Configuration
Memory

Setting Your Cache Type

By default, Auto-GPT set up with Docker Compose will use Redis as its memory backend. Otherwise, the default is LocalCache (which stores memory in a JSON file).

To switch to a different backend, change the MEMORY_BACKEND in .env to the value that you want:

  • local uses a local JSON cache file
  • pinecone uses the Pinecone.io account you configured in your ENV settings
  • redis will use the redis cache that you configured
  • milvus will use the milvus cache that you configured
  • weaviate will use the weaviate cache that you configured

Memory Backend Setup

Links to memory backends

Redis Setup

Important

If you have set up Auto-GPT using Docker Compose, then Redis is included, no further setup needed.

Caution

This setup is not intended to be publicly accessible and lacks security measures. Avoid exposing Redis to the internet without a password or at all!

  1. Launch Redis container

    docker run -d --name redis-stack-server -p 6379:6379 redis/redis-stack-server:latest
  2. Set the following settings in .env

    MEMORY_BACKEND=redis
    REDIS_HOST=localhost
    REDIS_PORT=6379
    REDIS_PASSWORD="PASSWORD"

    Replace "PASSWORD" by your password, omitting the double quotations ("").

    Optional configuration:

    • WIPE_REDIS_ON_START=False to persist memory stored in Redis between runs.
    • MEMORY_INDEX=<WHATEVER> to specify a name for the memory index in Redis. The default is auto-gpt.

Info

See redis-stack-server (opens in a new tab) for setting a password and additional configuration.

🌲 Pinecone API Key Setup

Pinecone lets you store vast amounts of vector-based memory, allowing the agent to load only relevant memories at any given time.

  1. Go to pinecone (opens in a new tab) and make an account if you don't already have one.
  2. Choose the Starter plan to avoid being charged.
  3. Find your API key and region under the default project in the left sidebar.

In the .env file set:

  • PINECONE_API_KEY
  • PINECONE_ENV (example: us-east4-gcp)
  • MEMORY_BACKEND=pinecone

Milvus Setup

Milvus (opens in a new tab) is an open-source, highly scalable vector database to store huge amounts of vector-based memory and provide fast relevant search. It can be quickly deployed with docker, or as a cloud service provided by Zilliz Cloud (opens in a new tab).

  1. Deploy your Milvus service, either locally using docker or with a managed Zilliz Cloud database:

    • Install and deploy Milvus locally (opens in a new tab)

    • Set up a managed Zilliz Cloud database

      1. Go to Zilliz Cloud (opens in a new tab) and sign up if you don't already have account.
      2. In the Databases tab, create a new database.
        • Remember your username and password
        • Wait until the database status is changed to RUNNING.
      3. In the Database detail tab of the database you have created, the public cloud endpoint, such as: https://xxx-xxxx.xxxx.xxxx.zillizcloud.com:443.
  2. Run pip3 install pymilvus to install the required client library. Make sure your PyMilvus version and Milvus version are compatible (opens in a new tab) to avoid issues. See also the PyMilvus installation instructions (opens in a new tab).

  3. Update .env:

    • MEMORY_BACKEND=milvus
    • One of:
      • MILVUS_ADDR=host:ip (for local instance)
      • MILVUS_ADDR=https://xxx-xxxx.xxxx.xxxx.zillizcloud.com:443 (for Zilliz Cloud)

    The following settings are optional:

    • MILVUS_USERNAME='username-of-your-milvus-instance'
    • MILVUS_PASSWORD='password-of-your-milvus-instance'
    • MILVUS_SECURE=True to use a secure connection. Only use if your Milvus instance has TLS enabled. Note: setting MILVUS_ADDR to a https:// URL will override this setting.
    • MILVUS_COLLECTION to change the collection name to use in Milvus. Defaults to autogpt.

Weaviate Setup

Weaviate (opens in a new tab) is an open-source vector database. It allows to store data objects and vector embeddings from ML-models and scales seamlessly to billion of data objects. To set up a Weaviate database, check out their Quickstart Tutorial (opens in a new tab).

Although still experimental, Embedded Weaviate (opens in a new tab) is supported which allows the Auto-GPT process itself to start a Weaviate instance. To enable it, set USE_WEAVIATE_EMBEDDED to True and make sure you pip install "weaviate-client>=3.15.4".

Install the Weaviate client

Install the Weaviate client before usage.

$ pip install weaviate-client

Setting up environment variables

In your .env file set the following:

MEMORY_BACKEND=weaviate
WEAVIATE_HOST="127.0.0.1" # the IP or domain of the running Weaviate instance
WEAVIATE_PORT="8080" 
WEAVIATE_PROTOCOL="http"
WEAVIATE_USERNAME="your username"
WEAVIATE_PASSWORD="your password"
WEAVIATE_API_KEY="your weaviate API key if you have one"
WEAVIATE_EMBEDDED_PATH="/home/me/.local/share/weaviate" # this is optional and indicates where the data should be persisted when running an embedded instance
USE_WEAVIATE_EMBEDDED=False # set to True to run Embedded Weaviate
MEMORY_INDEX="Autogpt" # name of the index to create for the application

View Memory Usage

View memory usage by using the --debug flag :)

🧠 Memory pre-seeding

Memory pre-seeding allows you to ingest files into memory and pre-seed it before running Auto-GPT.

$ python data_ingestion.py -h 
usage: data_ingestion.py [-h] (--file FILE | --dir DIR) [--init] [--overlap OVERLAP] [--max_length MAX_LENGTH]
 
Ingest a file or a directory with multiple files into memory. Make sure to set your .env before running this script.
 
options:
  -h, --help               show this help message and exit
  --file FILE              The file to ingest.
  --dir DIR                The directory containing the files to ingest.
  --init                   Init the memory and wipe its content (default: False)
  --overlap OVERLAP        The overlap size between chunks when ingesting files (default: 200)
  --max_length MAX_LENGTH  The max_length of each chunk when ingesting files (default: 4000)
 
# python data_ingestion.py --dir DataFolder --init --overlap 100 --max_length 2000

In the example above, the script initializes the memory, ingests all files within the Auto-Gpt/autogpt/auto_gpt_workspace/DataFolder directory into memory with an overlap between chunks of 100 and a maximum length of each chunk of 2000.

Note that you can also use the --file argument to ingest a single file into memory and that data_ingestion.py will only ingest files within the /auto_gpt_workspace directory.

The DIR path is relative to the auto_gpt_workspace directory, so python data_ingestion.py --dir . --init will ingest everything in auto_gpt_workspace directory.

You can adjust the max_length and overlap parameters to fine-tune the way the documents are presented to the AI when it "recall" that memory:

  • Adjusting the overlap value allows the AI to access more contextual information from each chunk when recalling information, but will result in more chunks being created and therefore increase memory backend usage and OpenAI API requests.
  • Reducing the max_length value will create more chunks, which can save prompt tokens by allowing for more message history in the context, but will also increase the number of chunks.
  • Increasing the max_length value will provide the AI with more contextual information from each chunk, reducing the number of chunks created and saving on OpenAI API requests. However, this may also use more prompt tokens and decrease the overall context available to the AI.

Memory pre-seeding is a technique for improving AI accuracy by ingesting relevant data into its memory. Chunks of data are split and added to memory, allowing the AI to access them quickly and generate more accurate responses. It's useful for large datasets or when specific information needs to be accessed quickly. Examples include ingesting API or GitHub documentation before running Auto-GPT.

Attention

If you use Redis for memory, make sure to run Auto-GPT with WIPE_REDIS_ON_START=False

For other memory backends, we currently forcefully wipe the memory when starting Auto-GPT. To ingest data with those memory backends, you can call the data_ingestion.py script anytime during an Auto-GPT run.

Memories will be available to the AI immediately as they are ingested, even if ingested while Auto-GPT is running.