROOT path - Copy user_config.yaml to scripts/ folder for self-contained script usage Path Fixes: - Update manage_memories.py to import sre_agent module from correct relative path - Fix .memory_id file path resolution for new script location - Update configure_gateway.sh PROJECT_ROOT to point to correct parent directory - Add fallback logic to find user_config.yaml in scripts/ or project root Script Improvements: - Update help text and examples to use 'uv run python scripts/' syntax - Make manage_memories.py executable with proper permissions - Maintain backward compatibility for custom config file paths - Self-contained scripts folder with all required dependencies Verification: - All scripts work correctly from new location - Memory management functions operate properly - Gateway configuration handles paths correctly - User preferences loading works from scripts directory * docs: update SSL certificate paths to use /opt/ssl standard location - Update README.md to reference /opt/ssl for SSL certificate paths - Update docs/demo-environment.md to use /opt/ssl paths - Clean up scripts/configure_gateway.sh SSL fallback paths - Remove duplicate and outdated SSL path references - Establish /opt/ssl as the standard SSL certificate location This ensures consistent SSL certificate management across all documentation and scripts, supporting the established /opt/ssl directory with proper ubuntu:ubuntu ownership. * feat: enhance memory system with infrastructure parsing fix and user personalization analysis Infrastructure Memory Parsing Improvements: - Fix infrastructure memory parsing to handle both JSON and plain text formats - Convert plain text memories to structured InfrastructureKnowledge objects - Change warning logs to debug level for normal text-to-structure conversion - Ensure all infrastructure memories are now retrievable and usable User Personalization Documentation: - Add comprehensive memory system analysis comparing Alice vs Carol reports - Create docs/examples/ folder with real investigation reports demonstrating personalization - Document side-by-side communication differences based on user preferences - Show how same technical incident produces different reports for different user roles Example Reports Added: - Alice's technical detailed investigation report (technical role preferences) - Carol's business-focused executive summary report (executive role preferences) - Memory system analysis with extensive side-by-side comparisons This demonstrates the memory system's ability to: - Maintain technical accuracy while adapting presentation style - Apply user-specific escalation procedures and communication channels - Build institutional knowledge about recurring infrastructure patterns - Personalize identical technical incidents for different organizational roles * feat: enhance memory system with automatic pattern extraction and improved logging ## Memory System Enhancements - **Individual agent memory integration**: Every agent response now triggers automatic memory pattern extraction through on_agent_response() hooks - **Enhanced conversation logging**: Added detailed message breakdown showing USER/ASSISTANT/TOOL message counts and tool names called - **Fixed infrastructure extraction**: Resolved hardcoded agent name issues by using SREConstants for agent identification - **Comprehensive memory persistence**: All agent responses and tool executions stored as conversation memory with proper session tracking ## Tool Architecture Clarification - **Centralized memory access**: Confirmed only supervisor agent has direct access to memory tools (retrieve_memory, save_*) - **Individual agent focus**: Individual agents have NO memory tools, only domain-specific tools (5 tools each for metrics, logs, k8s, runbooks) - **Automatic pattern recognition**: Memory capture happens automatically through hooks, not manual tool calls by individual agents ## Documentation Updates - **Updated memory-system.md**: Comprehensive design documentation reflecting current implementation - **Added example analyses**: Created flight-booking-analysis.md and api-response-time-analysis.md in docs/examples/ - **Enhanced README.md**: Added memory system overview and personalized investigation examples - **Updated .gitignore**: Now ignores entire reports/ folder instead of just .md files ## Implementation Improvements - **Event ID tracking**: All memory operations generate and log event IDs for verification - **Pattern extraction confirmation**: Logs confirm pattern extraction working for all agent types - **Memory save verification**: Comprehensive logging shows successful saves across all memory types - **Script enhancements**: manage_memories.py now handles duplicate removal and improved user management * docs: enhance memory system documentation with planning agent memory usage examples - Add real agent.log snippets showing planning agent retrieving and using memory context - Document XML-structured prompts for improved Claude model interaction - Explain JSON response format enforcement and infrastructure knowledge extraction - Add comprehensive logging and monitoring details - Document actor ID design for proper memory namespace isolation - Fix ASCII flow diagram alignment for better readability - Remove temporal framing and present features as current design facts * docs: add AWS documentation links and clean up memory system documentation - Add hyperlink to Amazon Bedrock AgentCore Memory main documentation - Link to Memory Getting Started Guide for the three memory strategies - Remove Legacy Pattern Recognition section from documentation (code remains) - Remove Error Handling and Fallbacks section to focus on core functionality - Keep implementation details in code while streamlining public documentation * docs: reorganize memory-system.md to eliminate redundancies - Merged Memory Tool Architecture and Planning sections into unified section - Consolidated all namespace/actor_id explanations in architecture section - Combined pattern recognition and memory capture content - Created dedicated Agent Memory Integration section with examples - Removed ~15-20% redundant content while improving clarity - Improved document structure for better navigation * style: apply ruff formatting and fix code style issues - Applied ruff auto-formatting to all Python files - Fixed 383 style issues automatically - Remaining issues require manual intervention: - 29 ruff errors (bare except, unused variables, etc.) - 61 mypy type errors (missing annotations, implicit Optional) - Verified memory system functionality matches documentation - Confirmed user personalization working correctly in reports * docs: make benefits section more succinct in memory-system.md - Consolidated 12 bullet points into 5 focused benefits - Removed redundant three-category structure (Users/Teams/Operations) - Maintained all key value propositions while improving readability - Reduced section length by ~60% while preserving essential information * feat: add comprehensive cleanup script with memory deletion - Added cleanup.sh script to delete all AWS resources (gateway, runtime, memory) - Integrated memory deletion using bedrock_agentcore MemoryClient - Added proper error handling and graceful fallbacks - Updated execution order: servers → gateway → memory → runtime → local files - Added memory deletion to README.md cleanup instructions - Includes confirmation prompts and --force option for automation * fix: preserve .env, .venv, and reports in cleanup script - Modified cleanup script to only remove AWS-generated configuration files - Preserved .env files for development continuity - Preserved .venv directories to avoid reinstalling dependencies - Preserved reports/ directory containing investigation history - Files removed: gateway URIs, tokens, agent ARNs, memory IDs only - Updated documentation to clarify preserved vs removed files * fix: use correct bedrock-agentcore-control client for gateway operations - Changed boto3 client from 'bedrock-agentcore' to 'bedrock-agentcore-control' - Fixes 'list_gateways' method not found error during gateway deletion - Both gateway and runtime deletion now use the correct control plane client * docs: add memory system initialization timing guidance - Added note that memory system takes 10-12 minutes to be ready - Added steps to check memory status with list command after 10 minutes - Added instruction to run update command again once memory is ready - Provides clear workflow for memory system setup and prevents user confusion * docs: comprehensive documentation update and cleanup - Remove unused root .env and .env.example files (not referenced by any code) - Update configuration.md with comprehensive config file documentation - Add configuration overview table with setup instructions and auto-generation info - Consolidate specialized-agents.md content into system-components.md - Update system-components.md with complete AgentCore architecture - Add detailed sections for AgentCore Runtime, Gateway, and Memory primitives - Remove cli-reference.md (excessive documentation for limited use) - Update README.md to reference configuration guide in setup section - Clean up documentation links and organization The documentation now provides a clear, consolidated view of the system architecture and configuration with proper cross-references and setup guidance. * feat: improve runtime deployment and invocation robustness - Increase deletion wait time to 150s for agent runtime cleanup - Add retry logic with exponential backoff for MCP rate limiting (429 errors) - Add session_id and user_id to agent state for memory retrieval - Filter out /ping endpoint logs to reduce noise - Increase boto3 read timeout to 5 minutes for long-running operations - Add clear error messages for agent name conflicts - Update README to clarify virtual environment requirement for scripts - Fix session ID generation to meet 33+ character requirement These changes improve reliability when deploying and invoking agents, especially under heavy load or with complex queries that take time. * chore: remove accidentally committed reports folder Removed 130+ markdown report files from the reports/ directory that were accidentally committed. The .gitignore already includes reports/ to prevent future commits of these generated files.
12 KiB
12 KiB
System Components
The SRE Agent system is built on three core Amazon Bedrock AgentCore components that work together to provide a scalable, secure, and intelligent infrastructure management solution.
Architecture Overview
┌──────────────────────────────────┐ ┌─────────────────────────────────────┐
│ AgentCore Memory │ │ AgentCore Runtime │
│ • User Preferences │ │ ┌──────────────────────────────────┐
│ • Infrastructure Knowledge │ │ │ Multi-Agent System (LangGraph) │
│ • Investigation Summaries │ │ │ ┌─────────────────┐ │
└─────────────┬────────────────────┘ │ │ │ Supervisor │ │
│ │ │ │ Agent │ │
└──────────────────────►│ │ └────────┬────────┘ │
│ │ │ │
│ │ ┌────────┴────┬─────┬─────┬───┐ │
│ │ ▼ ▼ ▼ ▼ ▼ │
│ │┌──────┐ ┌──────┐ ┌─────┐ ┌───┐│ │
│ ││ K8s │ │ Logs │ │Metr-│ │Run││ │
│ ││Agent │ │Agent │ │ics │ │bks││ │
│ │└───┬──┘ └──┬───┘ └──┬──┘ └─┬─┘│ │
│ └────┼────────┼────────┼──────┼──┘ │
└───────┼────────┼────────┼──────┼────┘
│ │ │ │
▼ ▼ ▼ ▼
┌────────────────────────────────────────────────────────────────────────┐
│ AgentCore Gateway │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────────────────────┐ │
│ │ MCP Tools │ │ Auth │ │ API Translation │ │
│ └──────┬──────┘ └──────┬──────┘ └─────────┬───────────────────┘ │
└─────────┼────────────────┼───────────────────┼─────────────────────────┘
│ │ │
▼ ▼ ▼
Backend APIs Identity Provider OpenAPI Specs
Multi-Agent System
Built on LangGraph, the multi-agent system orchestrates specialized agents for complex investigations.
Supervisor Agent
The central coordinator that:
- Analyzes incoming queries and determines investigation strategy
- Routes tasks to appropriate specialist agents based on expertise
- Coordinates multi-step investigations across multiple domains
- Manages all memory interactions - only the supervisor has direct access to AgentCore Memory
- Personalizes investigations based on retrieved user preferences and past investigations
Specialist Agents
Each agent focuses on a specific domain with dedicated tools:
Kubernetes Infrastructure Agent
Handles container orchestration and cluster operations. This agent investigates issues across pods, deployments, services, and nodes by examining cluster state, analyzing pod health, resource utilization, and recent events.
Capabilities:
- Check pod status across namespaces
- Examine deployment configurations and rollout history
- Investigate cluster events for anomalies
- Analyze resource usage patterns
- Monitor node health and capacity
Application Logs Agent
Processes log data to find relevant information. This agent understands log patterns, identifies anomalies, and correlates events across multiple services.
Capabilities:
- Full-text search with regex support
- Error log aggregation and categorization
- Pattern detection for recurring issues
- Time-based correlation of events
- Statistical analysis of log volumes
Performance Metrics Agent
Monitors system metrics and identifies performance issues. This agent understands relationships between different metrics and provides both real-time analysis and historical trending.
Capabilities:
- Application performance metrics (response times, throughput)
- Error rate analysis with thresholding
- Resource utilization metrics (CPU, memory, disk)
- Availability and uptime monitoring
- Trend analysis for capacity planning
Operational Runbooks Agent
Provides access to documented procedures, troubleshooting guides, and best practices. This agent helps standardize incident response by retrieving relevant procedures based on the current situation.
Capabilities:
- Incident-specific playbooks for common scenarios
- Detailed troubleshooting guides with step-by-step instructions
- Escalation procedures with contact information
- Common resolution patterns for known issues
- Best practices for system operations
Search Agent
Provides cross-domain information retrieval capabilities:
Capabilities:
- Unified search across all infrastructure domains
- Context-aware result ranking and filtering
- Cross-reference information between different agent domains
Agent Collaboration
The supervisor coordinates complex investigations by:
- Breaking down queries into specialized tasks
- Routing tasks to appropriate agents in parallel or sequence
- Aggregating results from multiple agents
- Applying memory-based personalization to findings
- Generating unified, context-aware reports
Amazon Bedrock AgentCore
The system leverages three fundamental AgentCore primitives that provide enterprise-grade AI infrastructure:
1. AgentCore Runtime
A serverless execution environment designed specifically for AI agents:
- Managed Infrastructure: Fully managed compute with automatic scaling from zero to thousands of concurrent sessions
- Container-based Deployment: Supports ARM64 Docker containers with built-in security and isolation
- Enterprise Integration: Native AWS IAM support with session-level security boundaries
- Multi-model Support: Compatible with Amazon Bedrock models and external LLM providers
- Production Features: Built-in monitoring, logging, debugging, and observability
2. AgentCore Gateway
A secure API bridge that enables agents to interact with backend systems:
- Protocol Translation: Converts REST/GraphQL/gRPC APIs into standardized MCP (Model Context Protocol) tools
- Universal Tool Interface: Provides a consistent interface for any agent framework to discover and use tools
- Enterprise Security: JWT-based authentication with support for multiple identity providers (Cognito, Auth0, Okta)
- API Management: Health monitoring, automatic retries, rate limiting, and error handling
- Schema-driven: Uses OpenAPI specifications to automatically generate tool definitions
3. AgentCore Memory
A persistent knowledge system that enables agents to learn and personalize over time:
- Event-based Storage: Immutable event log that accumulates knowledge without data loss
- Namespace Isolation: Automatic routing based on actor IDs for user and context separation
- Flexible Retention: Configurable retention policies for different memory types (30-90 days)
- Pattern Extraction: Automatic extraction of structured data from unstructured agent responses
- Cross-session Learning: Enables agents to build on past investigations and user interactions
Integration Architecture
Data Flow
- User Query → Supervisor Agent analyzes and retrieves relevant memories
- Investigation Planning → Supervisor routes to specialist agents
- Tool Execution → Agents access backend APIs through Gateway
- Response Processing → Memory system extracts and stores patterns
- Report Generation → Personalized report based on user preferences
Security Model
- IAM Integration: Full AWS IAM support with
BedrockAgentCoreFullAccesspolicy - JWT Authentication: Bearer tokens for gateway communication
- SSL/TLS Required: All endpoints must use HTTPS
- Namespace Isolation: Memory events isolated by actor ID
Tool Domains
The gateway provides 20 specialized tools across 4 domains:
Kubernetes Operations (5 tools)
get_pod_status: Monitor pod health and stateget_deployment_status: Check deployment rollout statusget_cluster_events: Retrieve recent cluster eventsget_resource_usage: Analyze CPU/memory utilizationget_node_status: Monitor node health and capacity
Log Analysis (5 tools)
search_logs: Full-text search across log streamsget_error_logs: Extract error and exception logsanalyze_log_patterns: Detect recurring patternsget_recent_logs: Retrieve latest log entriescount_log_events: Aggregate log event statistics
Metrics Collection (5 tools)
get_performance_metrics: Application performance dataget_error_rates: Error rate trends and spikesget_resource_metrics: Infrastructure resource usageget_availability_metrics: Service uptime and SLAsanalyze_trends: Historical trend analysis
Runbook Management (5 tools)
search_runbooks: Find relevant proceduresget_incident_playbook: Incident response guidesget_troubleshooting_guide: Step-by-step debuggingget_escalation_procedures: Contact and escalation pathsget_common_resolutions: Known issue solutions
Demo Environment
For evaluation and testing, the system includes a demo environment with:
- Mock API Servers: Simulated Kubernetes, logs, metrics, and runbooks APIs
- Realistic Data: Representative infrastructure scenarios and failure patterns
- Safe Testing: Isolated environment prevents production impact
- Full Feature Support: All agent capabilities available in demo mode
Development to Production
The architecture supports seamless progression from development to production:
Local Development → Container Testing → Production Deployment
(CLI) (Docker) (AgentCore Runtime)
↓ ↓ ↓
Gateway Only Gateway + Runtime Full Stack with Memory
This unified approach ensures consistent behavior across all deployment stages while providing the scalability and security required for enterprise production use.