How does AntTrail work?
og-memory was renamed to AntTrail, and we will refer to it as AntTrail in the future.
Get started
To start using AntTrail, clone this repo, check out to the dev branch, and you can see the full source code.
Deploy AntTrail with Docker
The Docker deployment is based on two configuration files deploy.env and ogmemory.yaml in the deploy folder. The deploy.sh script will read these two configuration files and pull docker images from the Internet, thus the local code changes will not be applied if you deploy AntTrail this way. To make it more clear, you can choose an empty folder and run the following command to download the necessary configuration files and deployment script.
1
curl -fsSL https://raw.gitcode.com/opengauss/oGMemory/raw/dev/deploy/install.sh | bash
This script downloads an ogmem-deploy folder with three files inside deploy.env, ogmemory.yaml, ogmemory.example.yaml.
The deployment script will start 3 containers for (1) opengauss database service; (2) openclaw service; (3) AntTrail memory service. To start these containers properly, several necessary configurations need to be set.
For the opengauss database container, you need to set the following variables in deploy.env. The variables OG_HOST_PORT, OG_CONTAINER_NAME should be customized to a unique value so that your port and container name do not conflict with other users on the same machine.
1
2
3
4
5
6
ENABLE_OPENGAUSS="true"
OG_CONTAINER_NAME="opengauss_yuanjian"
OG_HOST_PORT="35432"
OG_PORT="5432"
OG_IMAGE_REPO="swr.cn-north-4.myhuaweicloud.com/kunpeng-ai/opengauss-distributed"
OPENGAUSS_HOST_IP="127.0.0.1"
As the memory system heavily relies on LLMs and embedding models, we need to configure the following variables for the models to work.
Note that by default there is no proxy network inside the container, and you need to make sure the provider is reachable.
The model configuration will be used in ogmemory.yaml.
1
2
3
4
LLM_PROVIDER="openai"
LLM_API_KEY="sk-cp-THE_REST_OF_YOUR_API_KEY"
LLM_BASE_URL="http://api.minimaxi.com/v1"
LLM_MODEL="MiniMax-M2.7"
The embedding model needs to separately be set if you have a different provider. Find the following configuration in ogmemory.yaml and set it accordingly.
1
2
3
4
5
6
embedding:
provider: volcengine
model: "doubao-embedding-vision-250615"
base_url: "https://ark.cn-beijing.volces.com/api/coding/v3"
api_key: "ark-THE_REST_OF_YOUR_API_KEY"
multimodal: true
Come back to the deploy.env file. For OpenClaw and OGMEM, we need to set their corresponding container name and images. The image selection decides the version of both services.
1
2
3
4
OGMEM_CONTAINER_NAME="ogmem_yuanjian"
OGMEM_IMAGE="swr.cn-north-4.myhuaweicloud.com/kunpeng-ai/ogmemory:poc_57"
OPENCLAW_CONTAINER_NAME="openclaw_ogmem_yuanjian"
OPENCLAW_IMAGE="swr.cn-north-4.myhuaweicloud.com/kunpeng-ai/openclaw-ogmemory:poc1_416"
There are two folders that need to be mapped into the containers for the AGFS file system and OpenClaw home directory.
1
2
AGFS_DATA_DIR="/home/yuanjian/Development/memory-projects/agfs_data"
OPENCLAW_HOME_DIR="/home/yuanjian/Development/memory-projects/openclaw_dir"
Additionally, we need to set ports for OpenClaw gateway and AntTrail’s API.
1
2
GATEWAY_PORT="34589"
OGMEM_URL="http://127.0.0.1:3666"
At this point, the deploy.env file should be ready. Now let’s navigate to the ogmemory.yaml file. We’ve already set the embedding models. There are several more settings to be dialed in.
The vector database needs a connection string and a dimension parameter. The port should be the same as OG_HOST_PORT set in deploy.env. The dimension is determined by the embedding model you choose. For instance, the doubao-embedding-vision-250615 model encodes everything into a vector of dimension 1024.
1
2
3
4
5
6
vector_db:
type: opengauss
connection_string: "host=127.0.0.1 port=35432 dbname=postgres user=gaussdb password=YoUrpassWord$%23"
dimension: 1024
table_name: vector_index
pool_size: 5
The last one is the port of AntTrail service, which should be the same as the port you dialed in for OGMEM_URL.
1
2
3
service:
http_port: 3666
workers: 2
Congratulations! We finally finished configuring everything.
We can start the whole system with the following command.
1
bash deploy.sh -password "YoUrpassWord$%23"
The password is the user password for your opengauss user gaussdb. The script will set the password so that the connection_string you set earlier works properly.
If things went wrong, you can run bash deploy.sh cleanup to remove the containers and startover.
The cleanup script will only remove the containers but not delete the folder. So your AGFS folder and OpenClaw home folder are still there. If you want to start from fresh, manually remove those folders.
Deploy AntTrail locally without containers
This part is still under investigation.
Run the LoCoMo benchmark with AntTrail
Clone this repo for locomo evaluation.
There are several environment variables need to be set. So I create a env.sh file that can be sourced later.
1
2
3
4
5
export OPENAI_API_KEY="sk-cp-THE_REST_OF_YOUR_API_KEY"
export OPENAI_BASE_URL="https://api.minimaxi.com/v1"
export OPENCLAW_STATE_DIR=/home/yuanjian/Development/memory-projects/openclaw_dir
export JUDGE_MODEL="MiniMax-M2.7"
export OPENCLAW_GATEWAY_TOKEN="ogmem-default-token"
To start the evaluation, go into the scripts folder and run the following command. Remember to source env.sh or set the environment variabels manually before running the following command.
1
bash run_eval_small.sh --import-mode claw --gateway-url http://127.0.0.1:34589 --force-ingest --gateway-token "$OPENCLAW_GATEWAY_TOKEN"
The port is GATEWAY_PORT you set in the deploy.env file.