memnos-sdk 0.1.3

Lightweight client + LangChain/LangGraph adapters for memnos — backend memory for AI agents.

memnos-sdk

Lightweight Python client for memnos — governed, vendor-neutral backend memory for AI agents. Use it directly, or as a LangChain retriever, a LangGraph long-term-memory store, or a LlamaIndex retriever.

httpx-only (no server deps). Talks to a running memnos server over REST.

Add it to your project with uv (or plain pip — it's an importable library, not a CLI app):

uv pip install memnos-sdk                  # core client   (or: pip install memnos-sdk)
uv pip install 'memnos-sdk[langchain]'     # + LangChain retriever
uv pip install 'memnos-sdk[langgraph]'     # + LangGraph BaseStore
uv pip install 'memnos-sdk[llamaindex]'    # + LlamaIndex retriever
uv pip install 'memnos-sdk[all]'           # everything

Core client (sync + async)

from memnos_sdk import MemnosClient

with MemnosClient(base_url="http://127.0.0.1:8900", token="mnk_...", namespace="org:acme") as mem:
    mem.remember("We chose PostgreSQL + pgvector for the memory store")
    ctx = mem.context("what database did we choose?")   # ready-to-inject; no LLM at query time
    rows = mem.recall("database decision")["memories"]   # ranked memories w/ scores + dates

from memnos_sdk import AsyncMemnosClient

async with AsyncMemnosClient(token="mnk_...", namespace="org:acme") as mem:
    await mem.remember("...")
    print(await mem.context("..."))

A token + namespace come from your memnos admin (memnos token <principal>, memnos grant <principal> <namespace>). Every call is namespace-scoped and audited server-side.

LangChain

from memnos_sdk import MemnosClient
from memnos_sdk.integrations.langchain import MemnosRetriever

retriever = MemnosRetriever(client=MemnosClient(token="mnk_...", namespace="org:acme"))
docs = retriever.invoke("auth token expiry policy")     # drop into any RAG chain
retriever.save("JWT tokens expire after 15 minutes in prod")

LangGraph (long-term memory)

from memnos_sdk import MemnosClient
from memnos_sdk.integrations.langgraph import MemnosStore

store = MemnosStore(MemnosClient(token="mnk_..."))
graph = builder.compile(store=store)
# in a node:  store.search(("org","acme"), query="...")  ·  store.put(("org","acme"), key, {"text": "..."})

memnos is semantic memory: put→remember, search→hybrid+reranked recall. Exact-key get is best-effort (use search).

LlamaIndex

from memnos_sdk import MemnosClient
from memnos_sdk.integrations.llamaindex import MemnosRetriever

retriever = MemnosRetriever(client=MemnosClient(token="mnk_...", namespace="org:acme"))
nodes = retriever.retrieve("auth token expiry policy")   # NodeWithScore[]; drop into a query engine
retriever.save("JWT tokens expire after 15 minutes in prod")

API surface

remember(text) · recall(query) -> {memories, context} · context(query) -> str · consolidate() · feedback(query, helpful) · healthy(). Async mirror on AsyncMemnosClient.

Apache-2.0.