Skip to main content

AgentScope

Use PowerMem as a long-term memory tool provider for AgentScope agents through AgentScope's MCP client support. This keeps the integration lightweight: PowerMem owns memory storage and retrieval, while AgentScope discovers and calls PowerMem MCP tools through its MCP client.

AgentScope supports HTTP and stdio MCP servers, stateful and stateless MCP clients. PowerMem exposes the same memory tools through powermem-mcp, so no AgentScope-specific storage adapter is required.

Prerequisites​

  • PowerMem installed with MCP support:
pip install "powermem[server]"
  • PowerMem configured with your storage, LLM, and embedding providers.
  • AgentScope installed in the application environment.

Start PowerMem MCP​

Use streamable HTTP for a long-running local or remote AgentScope process:

powermem-mcp streamable-http 8848

PowerMem serves MCP at:

http://localhost:8848/mcp

If your AgentScope process should launch PowerMem itself, stdio is also available:

powermem-mcp stdio

Connect PowerMem tools in AgentScope​

import asyncio

from agentscope.mcp import HttpMCPConfig, MCPClient


async def list_powermem_tools() -> None:
    powermem = MCPClient(
        name="powermem",
        is_stateful=False,
        mcp_config=HttpMCPConfig(url="http://localhost:8848/mcp"),
    )

    tools = await powermem.list_tools()
    print([tool.name for tool in tools])


asyncio.run(list_powermem_tools())

The discovered tools include the standard PowerMem MCP memory operations:

  • add_memory
  • search_memories
  • get_memory_by_id
  • update_memory
  • delete_memory
  • delete_all_memories
  • list_memories

AgentScope wraps MCP tools with a model-facing name such as mcp__powermem__search_memories; use the original PowerMem tool name when calling get_tool().

Function-level calls​

AgentScope can obtain a single callable MCP tool by name. This is useful when a workflow wants explicit memory read/write steps.

import asyncio

from agentscope.mcp import HttpMCPConfig, MCPClient


async def search_powermem() -> None:
    powermem = MCPClient(
        name="powermem",
        is_stateful=False,
        mcp_config=HttpMCPConfig(url="http://localhost:8848/mcp"),
    )

    search_memories = await powermem.get_tool("search_memories")

    result = await search_memories(
        query="dragonfruit-zx9",
        user_id="agentscope-demo-user",
        limit=5,
    )
    print(result)


asyncio.run(search_powermem())

Verify​

  1. Start powermem-mcp streamable-http 8848.
  2. Run the AgentScope snippet above.
  3. Confirm search_memories and add_memory appear in the tool names printed by list_tools().
  4. Add a probe memory with add_memory and search for the same token with search_memories.

Troubleshooting​

  • If AgentScope cannot list tools, confirm http://localhost:8848/mcp matches the PowerMem MCP transport and port.
  • If memory calls fail, validate the PowerMem .env first. The MCP server depends on the same storage, LLM, and embedding configuration as the Python SDK.
  • If the AgentScope process runs in a different shell or container, pass the same PowerMem environment variables there, including POWERMEM_API_KEY when auth is enabled.
  • Use the generic MCP client guide for stdio, SSE, and auth variations.

References​