Python client SDK for AtomicMemory memory and artifact storage.

Docs: docs.atomicstrata.ai

AtomicMemory Core currently reaches cost-Pareto SOTA on BEAM-100K, BEAM-1M, and LoCoMo10, with BEAM-10M parity against the strongest published Mem0-new result. This package brings that memory layer to Python services, agents, notebooks, and evaluation workflows.

A backend-agnostic memory and storage client: ingest conversations and documents, search them semantically, package retrieval-ready context, register or upload raw artifacts, and access AtomicMemory-specific features (lifecycle, audit, lessons, agents/trust, runtime config) through typed namespace handles.

This is a Python port of the TypeScript atomicmemory-sdk. It mirrors the public surface 1:1 while staying idiomatic to Python (Pydantic models, httpx sync + async clients, match statements, snake_case).

Stable release — 1.0.0 on PyPI.

pip install atomicmemory                    # core + local search + SQLite store
pip install 'atomicmemory[embeddings]'      # + sentence-transformers for local embeddings

Prerequisite: start atomicmemory-core first. Follow the Core Quickstart if you do not already have a backend at http://localhost:3050.

from atomicmemory import AtomicMemoryClient

with AtomicMemoryClient({
    "apiUrl": "http://localhost:3050",
    "apiKey": "server-api-key",
    "userId": "demo",
}) as client:
    client.memory.initialize()

    client.memory.ingest({
        "mode": "messages",
        "messages": [
            {"role": "user", "content": "I prefer aisle seats on flights."},
        ],
        "scope": {"user": "demo"},
    })

    page = client.memory.search({"query": "seat preference", "scope": {"user": "demo"}})
    for hit in page.results:
        print(hit.memory.content, hit.score)

    artifact = client.storage.put({
        "mode": "pointer",
        "uri": "https://example.com/manual.pdf",
        "contentType": "application/pdf",
    })
    print(artifact.artifact_id)

import asyncio
from atomicmemory import AsyncAtomicMemoryClient

async def main() -> None:
    async with AsyncAtomicMemoryClient({
        "apiUrl": "http://localhost:3050",
        "apiKey": "server-api-key",
        "userId": "demo",
    }) as client:
        await client.memory.initialize()
        results = await client.memory.search({"query": "seat preference", "scope": {"user": "demo"}})
        for hit in results.results:
            print(hit.memory.content)

asyncio.run(main())

When configured with the atomicmemory provider, the client exposes a typed handle for backend-specific routes:

trail = client.memory.atomicmemory.audit.trail(memory_id="mem-123", user_id="demo")
health = client.memory.atomicmemory.config.health()

Categories: lifecycle, audit, lessons, config, agents.

The memory namespace supports the same provider family as the TypeScript SDK:

• atomicmemory — AtomicMemory core backend.
• mem0 — Mem0 OSS or hosted backend.
• hindsight — Hindsight Cloud or self-hosted backend.

from atomicmemory import MemoryClient

with MemoryClient(
    providers={
        "hindsight": {
            "apiUrl": "http://localhost:8888",
            "apiVersion": "v1",
            "projectId": "default",
        }
    }
) as memory:
    memory.initialize()
    page = memory.search({"query": "seat preference", "scope": {"user": "demo"}})

The client.storage namespace mirrors the TypeScript SDK's direct storage API:

• capabilities() reports active backend support.
• put({"mode": "pointer", ...}) registers a pointer to caller-owned bytes.
• put({"mode": "managed", "body": b"...", ...}) uploads known-length bytes to the configured raw content store.
• get, get_content, head, delete, and verify address artifacts by artifact_id.
• stream_content streams large artifact bodies without buffering the entire response in memory.

Every storage request sends Authorization: Bearer <apiKey> and X-AtomicMemory-User-Id. The SDK never sends the legacy ?user_id= URL parameter.

uv sync --extra dev --extra embeddings
uv run pytest
uv run ruff check .
uv run ruff format --check .
uv run mypy atomicmemory --strict
uv run vulture atomicmemory tests .vulture_whitelist.py --min-confidence 90

Live provider tests are opt-in and are not required for normal development. They assume the backend is already running and configured with its own model.

ATOMICMEMORY_HINDSIGHT_INTEGRATION=1 \
HINDSIGHT_API_URL=http://localhost:8890 \
HINDSIGHT_TIMEOUT_SECONDS=120 \
uv run pytest tests/providers/hindsight/test_integration.py -m integration -ra

Apache-2.0