Geniffy for the OpenAI Agents SDK

Give an agent built with the OpenAI Agents SDK a memory of each of your users. Before each model call, the model gets the user's briefing: where things stand, what is due, the rules that apply, what happened, and what is known that bears on their latest message. After the run, save it, tool calls included, into one memory for the whole conversation. When nothing is known, the model is told so, and says so instead of guessing.

pip install openai-agents-geniffy

Set GENIFFY_API_KEY from API keys in the Geniffy app. Keep it on your server.

The memory

from dataclasses import dataclass

from agents import Agent, Runner, SQLiteSession
from geniffy_agents import GeniffyMemory


@dataclass
class User:
    id: str


memory = GeniffyMemory(space=lambda user: f"user_{user.id}")      # read from the context you run with
agent = Agent[User](name="Assistant", instructions="You are a helpful assistant.", model="gpt-5.5")


async def chat(user: User, chat_id: str, message: str) -> str:
    session = SQLiteSession(chat_id)                                # or your own store of the conversation
    result = await Runner.run(agent, message, context=user, session=session, run_config=memory.run_config())
    await memory.save(user, result, session=chat_id)               # the run, saved into that conversation
    return result.final_output

• Before each model call, the user's briefing goes in the instructions, after the agent's own: where things stand, what is due or was promised, the rules that apply, what happened, then what is known that bears on their latest message, each line dated. In a run with tools, Geniffy is asked once, not once per step.
• save() keeps the run that just finished: what the user said, each tool the agent called and what it returned, and the answer. A conversation is one memory however long it gets; the history a session hands the run is not sent again, and a save that fails goes with the conversation's next one. Without session=, each run is a memory of its own.
• If Geniffy can't be reached, the run goes on without memory, and on_error hears about it.

run_config() takes your own RunConfig and adds the memory to it; a call_model_input_filter you already have runs after it.

Options: project (a string, or a function of the context) keeps the briefing to one project and labels what is saved with it; session names the conversation the same way, so save() needs no session=; briefing=False reads only what bears on the user's latest message; budget_chars is the most the briefing adds (6,000 characters unless you say); instructions changes what the model is told about the memory; and client shares your own geniffy.AsyncGeniffy. Without one, a client is made for each event loop and shared. An account without the briefing switched on is given what bears on the latest message instead, and the briefing is tried again ten minutes later. remember(user, message, result) still saves a single exchange, without the tool calls.

Tools

To let the agent decide when to look something up or save it:

from geniffy_agents import geniffy_tools

agent = Agent[User](name="Assistant", model="gpt-5.5",
                    tools=geniffy_tools(space=lambda user: f"user_{user.id}"))    # recall and remember

The tools read the user from the run's context, so the model never sees or chooses whose memory it reads.

One space per user

space is required: the user this is for, as a string or a function of the context you run with. Each space is a memory of its own, and nothing else can read it. space=None is your own memory, never your users' data, and a blank space, or None from your function, is refused, so a user with no id never lands in your own memory. When a user deletes their account, forget them with AsyncGeniffy().forget_space(...) from the geniffy SDK.

Develop

pip install -e ".[test]" && pytest     # through the SDK's own Runner, with its ScriptedModel, and the real
                                       # geniffy client against a stand-in for the API

Security

Report a vulnerability to ops@geniffy.com, not in a public issue. See the security policy.