Memori BYODB vs MemoryRouter: Who Runs the Database?

John Rood··3 min read

Memori's memori.config.storage.build() initializes memory storage in the database your application supplies. A successful recall test proves the save-and-retrieve path. It does not prove those new memory tables are included in your backup or survive your restore procedure.

MemoryRouter's hosted vault takes database setup out of connecting your existing assistants. With Memori BYODB, your team operates the memory database. With MemoryRouter, supported clients connect to a service we operate.

What you take on with Memori BYODB

Memori's BYODB documentation shows a database connection factory passed into Memori and an LLM client registered with the SDK. Its quickstart uses SQLite. The documented database options also include MySQL and PostgreSQL.

Your memory records live in a database you choose. The SDK captures interactions from the registered client rather than asking your application to save each message manually.

Your team owns schema rollout and database recovery. If the rest of your application already uses that database, memory can fit into the same access and recovery process. The new tables still need to be part of that process.

Memori also has a hosted Cloud product that removes database configuration. This comparison concerns its BYODB route, not every Memori deployment.

What you connect with MemoryRouter

I built MemoryRouter for continuity across the AI tools you use. Its hosted MCP integration gives supported clients access to a vault chosen during OAuth sign-in. Context saved through one authorized client is available to another using that vault.

You do not supply a SQL connection or apply a memory schema. Saved memories go to MemoryRouter's servers; this is not local storage.

The MCP setup and cross-session proof separate connecting from capturing. A generic MCP connection exposes search and save tools. The model decides when to call them, and the host may require approval for a write. Connecting does not import old conversations. It also does not reproduce Memori's automatic capture through a registered LLM client.

A recovery check before real history

The synthetic decision for this proposed check is: "The Harbor demo exports newline-delimited JSON because the downstream job reads one record at a time."

For Memori BYODB, the acceptance sequence is:

  1. Use a disposable database and a test entity. Let the registered LLM client capture a conversation containing that decision. Confirm it becomes retrievable.
  2. Stop the application and take a consistent database backup using your database's procedure. Restore it into a separate disposable database, not over the original.
  3. Point the application at the restored database. Use the same test entity in a new conversation, without the old chat. Ask what Harbor exports and why.
  4. Inspect the retrieved memory for both the format and the reason. If they are absent, investigate the backup contents and retrieval path before adding real history.

For MemoryRouter, the corresponding client-side check is a save through a writable connection, followed by retrieval in a new conversation connected to that same vault. Check the actual search result for the format and reason, rather than accepting a plausible model answer. That checks the connected workflow. It does not test MemoryRouter's infrastructure recovery.

Create a MemoryRouter account to save your first project decision in a shared vault.