Memory for Claude Agent SDK agents

• Which value is current. Harny stores a changed value with its timestamp so the newest value is the current one and the earlier value stays answerable by date.
• When a fact happened. You send a timestamp with each memory, set to when it happened. You can then ask what was true on a date with an as-of recall using queryTimestamp.
• One person across sources. Harny gives each person one private memory that the app's AI can read.

The code block connects the Harny MCP server to the agent. The key in the header is a person's key (fsk_), which names its own user, so calls are scoped to that person.

Keep the key on your server and do not put it in a browser.

import os
from claude_agent_sdk import ClaudeAgentOptions, query

options = ClaudeAgentOptions(
    mcp_servers={
        "harny": {
            "type": "http",
            "url": "https://api.harny.ai/v1/mcp/mcp",
            "headers": {"Authorization": f"Bearer {os.environ['HARNY_USER_KEY']}"},  # the signed-in person's key, set by your application
        }
    },
    allowed_tools=["mcp__harny__assemble_context", "mcp__harny__ingest_text"],
)

async def ask(question: str):
    async for message in query(prompt=question, options=options):
        print(message)

With a bearer key, the server lists the tools that key may call. A person's key (starts with fsk_) reaches the memory tools for that person only.

• assemble_context is the Recall Memory tool. It returns ranked context from the person's memory, scoped to that person.
• ingest_text is the Ingest Memory tool. It writes a note or fact into that person's memory.
• describe is available with any key.

• Use the SDK's own session state for the working context of a single run, such as the current task and the messages in the session.
• Add Harny when a fact must persist across sessions, change over time, or be answered as of a past date for the same person.