fly.toml.example

# Fly.io deployment template for mnemon.
#
# Setup:
#   cp fly.toml.example fly.toml
#   # Edit the three REPLACE_ME placeholders below, then:
#   fly launch --copy-config --no-deploy   # creates the app, don't deploy yet
#   fly volume create mnemon_data --size 1 --region <your-region>
#   fly secrets set MNEMON_LOCAL_TOKEN=... MNEMON_AS_ENABLED=true MNEMON_AS_PASSPHRASE=...
#   fly deploy
#
# See the "Self-host on Fly.io" section of README.md for the full runbook.

app = "REPLACE_ME_fly_app_name"      # must be globally unique on Fly
primary_region = "sjc"                # pick one close to you: fly platform regions

[build]

[env]
  PORT = "8080"
  MNEMON_VAULT_DIR = "/data"
  MNEMON_PUBLIC_URL = "https://REPLACE_ME_fly_app_name.fly.dev"
  MNEMON_ALLOWED_HOSTS = "REPLACE_ME_fly_app_name.fly.dev"

[http_service]
  internal_port = 8080
  force_https = true
  auto_stop_machines = "stop"
  auto_start_machines = true
  min_machines_running = 0

  # First request after a cold start loads the FastEmbed bge-small-en-v1.5
  # ONNX model (~15–25s). grace_period must be longer than that to avoid
  # false-positive "app not listening on 0.0.0.0:8080" warnings during
  # rolling deploys and `fly secrets set` restarts.
  [[http_service.checks]]
    interval = "30s"
    timeout = "5s"
    grace_period = "40s"
    method = "GET"
    path = "/health"

[mounts]
  source = "mnemon_data"
  destination = "/data"

# Memory must be at least 1GB because the first memory_search call loads
# the FastEmbed bge-small-en-v1.5 ONNX model (~130MB resident) into RAM
# on top of Python, uvicorn, FastMCP, and the vector store. The Fly
# default (shared-cpu-1x:256mb) OOM-kills the mnemon process on first
# search.
[[vm]]
  memory = "1gb"
  cpu_kind = "shared"
  cpus = 1