The brain behind
AI workflow orchestration
Turn chaotic multi-step AI tasks into structured, observable, retryable workflows — with cognitive memory and discipline-aware routing.
Built for Hermes Agent · Works with OpenClaw
Multi-step AI workflows are chaos
Existing tools force you to choose between flexibility and reliability. You get either spaghetti prompt chains or rigid pipelines that break at scale.
No Observability
When an 8-step agent chain fails at step 5, you're left guessing. No traces, no state snapshots, no way to resume.
Amnesiac Agents
Every invocation starts from zero. No episodic memory, no learned patterns, no cognitive decay modeling.
One-Size-Fits-All
Quick tasks waste deep reasoning. Complex tasks get shallow treatment. No discipline-aware routing.
No Learning
Repeated patterns stay manual. Teams copy-paste prompts instead of evolving reusable skills automatically.
Built different, by design
Four pillars that make SoloFlow the most advanced open-source AI orchestration engine.
Cognitive Memory System
Four-tier cognitive memory architecture inspired by human cognition. Working memory holds active context, episodic memory stores experiences, and semantic memory accumulates abstracted knowledge — all governed by the Ebbinghaus forgetting curve.
R(t) = base × e(-t / stability)
Discipline-Aware Routing
Not all tasks are equal. SoloFlow automatically classifies incoming tasks by complexity and routes them to the right agent discipline — from quick single-shot responses to multi-agent deep reasoning sessions. No manual configuration needed.
DAG + FSM Hybrid Architecture
Why choose between workflow graphs and state machines? SoloFlow gives you both — the expressiveness of DAGs for defining task relationships, combined with the rigor of finite state machines for execution guarantees. Retry, pause, resume, branch — all first-class.
expressiveness(DAG) + rigor(FSM) = reliability
Skill Auto-Evolution
Stop reinventing workflows. SoloFlow detects repeated patterns across your agent runs and automatically packages them into versioned, reusable skills — which can be installed as Hermes skills for reuse. Your workflows literally improve themselves.
pattern → detect → package → skill
How the orchestration works
A deep dive into the DAG + FSM hybrid engine that powers every SoloFlow workflow.
Grounded in cognitive science
SoloFlow's memory system is modeled after human memory research, featuring the Ebbinghaus forgetting curve for realistic memory decay. Built on lobster-press cognitive memory engine.
Ebbinghaus Forgetting Curve
R(t) = base × e-t/stability
Working Memory
Active task context. Holds 7±2 items in focus. Fast access, short-lived. Cleared after task completion.
Episodic Memory
Time-stamped experiences from past runs. Enables "remember when" reasoning. Decays via Ebbinghaus curve with periodic consolidation.
Semantic Memory
Abstracted, persistent knowledge. Patterns distilled from episodic experience. Survives sessions. Forms the foundation of skill evolution.
Powered by lobster-press
lobster-press is SoloFlow's cognitive memory engine, built on cognitive science with DAG lossless compression, adaptive forgetting curve, and semantic memory capabilities.
Core Python Files
CMV Triple-Pass Lossless
Three-verification mechanism ensures zero loss in memory compression
C-HLR+ Adaptive Forgetting
Intelligently regulates memory decay based on Ebbinghaus curve
Focus Active Compression
Smart detection and triggering of memory compression timing
R³Mem Reversible 3-Layer
Working → Episodic → Semantic memory compression
WMR Tool Framework
Modular tool registration and invocation framework
How we compare
A feature-by-feature comparison with the most popular AI orchestration tools.
| Feature | ⚡ SoloFlow | CrewAI | LangGraph | AutoGPT | n8n |
|---|---|---|---|---|---|
| Discipline-Aware Routing | ✕ | ✕ | ✕ | ✕ | |
| Cognitive Memory System | ✕ | ✕ | ✕ | ✕ | |
| Skill Auto-Evolution | ✕ | ✕ | ✕ | ✕ | |
| DAG + FSM Hybrid | Partial | ✕ | |||
| Visual Builder | ✕ | ✕ | ✕ | ||
| Multi-user / RBAC | ✕ | ✕ | ✕ | ||
| Hermes Plugin Interface | ✕ | ✕ | ✕ | ✕ | |
| Lean runtime dependencies | ✕ | ✕ | ✕ | ✕ |
Up and running in seconds
Install SoloFlow, define your first workflow, and watch it orchestrate. It's that simple.
# Clone and install SoloFlow
git clone https://github.com/SonicBotMan/SoloFlow.git
cd SoloFlow
# One-command install
bash install.sh
# Or manual install
cp plugins/soloflow.py ~/.hermes/plugins/
cp -r skills/meta/soloflow ~/.hermes/skills/meta/
hermes skills reload
# Verify installation
hermes skills list | grep soloflow
# → soloflow ✓# Create a DAG workflow with parallel branches
import asyncio
from pathlib import Path
from hermes_plugin.store.sqlite_store import SQLiteStore
from hermes_plugin.services.workflow_service import WorkflowService
from hermes_plugin.services.scheduler import Scheduler
async def main():
store = SQLiteStore(Path("soloflow.db"))
store.initialize()
ws = WorkflowService(store)
ws.set_scheduler(Scheduler(store, ws))
# Define workflow with parallel branches
wf = await ws.create_workflow(
name="research-report",
description="行业调研报告",
steps=[
{"id": "search_a", "name": "学术搜索", "discipline": "quick"},
{"id": "search_b", "name": "行业搜索", "discipline": "quick"},
{"id": "merge", "name": "合并分析", "discipline": "deep"},
],
edges=[("search_a", "merge"), ("search_b", "merge")],
)
# Execute with DAG parallelism
await ws.start_workflow(wf["id"])# Mark workflow boundaries
/soloflow begin research-report
# Perform your steps...
# SoloFlow watches and aggregates automatically
/soloflow end
# Propose a skill from detected pattern
/soloflow propose
# Generate and install the skill
/soloflow generate research-report
# → Skill installed to ~/.hermes/skills/
# → Run 'hermes skills reload' to activateReady to orchestrate intelligently?
Join developers building the next generation of AI-powered tools with SoloFlow.