CLAUDE.md Migration

Use memory_store:
- title: "pgvector HNSW non-transactional behaviour"
- memory_type: "incident"
- content: "pgvector's HNSW index is non-transactional. Rolling back a transaction does NOT undo index updates. If a test inserts a vector and the transaction rolls back, the vector remains in the HNSW index even though it's gone from the table. Always use DELETE or TRUNCATE for test cleanup — not transaction rollback. This affects any code that runs tests against a shared pgvector index."
- concepts: ["pgvector", "HNSW", "transactions", "testing", "rollback"]
- files: ["api/neuroloom_api/"]
- importance: 0.95
- tags: ["gotcha", "testing", "pgvector"]

Use memory_store:
- title: "SQLAlchemy workspace isolation requirement"
- memory_type: "convention"
- content: "Every database query must filter by workspace_id. No exceptions. This is a security invariant — not a performance optimization. pgvector cosine similarity queries do not enforce tenant isolation through the index; the WHERE workspace_id = :workspace_id clause is the only thing preventing cross-tenant data access. Any new query or endpoint that omits workspace_id is a security bug."
- concepts: ["workspace isolation", "multi-tenancy", "security", "SQLAlchemy", "pgvector"]
- files: ["api/neuroloom_api/"]
- importance: 1.0
- tags: ["security", "invariant"]

Use memory_store:
- title: "ARQ job serialization constraint"
- memory_type: "pattern"
- content: "ARQ background jobs must receive only JSON-serializable arguments. Never pass SQLAlchemy ORM objects to job functions — pass IDs as strings and reload from DB inside the job. The ARQ worker runs in a separate process with its own database connection pool and cannot deserialize ORM model instances from another process."
- concepts: ["ARQ", "background jobs", "serialization", "async"]
- files: ["api/neuroloom_api/workers/"]
- importance: 0.88
- tags: ["pattern", "arq"]

import os
import httpx

token = os.environ["MEMORIES_API_TOKEN"]
workspace_id = os.environ["MEMORIES_WORKSPACE_ID"]

# Memories to migrate from CLAUDE.md
migrations = [
    {
        "title": "pgvector HNSW non-transactional behaviour",
        "memory_type": "incident",
        "narrative": (
            "pgvector's HNSW index is non-transactional. Rolling back a transaction "
            "does NOT undo index updates. Always use DELETE or TRUNCATE for test cleanup — "
            "not transaction rollback. This affects all code that runs tests against a "
            "shared pgvector index."
        ),
        "concepts": ["pgvector", "HNSW", "transactions", "testing"],
        "source_files": ["api/neuroloom_api/"],
        "importance_score": 0.95,
        "tags": ["gotcha", "testing", "pgvector"],
    },
    {
        "title": "SQLAlchemy workspace isolation requirement",
        "memory_type": "convention",
        "narrative": (
            "Every database query must filter by workspace_id. No exceptions. "
            "pgvector similarity queries do not enforce tenant isolation through the index — "
            "the WHERE workspace_id = :workspace_id clause is the only guard against "
            "cross-tenant data access. Any query that omits workspace_id is a security bug."
        ),
        "concepts": ["workspace isolation", "multi-tenancy", "security", "SQLAlchemy"],
        "source_files": ["api/neuroloom_api/"],
        "importance_score": 1.0,
        "tags": ["security", "invariant"],
    },
    {
        "title": "ARQ job serialization constraint",
        "memory_type": "pattern",
        "narrative": (
            "ARQ background jobs must receive only JSON-serializable arguments. "
            "Never pass SQLAlchemy ORM objects to job functions — pass IDs as strings "
            "and reload from DB inside the job. The ARQ worker runs in a separate process "
            "with its own connection pool."
        ),
        "concepts": ["ARQ", "background jobs", "serialization", "async"],
        "source_files": ["api/neuroloom_api/workers/"],
        "importance_score": 0.88,
        "tags": ["pattern", "arq"],
    },
]

for memory in migrations:
    response = httpx.post(
        "https://api.neuroloom.dev/api/v1/memories/store",
        headers={"Authorization": f"Token {token}"},
        json={"workspace_id": workspace_id, **memory},
    )
    result = response.json()
    print(f"Stored: {result['id']}  {result['title']}")

curl -X POST https://api.neuroloom.dev/api/v1/memories/store \
  -H "Authorization: Token $MEMORIES_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "workspace_id": "'"$MEMORIES_WORKSPACE_ID"'",
    "title": "pgvector HNSW non-transactional behaviour",
    "memory_type": "incident",
    "narrative": "pgvector HNSW index is non-transactional. Rollback does NOT undo index updates. Use DELETE/TRUNCATE for test cleanup, not transaction rollback.",
    "concepts": ["pgvector", "HNSW", "transactions", "testing"],
    "source_files": ["api/neuroloom_api/"],
    "importance_score": 0.95,
    "tags": ["gotcha", "testing", "pgvector"]
  }'

Use memory_search with query "what happens when I roll back a pgvector transaction in tests"

Use memory_search with query "how do I avoid cross-tenant data in queries"

Use memory_search with query "how to pass data to background jobs"

for query in \
  "what happens when I roll back a pgvector transaction in tests" \
  "how do I avoid cross-tenant data in queries" \
  "how to pass data to background jobs"; do
  echo "Query: $query"
  curl -s -X POST https://api.neuroloom.dev/api/v1/memories/search \
    -H "Authorization: Token $MEMORIES_API_TOKEN" \
    -H "Content-Type: application/json" \
    -d "{
      \"workspace_id\": \"$MEMORIES_WORKSPACE_ID\",
      \"query\": \"$query\",
      \"limit\": 3
    }" | python3 -c "
import sys, json
data = json.load(sys.stdin)
for r in data['results']:
    print(f'  {r[\"score\"]:.2f}  {r[\"title\"]}')
"
  echo ""
done