Sulcus × CrewAI

Shared thermodynamic memory for multi-agent CrewAI crews. Every agent in the crew reads and writes to the same Sulcus memory graph, with automatic heat propagation across agent boundaries.

Install

pip install sulcus-crewai

Quick Start

from crewai import Agent, Crew, Task
from sulcus import Sulcus
from sulcus_crewai import SulcusSearchTool, SulcusStoreTool

client = Sulcus(api_key="sk-...")

# Tools — give to any agent
search = SulcusSearchTool(client=client)
store = SulcusStoreTool(client=client)

researcher = Agent(
    role="Researcher",
    tools=[search, store],
    goal="Find and store key findings",
)

writer = Agent(
    role="Writer",
    tools=[search],
    goal="Compose reports from existing research",
)

crew = Crew(agents=[researcher, writer], tasks=[...])
crew.kickoff()

Components

Tools (for agents)

Tool	Description
SulcusSearchTool	Search memories by natural language query. Returns results ranked by relevance + heat.
SulcusStoreTool	Store a new memory with type (episodic, semantic, preference, procedural).
SulcusContextTool	Build a structured context window from relevant memories within a token budget.

from sulcus_crewai import SulcusSearchTool, SulcusStoreTool, SulcusContextTool

search = SulcusSearchTool(client=client)
store = SulcusStoreTool(client=client)
context = SulcusContextTool(client=client)

agent = Agent(role="Analyst", tools=[search, store, context])

Storage (crew-level shared state)

from sulcus_crewai import SulcusStorage

storage = SulcusStorage(client=client, namespace="research-crew")

# Store findings
storage.save("market_size", "AI memory market estimated at $2.4B by 2027")

# Retrieve by semantic search
results = storage.load("market size")

# List recent
recent = storage.list_recent(limit=10)

# Filter by type
facts = storage.search_by_type("pricing", memory_type="semantic")

# Clean up
storage.forget(node_id="...")

Memory Types

Type	Use For	Decay
episodic	Events, conversations, findings	Fast (24h half-life)
semantic	Facts, data, knowledge	Slow (30d half-life)
preference	Opinions, settings, style	Slower (90d half-life)
procedural	Workflows, how-tos, recipes	Slowest (180d half-life)

How It Works

1. Researcher agent discovers something → calls sulcus_store → memory created in Sulcus graph
2. Writer agent needs context → calls sulcus_search → retrieves the researcher's findings
3. Next crew run → memories persist, ranked by heat (recency × frequency × relevance)
4. Over time → episodic findings decay; semantic facts stay; preferences are nearly permanent

All agents share one Sulcus tenant. The thermodynamic engine handles prioritization — no manual memory management needed.

Self-Hosted

client = Sulcus(
    api_key="your-key",
    server_url="http://localhost:4200",  # your Sulcus server
)

Examples

See examples/research_crew.py for a complete working example.

Links

• Sulcus Documentation
• Sulcus Python SDK
• All Integrations
• CrewAI Documentation