deepmem-client

Drop-in Mem0-compatible client for the DeepMem cloud memory API. Switch from Mem0 in one line - same class name, same method signatures, same filters={"user_id": ...} style as mem0.MemoryClient.

deepmem-client is a thin HTTP client for the DeepMem cloud memory API (https://deepmem.dev). It mirrors mem0.MemoryClient's class name and method signatures, so migrating an app from Mem0 is a one-line import change - no torch, no Qdrant, no server to run. Just pip install and point it at your cloud key.

Install

pip install deepmem-client

The one-line switch

# ── before: Mem0 cloud ──
from mem0 import MemoryClient
client = MemoryClient(api_key="m0-xxxxxxxx")
client.add(messages, user_id="alex")
client.search("What can Alex cook?", filters={"user_id": "alex"})

# ── after: DeepMem cloud ──
from deepmem import MemoryClient          # ← only this line changes
client = MemoryClient(api_key="dm_live_xxxxxxxx")   # key at https://deepmem.dev
client.add(messages, user_id="alex")
client.search("What can Alex cook?", filters={"user_id": "alex"})   # identical

That's it. The class name (MemoryClient), method names, and the filters={"user_id": ...} style all match mem0.MemoryClient, so the rest of your code stays untouched. (from deepmem import Memory works too as an alias for mem0's OSS Memory class, with search(query, user_id=...) style.)

Set the key via env and skip passing it each time:

export DEEPMEM_API_KEY=dm_live_xxxxxxxx

from deepmem import mem0_client as client   # singleton, reads DEEPMEM_API_KEY / DEEPMEM_BASE_URL
client.add(messages, user_id="alex")

Method reference (mirrors mem0.MemoryClient)

messages may be a list of {role, content} dicts, a single dict, or a string (just like mem0). search / get_all accept filters={"user_id": "..."} (mem0 style) or a top-level user_id= (convenience).

Behavioral differences from Mem0 (read this)

These are inherent to DeepMem and can't be hidden behind the client:

1. add(infer=True) is asynchronous on DeepMem cloud. add() returns immediately with pending=True and results=[]; extracted facts land a few seconds later. (Mem0 cloud's add is also async - it returns PENDING - so this is actually symmetric.) Pass infer=False for synchronous raw-text storage that is immediately searchable (no fact extraction) - useful when you want add then search to work back-to-back.

2. No graph relations. DeepMem uses hybrid vector retrieval (vector + BM25 + time-decay), not a knowledge graph. "relations" is always [].

3. reset differs. Mem0's reset() is account-wide (deletes ALL users); DeepMem's is per-user_id with a confirm guard, so this client's reset requires a user_id (arg / filters / default).

4. Self-hosted? Point the client at your own server: MemoryClient(api_key=None, host="http://localhost:8000"). Self-hosted DeepMem runs in open mode (no API key needed).

License

MIT. The full DeepMem server (with LLM extraction, Qdrant, BGE-M3 embeddings, MCP) is at github.com/deepmemteam/deepmem.