/MemGPT

Teaching LLMs memory management for unbounded context 📚🦙

Primary LanguagePythonApache License 2.0Apache-2.0

MemGPT logo

Try out our MemGPT chatbot on Discord!

Discord arXiv 2310.08560

🤖 Create perpetual chatbots with self-editing memory!


MemGPT demo video

🗃️ Chat with your data - talk to your SQL database or your local files!

SQL Database
MemGPT demo video for sql search
Local files
MemGPT demo video for sql search

📄 You can also talk to docs - for example ask about LlamaIndex!

MemGPT demo video for llamaindex api docs search
ChatGPT (GPT-4) when asked the same question:
GPT-4 when asked about llamaindex api docs
(Question from run-llama/llama_index#7756)

Quick setup

Join Discord and message the MemGPT bot (in the #memgpt channel). Then run the following commands (messaged to "MemGPT Bot"):

  • /profile (to create your profile)
  • /key (to enter your OpenAI key)
  • /create (to create a MemGPT chatbot)

Make sure your privacy settings on this server are open so that MemGPT Bot can DM you:
MemGPT → Privacy Settings → Direct Messages set to ON

set DMs settings on MemGPT server to be open in MemGPT so that MemGPT Bot can message you

You can see the full list of available commands when you enter / into the message box.

MemGPT Bot slash commands

What is MemGPT?

Memory-GPT (or MemGPT in short) is a system that intelligently manages different memory tiers in LLMs in order to effectively provide extended context within the LLM's limited context window. For example, MemGPT knows when to push critical information to a vector database and when to retrieve it later in the chat, enabling perpetual conversations. Learn more about MemGPT in our paper.

Running MemGPT locally

Install dependencies:

pip install -r requirements.txt

Extra step for Windows:

# only needed on Windows
pip install pyreadline

Add your OpenAI API key to your environment:

# on Linux/Mac
export OPENAI_API_KEY=YOUR_API_KEY
# on Windows
set OPENAI_API_KEY=YOUR_API_KEY

To run MemGPT for as a conversation agent in CLI mode, simply run main.py:

python3 main.py

To create a new starter user or starter persona (that MemGPT gets initialized with), create a new .txt file in /memgpt/humans/examples or /memgpt/personas/examples, then use the --persona or --human flag when running main.py. For example:

# assuming you created a new file /memgpt/humans/examples/me.txt
python main.py --human me.txt

GPT-3.5 support

You can run MemGPT with GPT-3.5 as the LLM instead of GPT-4:

python main.py --model gpt-3.5-turbo

Note that this is experimental gpt-3.5-turbo support. It's quite buggy compared to gpt-4, but it should be runnable.

Please report any bugs you encounter regarding MemGPT running on GPT-3.5 to cpacker#59.

main.py flags

--model
  select which model to use ('gpt-4', 'gpt-3.5-turbo-0613', 'gpt-3.5-turbo')
--persona
  load a specific persona file
--human
  load a specific human file
--first
  allows you to send the first message in the chat (by default, MemGPT will send the first message)
--debug
  enables debugging output
--archival_storage_faiss_path=<ARCHIVAL_STORAGE_FAISS_PATH>
  load in document database (backed by FAISS index)
--archival_storage_files="<ARCHIVAL_STORAGE_FILES_GLOB_PATTERN>"
  pre-load files into archival memory
--archival_storage_files_compute_embeddings="<ARCHIVAL_STORAGE_FILES_GLOB_PATTERN>"
  pre-load files into archival memory and also compute embeddings for embedding search
--archival_storage_sqldb=<SQLDB_PATH>
  load in SQL database

Interactive CLI commands

While using MemGPT via the CLI you can run various commands:

//
  enter multiline input mode (type // again when done)
/exit
  exit the CLI
/save
  save a checkpoint of the current agent/conversation state
/load
  load a saved checkpoint
/dump
  view the current message log (see the contents of main context)
/memory
  print the current contents of agent memory
/pop
  undo the last message in the conversation
/heartbeat
  send a heartbeat system message to the agent
/memorywarning
  send a memory warning system message to the agent

Example applications

Use MemGPT to talk to your Database!

MemGPT's archival memory let's you load your database and talk to it! To motivate this use-case, we have included a toy example.

Consider the test.db already included in the repository.

id name age
1 Alice 30
2 Bob 25
3 Charlie 35

To talk to this database, run:

python main.py  --archival_storage_sqldb=memgpt/personas/examples/sqldb/test.db

And then you can input the path to your database, and your query.

Please enter the path to the database. test.db
...
Enter your message: How old is Bob?
...
🤖 Bob is 25 years old.

Loading local files into archival memory

MemGPT enables you to chat with your data locally -- this example gives the workflow for loading documents into MemGPT's archival memory.

To run our example where you can search over the SEC 10-K filings of Uber, Lyft, and Airbnb,

  1. Download the .txt files from Hugging Face and place them in memgpt/personas/examples/preload_archival.

  2. In the root MemGPT directory, run

    python3 main.py --archival_storage_files="memgpt/personas/examples/preload_archival/*.txt" --persona=memgpt_doc --human=basic

If you would like to load your own local files into MemGPT's archival memory, run the command above but replace --archival_storage_files="memgpt/personas/examples/preload_archival/*.txt" with your own file glob expression (enclosed in quotes).

Enhance with embeddings search

In the root MemGPT directory, run

python3 main.py --archival_storage_files_compute_embeddings="<GLOB_PATTERN>" --persona=memgpt_doc --human=basic

This will generate embeddings, stick them into a FAISS index, and write the index to a directory, and then output:

  To avoid computing embeddings next time, replace --archival_storage_files_compute_embeddings=<GLOB_PATTERN> with
    --archival_storage_faiss_path=<DIRECTORY_WITH_EMBEDDINGS> (if your files haven't changed).

If you want to reuse these embeddings, run

python3 main.py --archival_storage_faiss_path="<DIRECTORY_WITH_EMBEDDINGS>" --persona=memgpt_doc --human=basic

Talking to LlamaIndex API Docs

MemGPT also enables you to chat with docs -- try running this example to talk to the LlamaIndex API docs!

  1. a. Download LlamaIndex API docs and FAISS index from Hugging Face.

    # Make sure you have git-lfs installed (https://git-lfs.com)
    git lfs install
    git clone https://huggingface.co/datasets/MemGPT/llamaindex-api-docs
    mv llamaindex-api-docs

    -- OR --

    b. Build the index:

    1. Build llama_index API docs with make text. Instructions here. Copy over the generated _build/text folder to memgpt/personas/docqa.
    2. Generate embeddings and FAISS index.
      cd memgpt/personas/docqa
      python3 scrape_docs.py
      python3 generate_embeddings_for_docs.py all_docs.jsonl
      python3 build_index.py --embedding_files all_docs.embeddings.jsonl --output_index_file all_docs.index
      
  2. In the root MemGPT directory, run

    python3 main.py --archival_storage_faiss_path=<ARCHIVAL_STORAGE_FAISS_PATH> --persona=memgpt_doc --human=basic

    where ARCHIVAL_STORAGE_FAISS_PATH is the directory where all_docs.jsonl and all_docs.index are located. If you downloaded from Hugging Face, it will be memgpt/personas/docqa/llamaindex-api-docs. If you built the index yourself, it will be memgpt/personas/docqa.

Support

If you have any further questions, or have anything to share, we are excited to hear your feedback!

  • By default MemGPT will use gpt-4, so your API key will require gpt-4 API access
  • For issues and feature requests, please open a GitHub issue or message us on our #support channel on Discord

Datasets

Datasets used in our paper can be downloaded at Hugging Face.

🚀 Project Roadmap

  • Release MemGPT Discord bot demo (perpetual chatbot)
  • Add additional workflows (load SQL/text into MemGPT external context)
  • Integration tests
  • Integrate with AutoGen (discussion)
  • Add official gpt-3.5-turbo support (discussion)
  • Release MemGPT family of open models (eg finetuned Mistral) (discussion)
  • CLI UI improvements (issue)
  • Add support for other LLM backends (issue)