Install the server

Both methods run the same server: a web interface at /, the MCP endpoint at /mcp/ and a health check at /healthz, all on one port. Keep that port on 127.0.0.1 and publish it through an HTTPS reverse proxy.

New in 14.6.0 tam setup --mode company can prepare the server for you (data directory, address, service or Compose, first admin, departments, providers). A server started without an administrator prints a one-time setup code and runs a web setup wizard at /dashboard/. See Setup wizard.

Option A: Docker Compose

From a checkout of the repository, with port 3738 free on the host:

git clone https://github.com/vbcherepanov/total-agent-memory.git
cd total-agent-memory
docker compose -f docker-compose.team.yml build
docker compose -f docker-compose.team.yml up -d

What the profile sets up:

• Container team-memory runs tam-team serve inside the container on port 3737, published on the host as 127.0.0.1:3738.
• Data lives in the Docker volume tam-team-memory-data (mounted at /team-data), model files in tam-team-model-cache. Both survive container rebuilds.
• A health check calls /healthz every 30 seconds; the container restarts unless stopped.
• It does not mount a personal single-user memory volume. Do not point it at an existing personal database; migrating one needs an owner check and a backup first.

Settings go in a .env file next to the compose file:

TAM_TEAM_PORT=3738
TAM_TEAM_PUBLISH_HOST=127.0.0.1
TAM_TEAM_MAX_WORKERS=3
TAM_TEAM_OPERATION_TIMEOUT=120
MEMORY_LLM_ENABLED=false

To run admin commands, use the same image with run --rm:

docker compose -f docker-compose.team.yml run --rm team-memory \
  python /app/src/team_memory/cli.py user-add alice 'Alice Moreau'

The next page, Users, teams & tokens, lists every command.

Option B: Python package

On Linux or macOS, in a dedicated virtual environment:

python3 -m venv /opt/tam-team/.venv
/opt/tam-team/.venv/bin/python -m pip install total-agent-memory
/opt/tam-team/.venv/bin/tam-team --root /srv/tam-team user-add alice 'Alice Moreau'
/opt/tam-team/.venv/bin/tam-team --root /srv/tam-team serve --host 127.0.0.1 --port 3738

On Windows PowerShell:

py -3 -m venv .venv
.\.venv\Scripts\python.exe -m pip install total-agent-memory
$env:PATH = "$PWD\.venv\Scripts;$env:PATH"
tam-team --root .\team-data serve --host 127.0.0.1 --port 3738

• --root is the data directory. Without it, tam-team uses $TAM_TEAM_DIR or ~/.tam-server. It is created with owner-only permissions.
• serve defaults to 127.0.0.1:3737 (or MCP_HTTP_HOST / MCP_HTTP_PORT). The examples use 3738 to match the Docker profile.
• From a source checkout, the same CLI is PYTHONPATH=src python -m team_memory.cli ….

The repository does not ship a systemd unit for the team server. Run serve under whatever process supervisor you use, with the same user that owns the data directory. Only one server can use a data directory at a time; a second one refuses to start.

Check that it is up

curl -s http://127.0.0.1:3738/healthz
# {"status":"ok","name":"total-agent-memory","version":"14.8.0","release_date":"…"}

The health response does not reveal paths or users. Then open http://127.0.0.1:3738/ in a browser on the server (or through an SSH tunnel) and sign in with a token’s contents.

LLM features on the server

Search and save work without any LLM. If you enable LLM-backed features, configure the provider on the server; it applies to all users’ content.

MEMORY_LLM_ENABLED=true
MEMORY_LLM_PROVIDER=ollama
MEMORY_LLM_MODEL=qwen2.5-coder:7b
OLLAMA_URL=http://host.docker.internal:11434

The Docker profile maps host.docker.internal to the host so a model server on the Docker host is reachable (override with TAM_TEAM_HOST_GATEWAY). For openai-compatible, also set MEMORY_LLM_API_BASE and, if needed, MEMORY_LLM_API_KEY. See Configuration.

Sizing

• Each scope that is in use gets its own worker process with the embedding model loaded. By default three stay warm (TAM_TEAM_MAX_WORKERS=3): typically one personal, one team and the shared scope.
• Plan at least 4 GiB of RAM for three warm workers and measure on your own data. On a Linux ARM64 test machine, three workers used about 2 GiB of resident memory together, two about 1.3 GiB, before the gateway, database cache and OS.
• On a small machine set TAM_TEAM_MAX_WORKERS=1 or 2. Fewer workers means more cold loads (and CPU) when people switch scopes; results are the same.
• Operations run one at a time per worker. The server accepts up to 32 requests in flight and answers 429 Server busy beyond that.