uams-sdk 1.0.2

Official Python SDK and MCP server for the Unified Agent Memory System (UAMS).

UAMS Shared Memory SDK

A unified, asynchronous Python SDK for the Unified Agent Memory System (UAMS). Designed to be shared natively by both Hermes and OpenClaw agents.

Features

• Unified Client: One standard interface for all agents.
• Async First: Built on httpx for non-blocking operations.
• Intelligent Caching: Built-in TTL caching with automatic invalidation on writes.
• Full API Coverage: Semantic search, graph traversal, context compression, and writes.
• MCP Adapter: Exposes UAMS as standard MCP tools, resources, and prompts for agent tool discovery.

Installation

pip install -e .

Usage

import asyncio
from uams_sdk import UAMSClient

async def main():
    client = UAMSClient(base_url="http://localhost:8000")
    task = "How do I fix Docker file locks?"

    # Default preflight before the agent works.
    memory = await client.begin_task(task)
    print(memory["context"])

    # Default write-back after durable work.
    await client.end_task(
        task=task,
        outcome="Documented the Docker file-lock workaround.",
        entities=["Docker File Locks"],
        tags=["#procedure"],
    )

asyncio.run(main())

MCP Server

Run the adapter from the repository root:

./uams mcp

Or run the installed console entry point:

UAMS_API_URL=http://localhost:8000 uams-mcp

Agent MCP config:

{
  "mcpServers": {
    "uams": {
      "command": "/absolute/path/to/unified_memory/uams",
      "args": ["mcp"],
      "env": {
        "UAMS_API_URL": "http://localhost:8000"
      }
    }
  }
}

Generate snippets from the repository root:

./uams mcp-config all
./uams mcp-config codex
./uams mcp-config json

Discovered MCP capabilities:

• Tools: begin_task, end_task, health, search_memory, get_context, get_procedures, remember, get_related_entities, summarize_memory, store_fix_summary
• Resource: uams://memory-policy
• Prompt: use_uams_memory

Agent default protocol:

Before each non-trivial task, call `begin_task`.
During work, call `search_memory` when more recall is needed.
After durable work, call `end_task`.
Never store raw transcripts; store distilled facts, decisions, fixes, and procedures.