Backup & restore

novamem's source of truth is Postgres — including the memory_relations co-occurrence edges, which ride along with every Postgres backup. Only the cold-tier (Qdrant) lives outside it, and it is reconstructible from the warm-tier embeddings.

What to back up

Store	Importance	Recovery
Postgres	Critical	Required for any restore. Holds memory_entries, memory_relations, users, sessions, tokens, projects, audit log.
Qdrant	Optional	Re-index from warm-tier entries via /v1/admin/reindex (planned). Until then, snapshot and restore.

In practice, back up both. Postgres restoration without Qdrant works (search degrades to keyword-only with degraded: true) but you want both for full fidelity.

Postgres

pg_dump for full daily snapshots

PGPASSWORD=$POSTGRES_PASSWORD pg_dump \
  -h postgres -U novamem -d novamem \
  --format=custom --compress=9 \
  --file novamem-$(date +%F).dump

Restore:

PGPASSWORD=$POSTGRES_PASSWORD pg_restore \
  -h postgres -U novamem -d novamem --clean --if-exists \
  novamem-2026-05-05.dump

WAL archive for point-in-time recovery

If you need PITR (e.g. recover from accidental delete to 5 minutes before), enable continuous archiving:

# postgresql.conf
wal_level = replica
archive_mode = on
archive_command = 'aws s3 cp %p s3://my-bucket/wal/%f'

Combine with a base backup via pg_basebackup and you can restore to any LSN.

Qdrant

# Snapshot a collection
curl -X POST http://qdrant:6333/collections/novamem_acme_default/snapshots

# Download
curl http://qdrant:6333/collections/novamem_acme_default/snapshots/<name> > snap.tgz

# Restore
curl -X PUT http://qdrant:6333/collections/novamem_acme_default/snapshots/upload \
  -F snapshot=@snap.tgz

Snapshot every collection (one per user/project scope × namespace). If your tenant count is large, automate via the Qdrant snapshot API.

Relations (memory_relations)

Nothing extra to do: the co-occurrence edges live in the memory_relations Postgres table (bitemporal valid_from/valid_to), so pg_dump / WAL archiving above covers them. Edges are also self-healing — new remember calls write fresh vector-neighbour edges via the async reconciler.

Disaster scenarios

Scenario	Recovery procedure
Postgres lost (no backup)	Total memory loss — entries and relations alike. Don't run without a backup strategy.
Postgres corrupt	Restore latest dump → restart novamem → cold-tier still has older vectors but new writes are coherent.
Qdrant lost	Restart with empty Qdrant → /v1/admin/reindex (planned) re-embeds every warm entry from content. Slow but lossless.
Whole cluster lost	Restore Postgres dump → Qdrant snapshots → start novamem. Verify with GET /health.

Operational tip

Time-bound your backup tests. A backup that hasn't been restored is hypothetical. Schedule a quarterly drill: restore yesterday's dump into a sandbox cluster, run /health, run a few memory_search calls, confirm signal scores match expectations.