jennah-sdk-py 0.65.0

jennah-sdk-py

The Python SDK for Jennah, the memory and context platform for AI agents.

pip install jennah-sdk-py

Store and recall a memory

Sign in once with the jnh CLI (jnh login), or set JENNAH_API_KEY. Then:

import uuid

from jennah import Client
from jennah.agent.v1 import agent_pb2, memory_pb2

with Client() as client:  # uses `jnh login`, or $JENNAH_API_KEY
    agent_id = f"quickstart-{uuid.uuid4().hex[:8]}"
    client.agents.CreateAgent(
        agent_pb2.CreateAgentRequest(agent_instance_id=agent_id)
    )
    agent = client.agent(agent_id)

    # Store a memory. The platform embeds raw_content for you.
    agent.vectors.upsert(memory_pb2.VectorChunk(
        chunk_id="pref-1",
        raw_content="The customer prefers invoices in Japanese yen, "
        "sent on the 5th.",
    ))

    # Recall it by meaning, not by keyword.
    result = agent.vectors.search(memory_pb2.SemanticQuery(
        query_text="what currency does the customer want to be "
        "billed in?",
        limit=3,
    ))
    for m in result.matches:
        print(f"{m.distance:.3f}  {m.raw_content}")

    client.agents.DeleteAgent(
        agent_pb2.DeleteAgentRequest(agent_instance_id=agent_id)
    )

0.235  The customer prefers invoices in Japanese yen, sent on the 5th.

Async

AsyncClient has the same operations, awaited, over grpc.aio. Nothing it does blocks the event loop, credential renewal included.

import asyncio

from jennah import AsyncClient
from jennah.agent.v1 import memory_pb2


async def main():
    async with AsyncClient() as client:
        query = memory_pb2.SemanticQuery(
            query_text="billing preferences", limit=3
        )
        result = await client.agent("my-agent").vectors.search(query)
        print(result.matches)


asyncio.run(main())

Use Client in scripts, notebooks and evaluation harnesses; use AsyncClient inside an agent framework or any other asyncio program.

Every operation, already authenticated

Each service in the API is a generated stub on the client, already bound to the credentialed connection, so every operation is callable without a wrapper:

from jennah.agent.v1 import scope_pb2

client.scopes.ListScopes(scope_pb2.ListScopesRequest())
client.memory.InspectMemory(...)
client.datasets.ListDatasets(...)

The services are agents, memory, scopes, datasets, schema, data, auth, approvals, billing and platform. For a service added after your SDK version, client.stub(SomeServiceStub) binds it the same way. Never build your own channel to reach one: you would lose credential renewal.

client.agent(id) adds the memory shortcuts, each one single-section call: memory.commit(...) and memory.query(...) (any request field, as keywords), logs.create/logs.recent, vectors.upsert/vectors.search, and graph.write/graph.query.

Credentials

The client takes the first credential it finds, in this order:

1. credentials=, a source your program supplies.
2. api_key=.
3. The JENNAH_API_KEY environment variable.
4. The session stored by jnh login.

A signed-in session renews itself: a rejected call is renewed once and retried, and the renewed session is written back to the shared credentials file, so jnh on the same machine stays signed in. client.credential reports what authenticated the client, and where it came from, without the secret.

Development

The generated code is not committed. With jennah-api checked out next to this repository:

scripts/dev-generate.sh      # generate and lay the stubs and conformance suite in place
pip install -e '.[test]'
pytest