Chat Memory

Lets an agent remember the conversation, across runs or within one run.

What it does

Chat Memory gives an agent a conversation history. Plug it into the Memory slot of a Tools Agent, a RAG Agent or a Chat Trigger. Before calling the model, the agent reads the earlier messages from the memory; after answering, it saves your message and its reply.

Conversations are stored in your browser, on this device. Every saved conversation is listed on the Memory page, where you can read, search, export or clear it.

• With a Chat Trigger, connect the same Chat Memory to the trigger and to its agent. The chat window then shows the history and reopening it continues the conversation. With This run only, or a Key that is an expression, the conversation isn’t known before a run, so the chat window keeps no history.
• When you publish or share the workflow, the picked Conversation is cleared (it only exists on your device). Whoever installs it gets a fresh, empty conversation. Keys are kept.
• When you delete a workflow, AWFlow offers to delete its own conversations. Shared keys, and conversations another workflow uses, are kept.

It replaces Local Memory, Persistent Chat Memory and Window Buffer Memory. Workflows that use them keep working and are turned into Chat Memory when you save them:

Old node	Becomes Chat Memory with
Local Memory	Conversation storage, the same conversation.
Persistent Chat Memory	Key storage, the same key, Shared scope.
Window Buffer Memory	This run only, Last N messages with N = 2 × the old window size.

When to use it

• A chatbot built with a Chat Trigger.
• An agent that runs again and again (on a schedule, on page load) and should remember earlier runs.
• One conversation per user or per thread, with a key such as {{ $input.userId }}.
• An agent that calls several tools in one run and should remember the exchange only for that run.

Inputs and settings

6 choices · pick storage and history sent to the model to see its settings

Conversation

As much as fitsLast N messages

Key

As much as fitsLast N messages

This run only

As much as fitsLast N messages

Settings for Conversation · As much as fits

Setting	Type	What it does
Conversation	Memory picker	The saved conversation to use. Leave empty to give this node its own conversation in this workflow.
Summarise older messages	Toggle	Messages that no longer fit are folded into a short summary the model still sees. Costs an extra model call now and then. Default: On
Remember tool results	Toggle	Keep what the agent’s tools returned with each reply, so later turns can use it without calling the tools again. Default: On

Showing settings for Conversation · As much as fits

The fields change with Storage and History sent to the model.

Storage decides where the conversation is kept:

Storage	The conversation…	Good for
Conversation (default)	is one saved conversation: the one you pick in Conversation, or this node’s own when you pick none.	A single chatbot, a recurring agent.
Key	is named by Key. Every run with the same key continues the same conversation.	One conversation per user, per ticket, per page.
This run only	lives only while the workflow runs. Nothing is saved.	Multi-step reasoning inside one run.

Key scope (with Key): This workflow (default) keeps the key private to this workflow, so two workflows using inbox don’t mix. Shared lets several workflows continue the same conversation.

History sent to the model:

• As much as fits (default): the newest messages that fit the connected model’s context window, keeping room for the reply.
• Last N messages: at most Messages messages (20 by default; a question and its answer are two).

Summarise older messages (on by default): messages that no longer fit are folded into a short summary the model still sees. It costs an extra model call now and then.

Remember tool results (on by default): the agent’s tool calls and results are saved with its replies, so a later message can refer to them (“what was the price you found?”) without running the tools again.

Outputs

This node has no output port. It plugs into a slot on another node instead of passing items along.

• Memory slot of Chat Trigger, Tools Agent, RAG Agent

Chat Memory has no output. It hands its memory to the node it is connected to.

Dependencies and credentials

None. This node needs no connection and has no slots for other nodes.

Example workflow

1. A Chat Trigger with Assistant name Shopping helper.
2. A Tools Agent with Prompt {{ $input.chatInput }}, a chat model such as Chat OpenAI in its Model slot, and Web Search in its Tools slot.
3. One Chat Memory, Storage Conversation, connected to both the Chat Trigger’s and the agent’s Memory slot.
4. Chat with it: ask for a product, then “which one was cheapest?”. The agent remembers the earlier answer and its search results.

Common issues

• The agent forgets between runs. Storage is This run only, which forgets everything when the run ends. Use Conversation or Key.
• “Pick a conversation, or save the workflow to give this memory its own.” With Conversation storage and none picked, the node needs a saved workflow. Save it, or pick a conversation.
• “Conversation Key is required”. Storage is Key but Key is empty, or its expression gave nothing. Fill it in.
• “This workflow is not saved yet, so the key cannot be scoped to it: memory lasts for this run only.” Save the workflow so a This workflow key can be kept.
• The chat window stays empty. The trigger and the agent must use the same Chat Memory node, with Conversation storage or a fixed Key.

Related nodes

• Chat Trigger: a chat window that runs your workflow.
• Tools Agent and RAG Agent: the agents that accept it.
• Memory page: read, slim, export or clear saved conversations.
• Memory & Context: how memory and context windows work.