AI memory for agents · backup and recovery

Keep the knowledge. Carry it forward.

We turn your agents’ discoveries into lasting memory—and give you a portable copy of the records you choose. Export useful knowledge, restore it in another Claude-Mem installation and verify that your agents can find it again.

Claude-Mem carrying a saved project note

AI memory in production

70 days of work. One searchable memory.

Alessandro used our open-source engine across three machines for 70 days. His customized deployment captured the decisions, fixes and discoveries behind the work—and published the data.

Observations captured
22,718
Bug fixes recorded
1,262
Decisions recorded
757

See how he made that knowledge searchable. Then give your own agents a memory to build on.

Claude-Mem’s mascot files a note in a shared card catalogue connected to three computers.
Alessandro’s published field study · March–May 2026

Reviewed November 19, 2026 · Native memory backup and recovery

What is in the copy?

Saved memory records

Native JSON export includes matching observations, session metadata, summaries and prompts. The file records export time, query, project scope and counts so you can inspect what was selected.

A separate work-state export

The local worker also documents an export for ordered work-state history. That export is not a work-state import or synchronization format. Keep its purpose separate from the native memory import script.

Your source files stay separate

A memory export is not a full disk or project backup. Keep source code, documents, media, configuration and credentials in their own backup process. Do not assume that importing notes restores the files they describe.

Sync has a different job

Cloud Sync helps connected clients access eligible memory records. Do not assume it provides a separate historical snapshot, restores every local file, or replaces a tested recovery copy.

A backup is useful when you can recover it.

  1. Choose the records to keep. A topical export selects matching records. An empty query exports all saved history within the selected project, or across projects when no project filter is supplied. Check the project name before exporting.
  2. Keep the export somewhere you control. The JSON contains readable memory content, including prompts and potentially private paths. Review its scope and use storage with access controls appropriate for your project. Keep a separate copy from the installation you are protecting.
  3. Restore into a separate installation first. Use the native import script with a running destination worker. Preserve the original installation and export while checking the result. This guide does not overwrite or reset your database for you.
  4. Read the result, including rejected rows. The importer distinguishes imported, already-present and rejected records. Rejected rows or request failures need investigation, even if other records were imported. Re-running an import can skip records already present; it is not proof that every record was restored.
  5. Search for a known decision. Find a distinctive record from the export at the destination, inspect its contents and project identity, then compare the imported counts and any rejection reasons. Check more than one record type when your export includes observations, summaries and prompts.

Start with one project.

Run these repository scripts from a current Claude-Mem source checkout with dependencies installed. The source worker must be running for export; the destination worker must be running for import. Replace my-app with your actual project name.

Source checkout

npx tsx scripts/export-memories.ts "" project-memory.json --project=my-app

Destination checkout

npx tsx scripts/import-memories.ts project-memory.json

Use a new output filename for each copy you want to retain. Keep the export until you have checked the destination records and any rejected rows.

For large exports, use matching script and worker versions. The importer splits the file into requests within the worker’s 5 MiB limit and resolves destination session identities before importing child records. If a later request fails, earlier accepted records remain. Read imported, skipped and rejected counts, fix the reported issue and rerun the same file; already restored records are skipped.

Check current requirements and output fields

What to check before calling it restored

  • The export has the expected project, query, record types and counts.
  • The destination reports imported, skipped and rejected records separately.
  • A search finds the records you selected and their contents match the saved copy.
  • Any source files those records refer to are available independently.
  • The next agent can retrieve the relevant history with its configured tools.

Backup questions

What happens with a large export?

We send the records in bounded requests, with session records first. Use matching script and worker versions. A record too large for one request is rejected without shortening it. Check the result and any rejection reasons; a later failure does not undo records already accepted.

Is this an automatic scheduled backup service?

No. This page describes the native export/import workflow. Set your own retention and storage process; no backup schedule is configured by reading this guide.

Will import replace an existing session?

Existing destination session metadata can be retained while new observations and summaries attach to that session. Import is not a byte-for-byte overwrite of a previous installation. Review the current documentation and verify the result.

Can this recover deleted project files?

No. Saved memory may describe a file or decision, but the memory export is not a copy of arbitrary project files. Recover those from your separate source or file backup.

Can I use this to import Mem0 or Supermemory data?

This workflow uses Claude-Mem’s native JSON format. A different provider’s export is not interchangeable with it.

Give your next installation a head start.

Bring the decisions, fixes and discoveries you have already captured. Keep a portable copy, restore the selected records and put that knowledge back within reach.