Skip to main content

​
Operations

​
Services in Docker Compose

The default docker-compose.yml starts:
  • api
  • worker
  • beat
  • postgres
  • redis
  • neo4j
It reads configuration from .env. Container startup also waits for Postgres and runs Alembic automatically before launching the service process.

​
Ports

  • API: 8000
  • Postgres: 5432
  • Redis: 6379
  • Neo4j HTTP: 7474
  • Neo4j Bolt: 7687

​
Celery schedules

The Celery beat schedule includes:
  • hourly_cortex
    • every hour at minute 0
  • sync_graph_projection
    • every 300 seconds

​
Graph projection behavior

Canonical graph mutations are written to PostgreSQL and recorded in the outbox. After the canonical commit succeeds, aCMF immediately dispatches an async graph projection task for the newly created outbox events. The graph sync worker then projects:
  • memory nodes
  • entities
  • relations
  • memory links
  • general graph edges
into Neo4j. Beat still runs the periodic sync as a backlog drain and retry path. Canonical writes do not depend on Neo4j being current.

​
Tests

Run: 
python3 -m pytest

​
Process smoke test

To send 10 varied /v1/process requests against a running API: 
python3 scripts/process_smoke_test.py --base-url http://localhost:8000 --pretty
Useful flags:
  • --prefix demo to avoid reusing the default ids
  • --timeout 30 to increase request timeout

​
Useful sanity check

python3 -c 'from app.main import app; print(app.title, app.version)'

​
Important current behavior

  • /v1/process returns immediately after enqueuing.
  • Read paths can abstain when evidence is weak or irrelevant.
  • Read-path diagnostics are relevance-gated and do not expose unrelated raw retrieval hits.
  • Snapshots only include user and global memories.
  • Neo4j is projection-only and can be rebuilt from canonical state.