Python SDK reference

Package: agentmem-py (Apache-2.0).


pip install agentmem-py

Python ≥ 3.8. Depends only on httpx.

Synchronous client


import os
from agentmem_py import MemoryStore
 
mem = MemoryStore(
    api_key=os.environ["AGENTMEM_API_KEY"],
    base_url=None,                  # None = use SDK default
)

Async client


import asyncio
from agentmem_py import AsyncMemoryStore
 
async def main():
    async with AsyncMemoryStore(api_key="sk-...") as mem:
        await mem.write(content="hello", agent_id="agent-1", scope="team")
        hits = await mem.search(query="hello", agent_id="agent-1")
        for h in hits:
            print(h.score, h.content)
 
asyncio.run(main())

Both clients have the same method surface; the async one returns coroutines.

Memory


result = mem.write(
    content="Customer prefers email",
    agent_id="support-bot",
    scope="team",
    role="executor",            # optional
    workflow_id="wf-123",       # optional
    expires_at=None,            # optional ISO-8601 string
)
# result.write_id, result.duplicate
 
hits = mem.search(
    query="customer preferences",
    agent_id="support-bot",
    top_k=5,
    scope=None,
    workflow_id=None,
    at_time=None,
    rerank=False,
)
# hits[i].score, .content, .agent_id, .scope, .role, .workflow_id,
# .created_at, .vector_clock

What’s exported


from agentmem_py import (
    MemoryStore, AsyncMemoryStore,
    MemoryHit, Conflict,
)

The Python SDK in v0.2 covers write, search, and conflict-detection responses. Other admin endpoints (keys, teams, policies, webhooks, grants, retention, BYOK) — call them via raw HTTP with httpx for now, or use the TypeScript SDK. Full parity is planned for v0.3.

Source

github.com/rooney011/agentmem-sdk (Python source under python/).