Neuronium

Overview#

Neuronium is a Python-based AI super agent framework developed by Dataism Lab. Its core features are deterministic execution and auditability, designed to address the unpredictability, debugging difficulties, and uncontrollable execution processes that traditional LLM Agents face when handling long-horizon, high-entropy complex tasks.

Core Features#

Planning System#

• HTN (Hierarchical Task Network): Hierarchical task network decomposition, breaking complex goals into executable subtasks
• Action Graph (DAG): Directed Acyclic Graph-based action planning, supporting parallelization and precise recovery

Execution Engine#

• Deterministic Execution: Core invariant — same input/state/tools produce identical execution traces
• Artifact Lineage Immutability: Content-Addressable Storage (CAS) with SHA-256 hashing, preventing tampering
• Multi-type Nodes: Model Node (LLM), MCP Tool Node, Code Node, Decision Node, Aggregate Node

Verification & Control#

• Verification Critic: Embedded quality assessment, validating at decision points rather than post-hoc checks
• Typed Contracts: JSON Schema validation for inter-node communication
• Intent Lifecycle: Commit → Execute → Control → Adapt four-state machine management

Memory System#

• Hybrid Memory: GraphRAG + agent retrieval dual mechanism as first-class architectural elements

Observability#

• Audit & Replay: Complete execution tracing, supporting deterministic replay from any checkpoint

Installation & Configuration#

# Basic installation
pip install -e .

# With Docker sandbox support
pip install -e ".[docker]"

Configuration file neuronium.toml:

[project]
name = "neuronium"
data_dir = ".neuronium"

[determinism]
canonical_json = "neuronium-v1"
default_random_seed = 0
llm_temperature = 0.0

[storage]
blob_backend = "fs_cas"
index_backend = "sqlite"

[llm]
provider = "openai"
model = "gpt-4.1-mini"

Usage#

CLI#

neuronium-agent run --objective "Write a fibonacci function in Python" --trace-export ./trace.jsonl
neuronium-agent run -o "Write fibonacci" --mode interactive

Python API#

from neuronium_agent.api import create_runner
from neuronium_agent.types import RunRequest

runner = create_runner()
handle = runner.start(RunRequest(objective="Write fibonacci"))
status = runner.get_status(handle)
runner.export_trace(handle, "jsonl", "trace.jsonl")

Use Cases#

• Long-horizon task solving: Complex tasks requiring multi-step planning and execution
• High-entropy environments: Real-world environments with high uncertainty requiring dynamic adaptation
• Code generation & execution: Supporting deterministic computation and sandboxed execution
• Quality-sensitive applications: AI applications requiring strict verification and quality assurance
• Compliance & audit requirements: Enterprise applications requiring complete decision tracing

Project Info#

• Organization: Dataism Lab
• Primary Language: Python (100%)
• Current Version: v0.1
• License: Apache-2.0