Integrate MemGPT in my codebase

[izevar]

Hello Everyone
I was looking for some documentation/guidance on how to integrate memgpt on my projects.
An API will be nice or a python library

any insight ?

[jaquielajoie]

Second this. I am trying to figure out how to make the CLI show up in a browser. I love the console app but would love to create my own frontend.

[BabellDev]

EDIT: this example was tested with memgpt 0.2.3 and no longer works with ~0.2.5; I have posted updated working code below

I've had some success invoking the memgpt agent from code, mostly by looking at how autogen integration was done. I will share my code here, which should at least serve as a starting point for your own projects.

This code depends on an existing config file. The easiest way to create the config file is via command line: memgpt configure

My code below expects a config file called openai_config, so you can copy the generated file and rename it. Alternatively, you can just go with the default and remove the line in the code that sets environment variable MEMGPT_CONFIG_PATH

import os

import memgpt.constants as constants
import memgpt.presets.presets as presets
import openai

from pathlib import Path
from memgpt.agent import Agent
from memgpt.config import AgentConfig
from memgpt.humans import humans
from memgpt.interface import CLIInterface as interface
from memgpt.persistence_manager import LocalStateManager
from memgpt.personas import personas

# Set your OpenAI API key
os.environ['OPENAI_API_KEY'] = 'TODO'

# Point to your config file (or remove this line if you just want to use the default)
os.environ['MEMGPT_CONFIG_PATH'] = Path.home().joinpath('.memgpt').joinpath('openai_config').as_posix()

# Necessary to avoid errors when LLM is invoked
openai.api_key = os.getenv('OPENAI_API_KEY')
openai.api_base = 'https://api.openai.com/v1'

agent_config = AgentConfig(
    name="BillsFriend",
    persona=personas.DEFAULT,
    human=humans.DEFAULT,
    model=constants.DEFAULT_MEMGPT_MODEL,
    #model="gpt-3.5-turbo-16k",
    preset=presets.DEFAULT_PRESET,
    model_endpoint_type="openai",
    model_endpoint="https://api.openai.com/v1",
)

skip_verify = agent_config.model != constants.DEFAULT_MEMGPT_MODEL

NEW_AGENT = True
if NEW_AGENT:
    persistence_manager = LocalStateManager(agent_config)

    memgpt_agent = presets.use_preset(
        preset_name=agent_config.preset,
        agent_config=agent_config,
        model=agent_config.model,
        persona=agent_config.persona,
        human=agent_config.human,
        interface=interface,
        persistence_manager=persistence_manager,
    )

    memgpt_agent.step(user_message="Hi my name is Bill Gates.  I love the Windows OS!", first_message=True, skip_verify=skip_verify)
    memgpt_agent.step(user_message="I am 68 years old, born October 28, 1955", first_message=False, skip_verify=skip_verify)
    memgpt_agent.save()
else:
    memgpt_agent = Agent.load_agent(interface, agent_config)
    memgpt_agent.step(user_message="What is my name?", first_message=False, skip_verify=skip_verify)
    memgpt_agent.step(user_message="What is my age?", first_message=False, skip_verify=skip_verify)

Run first with NEW_AGENT = True, which will create the agent and save the conversation; then run again with NEW_AGENT = False which will load the agent and ask some questions.

I'm really not sure if this is the best, or even correct way to use it; but at least it works and gives a jumping-off point for further exploration.

I've got an adaptation of the above code for Azure as well, if that's of interest.

[CUexter]

yea i think this is the way, the dev implemented a websocket server which uses memgpt logic but totally different from the cli interface

[BabellDev]

Yeah I had a quick look at the websocket client/server, and I can see that being a useful approach in certain scenarios, but I really wanted to use the memgpt client directly in my code.

[abkosar]

Hey @BabellDev! Appreciate this; however, I wonder how did you get from memgpt.personas import personas to work? Because when I run that line I get:

ImportError                               Traceback (most recent call last)
Cell In[19], line 1
----> 1 from memgpt.personas import personas

ImportError: cannot import name 'personas' from 'memgpt.personas' (/home/vscode/.local/lib/python3.9/site-packages/memgpt/personas/__init__.py)

Thanks!

[BabellDev]

Hi. It looks like the code i based this example on has changed recently. Could you please try with pymemgpt 0.2.3? Meanwhile, I will try to get it working again with the latest version, and will update the example if I'm successful.

btw I've been assisting another user with this demo code, and worked through various issues they were having. You can follow our discussion here: #480

Edit: if you are interested in updating the code yourself, take a look at how the persona is built here and then passed in the constructor of AgentConfig
https://github.com/cpacker/MemGPT/blob/d06a603c211a213790de789f63daeb356be507a0/memgpt/autogen/memgpt_agent.py#L42

[abkosar]

This is great actually thanks! My goal is to put this on a simple gradio app and allow users to chat with the file they upload. I am trying to figure out how to keep chatting with the agent. There is a run_agent_loop function but I haven't able to figure out how should I use it or if that is the write function to use. Do you have any guesses?

[BabellDev]

I haven't tried it, but you'd be on the right track looking at run_agent_loop in main.py.

I'm not sure of your use case, but that function has a lot of code for handing CLI commands, many of which I doubt you'd need (/dump etc). Personally, I would keep it simple and implement my own loop similar to this:
https://github.com/cpacker/MemGPT/blob/d06a603c211a213790de789f63daeb356be507a0/memgpt/main.py#L249

That snippet appears to be stripped-down handling of user input in a loop.

Hope it helps!

[abkosar]

Thanks, appreciate it! That helps!

[BabellDev]

@abkosar here is updated code that works with memgpt 0.2.5 (latest as of this comment).

You may need to generate a new config file using memgpt configure from a terminal.

Instructions are the same as the original code I posted.

import os
import openai

from pathlib import Path

from memgpt.agent import Agent as _Agent

from typing import Callable, Optional, List, Dict, Union, Any, Tuple

from memgpt.persistence_manager import LocalStateManager
import memgpt.constants as constants
import memgpt.utils as utils
import memgpt.presets.presets as presets
from memgpt.config import AgentConfig
from memgpt.interface import CLIInterface as interface

# Set your OpenAI API key
os.environ['OPENAI_API_KEY'] = 'TODO'

# Point to your config file (or remove this line if you just want to use the default)
os.environ['MEMGPT_CONFIG_PATH'] = Path.home().joinpath('.memgpt').joinpath('config').as_posix()

# Necessary to avoid errors when LLM is invoked
openai.api_key = os.getenv('OPENAI_API_KEY')
openai.api_base = 'https://api.openai.com/v1'

persona_desc = utils.get_persona_text(constants.DEFAULT_PERSONA)
user_desc = ""

# Create an AgentConfig option from the inputs
agent_config = AgentConfig(
    name="BillsFriend",
    persona=persona_desc,
    human=user_desc,
    preset="memgpt_chat",
    model="gpt-4",
    model_endpoint_type="openai",
    model_endpoint="https://api.openai.com/v1",
    context_window=8192,
)

skip_verify = agent_config.model != "gpt-4"

NEW_AGENT = False
if NEW_AGENT:
    persistence_manager = LocalStateManager(agent_config)

    memgpt_agent = presets.use_preset(
        preset_name=agent_config.preset,
        agent_config=agent_config,
        model=agent_config.model,
        persona=agent_config.persona,
        human=agent_config.human,
        interface=interface,
        persistence_manager=persistence_manager,
    )

    memgpt_agent.step(user_message="Hi my name is Bill Gates.  I love the Windows OS!", first_message=True, skip_verify=skip_verify)
    memgpt_agent.step(user_message="I am 68 years old, born October 28, 1955", first_message=False, skip_verify=skip_verify)
    memgpt_agent.save()
else:
    memgpt_agent = Agent.load_agent(interface, agent_config)
    memgpt_agent.step(user_message="What is my name?", first_message=False, skip_verify=skip_verify)
    memgpt_agent.step(user_message="What is my age?", first_message=False, skip_verify=skip_verify)

I will experiment with implementing REPL (loop) and post my results if successful

[cpacker]

that would be awesome!! feel free to DM me on discord (if you use discord) - I'm definitely more active there than the github discussion pages but I'll try to keep on top of notifications I get from this thread

[BabellDev]

I've actually never used discord, but maybe this is reason enough to finally give it a try. 😛 In any case, I'll keep this thread updated.

[cpacker]

@BabellDev did some light code edits for another github issue, may be of interest (though very messy, but the general gist of the issue is related): #689

[BabellDev]

Thanks @cpacker! I'll be looking at this a little later today

[cpacker]

@BabellDev let's start tracking discussions on it here: #697

[Priyansh-jsk]

[cpacker]

Try checking the code example I wrote in this github issue: #689

We are also planning on making the python code much easier to use: #697

Also, there's an API being build that you may like to use instead: https://memgpt.readme.io/reference

[Priyansh-jsk]

@cpacker I already check the code that you have written, but I do not want to use any openAPI key. I simply doing want to integrate MemGPT into Gradio interface. So i can run memgpt into Gradio interface.

I already try to create above code to work in Gradio interface but it is not working. I'm looking forward to hiring from you.
I hope you understand. Thanks!

[Priyansh-jsk]

yes @cpacker any update or anyone work on it?
Thanks!

[cpacker]

For future reference, a big (community) PR by @BabellDev was merged that adds a clean way to interact with MemGPT via python (via a "python client"): #713. This should be very useful for connecting to existing Python chat/web UI interfaces like gradio.

Documentation: https://memgpt.readme.io/docs/python_client

This is not in the main package yet, it will be added as part of 0.2.11. If you want to use this now, install from source or use pip install pymemgpt-nightly.