Persistent Memory in
n8n AI Workflows

n8n's AI Agent node is powerful — but every workflow run starts with a blank slate. Here's how to connect it to Kronvex so your agents remember users, past interactions, and accumulated knowledge across every run.

Why n8n agents forget

n8n's built-in "Window Buffer Memory" node keeps context within a single workflow execution. The moment the run ends — triggered by a webhook, a cron job, or manually — the memory is gone. The next run knows nothing about previous conversations.

For automations that interact with the same user over time (support bots, sales follow-ups, onboarding sequences), this is a fundamental limitation. Kronvex solves it with three HTTP Request nodes that take under 5 minutes to set up.

Prerequisites

• An n8n instance (cloud or self-hosted)
• A Kronvex account — get a free API key here
• Your Kronvex Agent ID (shown in the dashboard after creating an agent)

Store your API key as an n8n credential: Credentials → New → Header Auth, Name: Kronvex, Header Name: X-API-Key, Value: kv-your-key.

Node 1 — Store a memory

After your AI generates a response (or at any point where you want to store context), add an HTTP Request node:

{
  "content": "{{ $json.userMessage }}",
  "memory_type": "episodic",
  "session_id": "{{ $json.userId }}"
}

The session_id field scopes memories per user — critical for multi-user workflows. Use any stable user identifier: email, user ID, phone number.

Memory types:

• episodic — what happened ("User said they prefer email over calls")
• semantic — facts ("User is a freelance designer, 3 years experience")
• procedural — patterns ("User always responds within 2 hours")

Node 2 — Recall memories

At the start of your workflow (before the AI node), add a recall node to load relevant context:

{
  "query": "{{ $json.userMessage }}",
  "top_k": 5,
  "session_id": "{{ $json.userId }}"
}

The response contains a memories array. Each memory has content, memory_type, and a score (0–1 relevance). Access them in later nodes with {{ $json.memories[0].content }}.

Node 3 — Inject context (recommended)

Instead of manually formatting memories, use the /inject-context endpoint which returns a ready-to-use context string formatted for LLM prompts:

{
  "query": "{{ $json.userMessage }}",
  "top_k": 5,
  "session_id": "{{ $json.userId }}"
}

The response has a context string ready to prepend to your system prompt. In your AI Agent node's system message:

You are a helpful assistant.

{{ $('Inject Context').item.json.context }}

Answer the user's question based on the context above.

Full workflow pattern

Here's the recommended node order for a webhook-triggered AI workflow:

Production tips

• Always pass session_id — without it, all users share the same memory pool
• Store selectively — only remember information that matters long-term. Don't store every word; summarize instead
• Use TTL for temporary context — add "ttl_days": 30 to memories that should expire (e.g., support ticket context)
• Error handling — add a Continue On Fail on the /recall node so a memory outage never breaks your workflow
• Rate the n8n HTTP credential — set up a dedicated Kronvex agent per n8n environment (dev/staging/prod)
• GDPR — all data stays in EU (Frankfurt). For GDPR deletion requests, use DELETE /agents/{id}/memories to wipe a user's session