Multi-Agent Ralph Loop

Orchestration system for Claude Code with memory-driven planning, multi-agent coordination, Agent Teams integration, and automatic learning.

Overview

Ralph extends Claude Code with intelligent orchestration capabilities. It classifies tasks, routes to appropriate models, coordinates multiple agents, and maintains persistent memory across sessions.

Key capabilities:

• Agent Teams Integration - Multiple Claude Code instances working in parallel
• Task classification and routing
• Parallel memory search across multiple backends
• Multi-agent coordination with swarm mode
• Automatic learning from GitHub repositories
• Quality validation with adversarial review
• Session state management with checkpoints

Version

Current: v2.93.0

What's New in v2.93.0

• Memory System Optimization — Smart memory search re-enabled + cleanup policies implemented
  • Re-enabled smart-memory-search.sh hook (700+ lines of parallel memory search)
  • Implemented episodic memory cleanup: 30-day TTL (previously: NO cleanup)
  • Removed redundant storage: memvid.json (175KB) and semantic.json (62KB)
  • Streamlined to primary memory storage: claude-mem + episodic + handoffs + ledgers
  • Removed memvid integration from ralph CLI (tool definition + validation)

What's New in v2.92.0

• Removed Obsolete Dependencies — Cleaned up llm-tldr and claude-sneakpeek references
  • Removed llm-tldr integration (hook, tests, CLI command)
  • Removed claude-sneakpeek historical references
  • Updated context tree to use ast-grep backend instead of llm-tldr
  • Simplified architecture without deprecated tooling

What's New in v2.91.1

• Optional Code Quality Improvements — Enhanced defensive coding practices
  • API key validation: Format checking (minimum length, placeholder detection)
  • JSON error handling: Comprehensive try/except blocks in 2 files
  • Subprocess safety validation: All patterns confirmed safe
  • Console.log analysis: All uses verified as legitimate
  • Security grade: Maintained at A- with enhanced robustness
  • SQL injection warnings: Added to 7 test files for security awareness

What's New in v2.90.2

• MCP Servers Integration — Structural code search and Zai vision capabilities
  • ast-grep MCP: AST-based code search (13 MCP servers total)
    • Tool selection guide: ast-grep vs Grep vs Glob based on benchmark results
    • 4 cases where only ast-grep works: containment, blocks, negation, scope-aware
    • Complete rule reference (atomic, relational, composite, metavariables)
  • Zai MCP Servers: 4 servers with 13 tools
    • zai-mcp-server: 8 vision tools (ui_to_artifact, extract_text, diagnose_error, etc.)
    • web-search-prime: Web search with rich results
    • web-reader: Full webpage content extraction
    • zread: GitHub repository documentation (search_doc, read_file, get_repo_structure)

Previous Releases

• v2.89.2 - Hooks alignment with official...