Memory Layer

Operations

Memory Layer is local-first, but it still has operational concerns: service health, database availability, backups, logs, privacy, and upgrade safety.

Daily commands

memory health                          # backend service
memory doctor                          # config and dependencies
memory status --project <project-slug> # combined service, watcher, MCP, doctor

Operations checklist

Before making changes to a running instance:

  1. Know which service instance you are operating.
  2. Check database connectivity before blaming retrieval.
  3. Back up state before upgrades or migrations.
  4. Redact secrets before sharing logs.
  5. Confirm agent and MCP integrations are scoped to the intended project.

Service

The service runs as a systemd unit on Linux or a launchd agent on macOS. Use standard OS tools to start, stop, and inspect it. See Service setup for initial configuration.

Database

Memory Layer stores everything in PostgreSQL with pgvector. See PostgreSQL and pgvector for setup.

Backups and restore

What to back up

  • PostgreSQL database.
  • Global configuration.
  • Repo-local .mem/ project config.
  • .agents/ integration files if they are part of the project workflow.

Basic database backup

pg_dump "$DATABASE_URL" > memory-layer-backup.sql
psql "$DATABASE_URL" < memory-layer-backup.sql

For production-grade backups, prefer custom-format dumps, checksums, restore tests, and a retention policy. Verify a restore by running memory doctor against a disposable database.

Do not commit secrets, API keys, local database URLs, or runtime state.

Security and privacy

Memory Layer can be local-first, but external embedding or LLM providers may receive text depending on your configuration.

Check before enabling integrations

  • Which database stores project memory?
  • Which LLM provider is configured?
  • Which embedding provider is active?
  • Is MCP HTTP local and token-protected?
  • What does the watcher capture?
  • Are logs redacted before sharing?

Do not expose the MCP HTTP server to the public internet unless you have added an authentication and network control layer.

Multi-project use

One backend serves multiple projects. Each project has its own slug, config, watchers, and scoped queries. Use memory status --project <slug> to inspect each one independently.

Next

Read Reference, Troubleshooting, or MCP.

On this page