rlm-toolkit 1.0.0

Recursive Language Models Toolkit for processing unlimited context

RLM-Toolkit

Recursive Language Models Toolkit — библиотека для обработки контекстов произвольной длины (10M+ токенов) с использованием рекурсивных вызовов LLM.

🚀 Quick Start

pip install rlm-toolkit

from rlm_toolkit import RLM

# Simple usage with Ollama
rlm = RLM.from_ollama("llama3")
result = rlm.run(
    context=open("large_document.txt").read(),
    query="What are the key findings?"
)
print(result.answer)

✨ Features

Feature	Description
Infinite Context	Process 10M+ tokens with O(1) memory
InfiniRetri	🆕 Attention-based retrieval, 100% accuracy on 1M+ tokens
H-MEM	🆕 4-level hierarchical memory with LLM consolidation
Self-Evolving	🆕 LLMs that improve through usage (R-Zero pattern)
Multi-Agent	🆕 Decentralized P2P agents with Trust Zones
DSPy Optimization	🆕 Automatic prompt optimization
Secure REPL	CIRCLE-compliant sandboxed code execution
Multi-Provider	75 LLM providers (OpenAI, Anthropic, Google, Ollama, vLLM...)
Document Loaders	135+ sources (Slack, Jira, GitHub, S3, databases...)
Vector Stores	20+ stores (Pinecone, Chroma, Weaviate, pgvector...)
Embeddings	15+ providers (OpenAI, BGE, E5, Jina, Cohere...)
Cost Control	Budget limits, cost tracking
Observability	OpenTelemetry, Langfuse, LangSmith, W&B (12 backends)
Memory Systems	Buffer, Episodic, Hierarchical (H-MEM)

📋 Full Integration Catalog — 287+ production-ready integrations

🔥 InfiniRetri (NEW)

Attention-based infinite context retrieval — 100% accuracy on Needle-In-a-Haystack up to 1M+ tokens.

from rlm_toolkit.retrieval import InfiniRetriever

# Retrieve from 1M+ token documents
retriever = InfiniRetriever("Qwen/Qwen2.5-0.5B-Instruct")
answer = retriever.retrieve(
    context=million_token_doc,
    question="What is the secret code?"
)

# Or use automatic routing in RLM
from rlm_toolkit import RLM, RLMConfig

config = RLMConfig(
    use_infiniretri=True,
    infiniretri_threshold=100_000,  # Auto-switch at 100K tokens
)
rlm = RLM.from_ollama("llama3", config=config)
result = rlm.run(huge_document, "Summarize")  # Automatically uses InfiniRetri

Based on arXiv:2502.12962 — requires pip install infini-retri

🧠 Hierarchical Memory (H-MEM) (NEW)

Multi-level persistent memory with semantic consolidation — memories that learn and evolve.

from rlm_toolkit.memory import HierarchicalMemory, SecureHierarchicalMemory

# Basic H-MEM
hmem = HierarchicalMemory()
hmem.add_episode("User asked about weather")
hmem.add_episode("AI responded with forecast")
hmem.consolidate()  # Auto-creates traces, categories, domains

results = hmem.retrieve("weather")

# Secure H-MEM with encryption and trust zones
smem = SecureHierarchicalMemory(
    agent_id="agent-001",
    trust_zone="zone-secure"
)
smem.add_episode("Confidential data")
smem.grant_access("agent-002", "zone-secure")

4-Level Architecture:

Level 3: DOMAIN    → High-level knowledge
Level 2: CATEGORY  → Semantic categories  
Level 1: TRACE     → Consolidated memories
Level 0: EPISODE   → Raw interactions

Based on arXiv H-MEM paper (July 2025)

🧬 Self-Evolving LLMs (NEW)

LLMs that improve reasoning through usage — no human supervision required.

from rlm_toolkit.evolve import SelfEvolvingRLM, EvolutionStrategy
from rlm_toolkit.providers import OllamaProvider

# Create self-evolving RLM
evolve = SelfEvolvingRLM(
    provider=OllamaProvider("llama3"),
    strategy=EvolutionStrategy.CHALLENGER_SOLVER
)

# Solve with self-refinement
answer = evolve.solve("What is 25 * 17?")
print(f"Answer: {answer.answer}, Confidence: {answer.confidence}")

# Run training loop (generates challenges → solves → improves)
metrics = evolve.training_loop(iterations=10, domain="math")
print(f"Success rate: {metrics.success_rate}")

Strategies:

• SELF_REFINE — Iterative self-improvement
• CHALLENGER_SOLVER — R-Zero co-evolutionary loop
• EXPERIENCE_REPLAY — Learn from past solutions

Based on R-Zero (arXiv:2508.05004)

🤖 Multi-Agent Framework (NEW)

Decentralized P2P agents inspired by Meta Matrix — no central orchestrator bottleneck.

from rlm_toolkit.agents import MultiAgentRuntime, SecureAgent, EvolvingAgent

# Create runtime
runtime = MultiAgentRuntime()

# Register agents with Trust Zones
runtime.register(SecureAgent("analyst", "Data Analyst", trust_zone="internal"))
runtime.register(EvolvingAgent("solver", "Problem Solver", llm_provider=provider))

# Run message through agents
from rlm_toolkit.agents import AgentMessage
message = AgentMessage(content="Analyze this data", routing=["analyst", "solver"])
result = runtime.run(message)

Agent Types:

• SecureAgent — H-MEM Trust Zones integration
• EvolvingAgent — Self-improving via R-Zero
• SecureEvolvingAgent — Both combined

Based on Meta Matrix (arXiv 2025)

🎯 DSPy-Style Optimization (NEW)

Automatic prompt optimization — define what, not how.

from rlm_toolkit.optimize import Signature, Predict, ChainOfThought, BootstrapFewShot

# Define signature
sig = Signature(
    inputs=["question", "context"],
    outputs=["answer"],
    instructions="Answer based on context"
)

# Use with Chain of Thought
cot = ChainOfThought(sig, provider)
result = cot(question="What is X?", context="X is 42")

# Auto-optimize with few-shot selection
optimizer = BootstrapFewShot(metric=lambda p, g: p["answer"] == g["answer"])
optimized = optimizer.compile(Predict(sig, provider), trainset=examples)

Modules: Predict, ChainOfThought, SelfRefine
Optimizers: BootstrapFewShot, PromptOptimizer

Inspired by Stanford DSPy

📦 Installation

# Basic
pip install rlm-toolkit

# With all providers
pip install rlm-toolkit[all]

# Development
pip install -e ".[dev]"

🔧 Usage

Basic

from rlm_toolkit import RLM, RLMConfig

# With configuration
config = RLMConfig(
    max_iterations=50,
    max_cost=5.0,  # USD
)

rlm = RLM.from_openai("gpt-4o", config=config)
result = rlm.run(context, query)

With Memory

from rlm_toolkit.memory import EpisodicMemory

memory = EpisodicMemory(max_entries=1000)
rlm = RLM.from_ollama("llama3", memory=memory)

# Memory persists across runs
result1 = rlm.run(doc1, "Summarize this")
result2 = rlm.run(doc2, "Compare with previous")

With Observability

from rlm_toolkit.observability import Tracer, CostTracker

tracer = Tracer(service_name="my-app")
cost_tracker = CostTracker(budget=10.0)

rlm = RLM.from_openai("gpt-4o", tracer=tracer, cost_tracker=cost_tracker)

🔒 Security

RLM-Toolkit implements CIRCLE-compliant security:

• AST Analysis — Block dangerous imports before execution
• Sandboxed REPL — Isolated code execution with timeouts
• Virtual Filesystem — Quota-enforced file operations
• Attack Detection — Obfuscation and indirect attack patterns

from rlm_toolkit import RLMConfig, SecurityConfig

config = RLMConfig(
    security=SecurityConfig(
        sandbox=True,
        max_execution_time=30.0,
        max_memory_mb=512,
    )
)

📊 Benchmarks

Based on RLM paper methodology:

Benchmark	Score
OOLONG-Pairs	TBD
CIRCLE Security	~95%

🛠️ CLI

# Run a query
rlm run --model ollama:llama3 --context file.txt --query "Summarize"

# Interactive REPL
rlm repl --model openai:gpt-4o

# Cost tracking
rlm trace --session latest

📚 Documentation

• Getting Started
• API Reference
• Security Guide
• Examples

🤝 Contributing

# Clone repo
git clone https://github.com/sentinel-community/rlm-toolkit.git
cd rlm-toolkit

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Lint
ruff check rlm_toolkit/

📄 License

Apache 2.0 — see LICENSE

🙏 Acknowledgments

• RLM Paper by Zhang, Kraska, Khattab
• CIRCLE Benchmark
• SENTINEL Community