Designing per-customer agent instances with isolated memory in AgentScope

[nottydadie]

Hi everyone,

I’m working on a multi-agent system using AgentScope and wanted to get some guidance on architecture and best practices.

My use case requires:

Running agents per customer (tenant)
Each agent instance should maintain separate memory per customer
Additionally, each agent (within a customer context) should have its own isolated memory (not shared across agents or customers)

So effectively:

Customer A → Agent X (memory A-X), Agent Y (memory A-Y)
Customer B → Agent X (memory B-X), Agent Y (memory B-Y)

I’m trying to understand:

What’s the recommended way in AgentScope to scope or namespace memory like this?
Should I instantiate separate agent objects per customer, or is there a built-in way to parameterize memory/context?
How should memory persistence be handled (e.g., external store, vector DB, session-based memory)?
Are there any patterns for scaling this when the number of customers grows?
How would this interact with an orchestrator agent delegating tasks to sub-agents?

agentscope.init(
    studio_url=settings.AGENTSCOPE_STUDIO_URL,
    project=settings.AGENTSCOPE_PROJECT_NAME,
    logging_level="DEBUG"
)

Should I give separate run_id per run per customer?

OR

agent_a = ReActAgent(
    name="Agent A",
    sys_prompt="You are a helpful assistant.",
    model=MODEL,
    memory=RedisMemory(
        host=settings.REDIS_HOST,
        port=settings.REDIS_PORT,
        password=settings.REDIS_PASS,
        db=settings.REDIS_DB,
        user_id=user_id,
        session_id=session_id
    ),
    formatter=OpenAIChatFormatter(),
    toolkit=toolkit,
    long_term_memory=long_term_memory,
)

Just provide separate memory for each customer.

If anyone has implemented something similar or can point me to best practices/examples, I’d really appreciate it.

Thanks!