🧠Advanced Patterns
RLM, Ralph-RLM, GasTown, and multi-agent orchestration patterns
✨ REMBR for Multi-Agent Systems
RLM (Recursive Language Model)
Hierarchical task decomposition with persistent context
The RLM pattern breaks complex tasks into hierarchical subtasks, with each level storing findings in REMBR. Sub-agents search context, validate results, and synthesize insights.
How It Works
- •L0 agent receives complex task
- •Decomposes into L1 subtasks
- •Each L1 stores findings in REMBR
- •L0 searches context, synthesizes
REMBR Tools Used
- •
create_context- Task workspaces - •
store_memory- Save findings - •
search_context- Retrieve results - •
create_snapshot- State handoff
Ralph-RLM (Acceptance-Driven RLM)
Loops until explicit criteria are met with stuck detection
The Ralph-RLM pattern combines RLM decomposition with validation loops. Tasks iterate until all acceptance criteria pass, with automatic stuck detection and plan regeneration.
How It Works
- •Store acceptance criteria in REMBR
- •Execute RLM decomposition
- •Validate against criteria
- •Loop until all criteria met or stuck
Key Features
- •Explicit acceptance criteria
- •Automatic stuck detection
- •Plan regeneration on stall
- •Progress tracking in REMBR
GasTown (Multi-Agent Collaboration)
Parallel agents with shared context and conflict resolution
The GasTown pattern enables parallel agents to collaborate on tasks using REMBR as a shared knowledge base with conflict detection and resolution.
How It Works
- •Multiple agents share REMBR context
- •Each agent stores findings
- •Contradiction detection via REMBR
- •Coordinator resolves conflicts
REMBR Tools Used
- •
detect_contradictions - •
get_memory_graph - •
infer_causality - •
compare_snapshots
🔗 Combining Patterns
Mix and match for complex workflows