CrewAI Persistent Memory Backend

Why CrewAI Needs Better Memory

CrewAI crews are powerful for multi-agent workflows, but their built-in memory resets between kickoff() calls. For crews that run repeatedly — customer support, research, monitoring — this means starting from scratch every time.

Dakera provides:

• Persistent cross-run memory — crews retain knowledge indefinitely
• Namespace isolation — each agent gets its own memory space
• Shared crew knowledge — agents can read from shared namespaces
• Hybrid retrieval — HNSW vector + BM25 + cross-encoder reranking
• Knowledge graphs — track entities and relationships discovered by any agent

How It Works

pip install dakera crewai

Code Example: CrewAI + Dakera

from crewai import Agent, Task, Crew
from crewai.tools import tool
from dakera import Dakera

memory = Dakera(base_url="http://localhost:3300", api_key="dk-...")

@tool("recall_memory")
def recall_memory(query: str, agent_name: str) -> str:
    """Recall relevant memories for this agent."""
    results = memory.memory.recall(
        query=query,
        namespace=agent_name,
        top_k=5
    )
    return "\n".join([r["content"] for r in results["results"]])

@tool("store_memory")
def store_memory(content: str, agent_name: str) -> str:
    """Store a new memory for this agent."""
    memory.memory.store(
        content=content,
        namespace=agent_name,
        metadata={"source": "crewai"}
    )
    return f"Memory stored for {agent_name}"

# Create agents with persistent memory
researcher = Agent(
    role="Research Analyst",
    goal="Research topics and remember findings across runs",
    tools=[recall_memory, store_memory],
    backstory="You are a researcher who builds on previous findings."
)

writer = Agent(
    role="Content Writer",
    goal="Write content informed by research history",
    tools=[recall_memory, store_memory],
    backstory="You are a writer who maintains consistent style and builds on past work."
)

# Tasks that leverage persistent memory
research_task = Task(
    description="Research {topic}. First recall what you already know, then find new information and store it.",
    agent=researcher,
    expected_output="Research summary with new findings stored to memory"
)

write_task = Task(
    description="Write about {topic} using research findings. Store the final piece to memory.",
    agent=writer,
    expected_output="Written content that builds on previous work"
)

crew = Crew(agents=[researcher, writer], tasks=[research_task, write_task])
result = crew.kickoff(inputs={"topic": "AI agent architectures"})

Multi-Agent Memory Architecture

What Dakera Adds to CrewAI