Add Memory to Conversational Workflows with Window Buffer Memory

By default, each LLM call in n8n is stateless. The Window Buffer Memory sub-node adds multi-turn memory by keeping a sliding window of recent messages and injecting them into every new prompt. Attach it to an AI Agent node, set a Context Window Length, and give each conversation a stable Session ID so separate threads keep independent history.

What is Window Buffer Memory in n8n?

By default, each LLM call in n8n is stateless -- the model has no awareness of previous messages in the conversation. The Window Buffer Memory sub-node fixes this by maintaining a sliding window of recent messages that get injected into each new prompt, enabling multi-turn conversations.

Real-world example: A Slack bot that remembers the last 10 messages in a conversation thread so users can ask follow-up questions without repeating context.

How do you configure Window Buffer Memory on an AI Agent node?

AI Agent node with memory configuration:

Setting	Value
Agent Type	Tools Agent
Memory Type	Window Buffer Memory
Context Window Length	10 (messages)
Session ID	{{ $json.thread_ts }}

The Session ID is critical -- it determines how conversations are tracked. Using Slack's thread_ts means each thread maintains its own independent memory.

// Conversation flow example:

User: "What were our Q1 sales numbers?"
AI: "Q1 sales totaled $2.3M, up 15% from Q4..."

User: "How does that compare to the same quarter last year?"
AI: "Compared to Q1 last year ($1.8M), this represents a 28% YoY increase..."
// ↑ The model knows "that" refers to Q1 sales because of conversation memory

User: "Break it down by region"
AI: "Here's the Q1 regional breakdown: North America $1.1M, EMEA $720K..."
// ↑ Still maintains context about Q1 sales discussion

What's the difference between Window, Summary, and Token Buffer Memory?

Warning: Memory and Cost. Each message in the buffer is sent as part of the prompt, consuming tokens. A 10-message window with long messages can add significant token cost. For cost-sensitive workflows, reduce the window to 5 or use the Summary Buffer Memory sub-node, which compresses older messages into a summary.

Window Buffer Memory:   [msg1, msg2, ..., msg10]  → All sent verbatim
Summary Buffer Memory:  [summary_of_msg1-8, msg9, msg10]  → Compressed
Token Buffer Memory:    [as many messages as fit in N tokens] → Token-capped

Memory transforms one-shot Q&A into genuine conversational assistants that feel natural to interact with.

Related: Use Manual Trigger During Development Instead of Webhook or Schedule · Flatten Deeply Nested API Responses