CrewAI integration

agentmem-crewai is the Python integration for CrewAI . It ships two pre-built tools that match the shape of CrewAI’s built-in RecallMemoryTool / RememberTool — agents get hybrid retrieval and Postgres-native storage with one config object, no changes to the rest of your Crew code.


pip install agentmem-crewai

Quick start


from crewai import Agent, Crew, Task
from agentmem_crewai import AgentMemConfig, create_agentmem_tools
 
config = AgentMemConfig(
    api_key="sk-...",
    agent_id="support-bot",
    scope="team",        # optional default
)
 
support = Agent(
    role="Support specialist",
    goal="Help customers and remember what they prefer.",
    backstory="...",
    tools=create_agentmem_tools(config),
)
 
crew = Crew(agents=[support], tasks=[Task(...)])
crew.kickoff()

The agent gets two tools: Search memory (recall) and Save to memory (remember). Same names as CrewAI’s built-ins — agents pick them up automatically.

Tools

Tool	Purpose	Input
`AgentMemRecallTool`	Hybrid search (semantic + keyword + graph).	`queries: list[str]`
`AgentMemRememberTool`	Persist one or more facts.	`contents: list[str]`

Both shapes are identical to CrewAI’s built-ins (same name, args_schema, schema fields). The model doesn’t need re-prompting — it’ll use these tools without any tuning.

Helpers (without tools)

For pre-task setup, post-task extraction, or any flow outside the agent loop:


from agentmem_crewai import add_memory, recall_memories, search_memories
 
# Pre-task: inject memories into a task description
context = recall_memories("what does this customer prefer?", config)
# → "Relevant memories...\n- (score=0.85) ..."
 
# Raw hits for custom rendering
hits = search_memories("escalation policy", config)
 
# Post-task: persist a finding
add_memory("Customer asked about Q4 pricing.", config)

Config

Field	Default	Purpose
`api_key`	—	Required.
`agent_id`	—	Required.
`base_url`	hosted	Self-host override.
`scope`	`private`	Default visibility for writes.
`role`	—	Default agent role.
`workflow_id`	—	Group memories by workflow / session.
`top_k`	5	Hits per recall.
`rerank`	false	Re-score with Gemini (+0.5-1.5s).
`min_score`	0.0	Drop hits below this score.

What this replaces

CrewAI’s built-in memory uses LanceDB locally (or Mem0 if you configure that provider). This package gives you AgentMem’s Postgres-native memory through the same tool interface — no Memory class rewrite, no separate storage stack, no per-agent LanceDB files.