MemorySync Chat Memory: a community node that keeps n8n AI Agent conversations across restarts

I work on MemorySync, and we built a community node package for n8n: n8n-nodes-memorysync. It installs two nodes:

  • MemorySync Chat Memory plugs into an AI Agent’s Memory port. The conversation history is stored in MemorySync instead of process memory, so sessions survive restarts, redeploys and weeks between chats. Durable facts are extracted from the user’s turns and stay available as that user’s memories.
  • MemorySync gives workflows six operations: Add Memory, Add Conversation Turn, Search Memories, Recall Context, Get Many and Delete Memories. Every operation is scoped by End User ID, and you can also hand the node to an agent as a tool.

Install (self-hosted, instance owner): Settings → Community Nodes → Install → n8n-nodes-memorysync. The Chat Memory sub-node needs n8n 2.16.0 or later. On n8n Cloud the node isn’t available yet; MemorySync says verification is in progress.

Guide: n8n Workflow & Agent Memory | MemorySync Docs
Step-by-step walkthrough (Tips & Tricks): How to give an n8n AI agent long-term memory

Happy to answer questions or take feedback on the node.

Interesting project! Persistent memory is useful, especially when AI agents need to maintain context across restarts and long gaps between conversations.

I’m exploring reliable n8n workflows with FastAPI and PostgreSQL, with a focus on idempotency, error handling, and recovery.

I’m curious about two things:

  1. How does MemorySync handle concurrent updates when the same End User ID has multiple active conversations?

  2. What safeguards are available to prevent one user’s memories from being retrieved in another user’s session?

Also, is there a recommended approach for backing up or exporting memories if a workflow needs to migrate to another environment?

Thanks for sharing!

For the durable facts extracted from user turns, does recall stop presenting an older fact as current when the user corrects it? For example, “I live in Berlin” followed later by “Actually, I live in Madrid, not Berlin.”

Thanks, both, great questions.

@Shital_Katre

  • Export/backup: there’s a per-user export endpoint, so you can pull a user’s memories out whenever you want (see the export docs in the MemorySync docs). You can also forget memories by ID or filter (with a dry run first), purge one end user completely (see the purge docs in the MemorySync docs), and deleted items go to a trash you can recover from. Your data isn’t locked in.
  • Isolation: every call has to name an end user (X-End-User-ID), or the API rejects it. Reads, searches and deletes are filtered to that end user, project and environment, so a query for User A only searches User A’s memories (see the scopes in the MemorySync docs). One tip: don’t use the session ID as an access boundary, and set the end-user ID on your server, not from the browser.
  • Concurrent updates: quota checks and memory tier changes are atomic, and near-duplicate writes get deduplicated. In n8n, the Chat Memory node gives each turn a deterministic idempotency seed, so a retried execution doesn’t store the same turn twice (see the n8n guide in the MemorySync docs).

@Ricart_Juncadella Yes, for one-value facts like where someone lives, their job title or a preference. “Actually, I live in Madrid” replaces the Berlin fact when it’s stored (see the FAQ in the MemorySync docs). Facts that can be true side by side are kept with timestamps, and you can forget the old one explicitly if you want. If you try it with your Berlin/Madrid case, I’d love to hear how it behaves in your workflow.

Thanks for the detailed explanation! The server-side end-user ID requirement and scoped access controls are especially important for secure multi-user workflows.

I’d be interested to know whether the export endpoint supports bulk or automated scheduled exports, and whether the n8n idempotency seed can be configured or inspected for debugging retries.

Appreciate the clarification!

Glad it helps, @Shital_Katre!

Scheduled exports: the export endpoint is per end user, and I’m not aware of a built-in scheduler on our side. The easy route is to use n8n itself: a Schedule Trigger, then a loop over your end-user IDs, then an HTTP Request to the export endpoint for each one, then write the results to S3, Drive or wherever you keep backups. That gives you automated backups on whatever schedule you like.

Idempotency seed: it isn’t configurable. It’s derived deterministically from the turn’s role, session and content, so the same turn always gets the same seed. That’s what makes retries safe. To debug it, look at the add step’s output: a replayed turn comes back with alreadyExists: true instead of writing a second copy. The node is open on npm and GitHub (n8n-nodes-memorysync), so you can read exactly how the seed is built.

There’s more detail in the troubleshooting section of the n8n guide in our docs.