crewAI Documentation | RealityOptimizer

Concept 1021 lines 40KB Benchmark 94.2

Memory

CrewAI's memory system enables agents to retain information across tasks and sessions. The unified `Memory` class replaces separate short-term, long-term, entity, and external memory stores.

Core API

from crewai import Memory

memory = Memory()

memory.remember("PostgreSQL chosen for user database.")

matches = memory.recall("What database did we choose?")

memory.forget(scope="/project/old")

How It Works

**LLM-based metadata extraction** — when saving a memory, the system infers scope, categories, and importance automatically.

**Composite scoring** — retrieval uses semantic similarity + recency + importance, with tuneable weights.

**Self-organizing scope trees** — hierarchical memory namespaces group related concepts.

| Feature | Implementation |

|---|---|

| Storage | In-memory dict / Mem0 / Chroma |

| Scoring | Semantic + Recency + Importance |

| Scopes | Hierarchical tree |

| Forgetting | Scoped deletion |

| Weights | `recency_weight=0.5`, `half_life_days=7` |

**Note:** This documentation was extracted from crewAI docs v1.14.4. See the [full memory docs](memory.md) for complete API reference.

View full Memory documentation →

Concept 808 lines 39KB Benchmark 91.5

📄 Raw

Agents

An Agent is the fundamental building block of CrewAI. Each agent has a role, goal, backstory, tools, and optional memory.

Creating an Agent

from crewai import Agent

researcher = Agent(

role="Senior Researcher",

goal="Find the latest AI trends",

backstory="You are a research analyst with 10+ years of experience.",

tools=[search_tool],

memory=True,

llm="gpt-4"

)

Key Attributes

| Attribute | Description |

|---|---|

| `role` | Defines the agent's function within the crew |

| `goal` | Clear, actionable objective |

| `backstory` | Personality and context for the agent |

| `tools` | Python functions the agent can call |

| `memory` | Whether the agent retains context |

| `llm` | Language model provider |

Agents can be configured with custom LLMs, delegated tasks, and equipped with skills for specialized workflows.

See [full agents docs](agents.md) for complete reference.

View full Agents documentation →

Concept 1143 lines 49KB Benchmark 90.8

📄 Raw

Tasks

Tasks define what agents should do. Each task has a description, expected output format, optional agent assignment, and tools.

Creating a Task

from crewai import Task

task = Task(

description="Analyze competitor documentation",

expected_output="Markdown report with 5 key findings",

agent=researcher,

tools=[scrape_tool],

context=[previous_task.output]

)

Task Types

| Type | Description |

|---|---|

| **Simple** | Single-step task with direct output |

| **Conditional** | Branching based on intermediate results |

| **Callback** | Execute function on completion |

| **Human-in-the-loop** | Pause for human approval |

Tasks support chaining via `context` (access previous task outputs), callbacks, and guardrails for output validation.

See [full tasks docs](tasks.md) for complete API.

View full Tasks documentation →

Concept 561 lines 31KB Benchmark 89.3

📄 Raw

Crews

A Crew orchestrates multiple agents working together. Define process types to control execution flow.

Creating a Crew

from crewai import Crew, Process

crew = Crew(

agents=[researcher, writer],

tasks=[task1, task2],

process=Process.sequential,

memory=True

)

crew.kickoff()

Process Types

| Process | Behavior |

|---|---|

| `sequential` | Tasks execute one after another |

| `hierarchical` | Manager agent assigns and reviews |

| `custom` | Define your own orchestration |

Crews support before/after kickoff hooks, result management, and resumable execution via checkpointing.

See [full crews docs](crews.md) for complete reference.

View full Crews documentation →

Concept 355 lines 21KB Benchmark 88.7

📄 Raw

Tools

Tools are Python functions that agents can call. CrewAI provides 100+ built-in tools and supports custom tool creation.

Built-in Tool Categories

| Category | Examples |

|---|---|

| **File & Document** | PDF, DOCX, CSV readers |

| **Web Scraping** | Browser automation, HTTP requests |

| **Search & Research** | Serper, Wikipedia, arXiv |

| **Database** | SQL queries, vector search |

| **AI/ML** | Image generation, text-to-speech |

| **Cloud** | S3, GCS, GitHub API |

Custom Tools

from crewai.tools import BaseTool

class MyTool(BaseTool):

name: str = "my_tool"

description: str = "Does something useful"

def _run(self, query: str) -> str:

return f"Result for {query}"

See [full tools docs](tools.md) for complete reference.

View full Tools documentation →

Integration 109 lines 5KB Benchmark 85.1

📄 Raw

MCP Integration

CrewAI supports Model Context Protocol (MCP) servers as tools, enabling agents to use any MCP-compatible service.

Supported Transports

| Transport | Use Case |

|---|---|

| **stdio** | Local MCP servers (e.g., filesystem, sqlite) |

| **SSE** | Remote servers with Server-Sent Events |

| **Streamable HTTP** | Full HTTP transport for complex interactions |

Multi-Server Setup

Connect multiple MCP servers simultaneously. Each server's tools are automatically registered with the assigned agents.

Security

- Server-specific authentication

- Tool permission scoping

- Connection validation before task execution

See [full MCP docs](mcp-integration.md) for DSL integration and security considerations.

View full MCP documentation →

Concept 127 lines 12KB Benchmark 87.2

📄 Raw

Processes

Processes define how crews execute tasks — the orchestration pattern that determines agent interaction.

Available Processes

| Process | Description |

|---|---|

| `sequential` | Linear pipeline. Each task completes before the next starts. |

| `hierarchical` | Manager agent breaks down work, assigns to workers, reviews output. |

| `parallel` | Tasks execute concurrently when possible. |

Choosing a Process

- **Sequential** — reliable for dependent tasks

- **Hierarchical** — best for complex projects requiring coordination

- **Parallel** — fastest for independent tasks

Process selection affects crew memory scope, callback timing, and result aggregation.

See [full processes docs](processes.md) for complete reference.

View full Processes documentation →