Geniffy for LlamaIndex

Give a LlamaIndex agent a memory of each of your users. Add one memory block, and before the agent answers, LlamaIndex puts the user's briefing into the system message: where things stand, what is due, the rules that apply, what happened, and what is known that bears on their message. After each 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 llama-index-memory-geniffy

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

The memory block

from llama_index.core.agent.workflow import FunctionAgent
from llama_index.core.memory import Memory
from llama_index.llms.anthropic import Anthropic
from llama_index.memory.geniffy import GeniffyMemoryBlock

agent = FunctionAgent(llm=Anthropic(model="claude-opus-5-5"), system_prompt="You are a helpful assistant.")


async def chat(user_id: str, chat_id: str, message: str) -> str:
    block = GeniffyMemoryBlock(space=f"user_{user_id}")       # the user from your own sign-in
    memory = Memory.from_defaults(session_id=chat_id, memory_blocks=[block])
    reply = await agent.run(user_msg=message, memory=memory)
    await block.save(memory)                                  # the run, saved into that user's conversation
    return str(reply)

• Before the agent answers, the user's briefing goes in the system message, inside LlamaIndex's <memory> section: 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(memory) keeps the run that just finished: what the user said, each tool the agent called and what it returned, and the answer. The Memory's session_id is the conversation, and a conversation is one memory however long it gets. Saving the same run twice sends nothing the second time, and a save that fails goes with the next one. An app that keeps the history itself passes its messages and session= its id instead.
• If Geniffy can't be reached, the agent goes on without memory, and on_error hears about it.

Options: project keeps the briefing to one project and labels what is saved with it; 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 by every block. 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.

LlamaIndex hands a block only the messages that overflow its short-term history, which most conversations never reach, so save() is how each run gets to Geniffy. To save the overflow instead, pass accept_short_term_memory=True and don't call save(), or a run is saved twice. remember(message, reply) still saves a single exchange, without the tool calls.

Tools

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

from llama_index.memory.geniffy import geniffy_tools

agent = FunctionAgent(llm=llm, tools=geniffy_tools(space=f"user_{user_id}"))   # recall and remember

Make the tools per user (per request), so the model never sees or chooses whose memory it reads.

One space per user

space is required: the user this is for. 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 is refused, so a user with no id can't end up in it. When a user deletes their account, forget them with AsyncGeniffy().forget_space(...) from the geniffy SDK.

Develop

pip install -e ".[test]" && pytest     # through LlamaIndex's own FunctionAgent and Memory, with a scripted LLM
                                       # 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.