> ## Documentation Index
> Fetch the complete documentation index at: https://docs.flowx.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Memory capabilities
> Persist conversation context across messages with automatic session memory retrieval and summarization.
## Overview
Session memory allows AI nodes in conversational workflows to access previous messages from the current conversation. When enabled, FlowX automatically retrieves conversation history and injects it into the LLM's context — giving the AI agent awareness of what was discussed earlier in the session.
Memory is **per-session** (identified by Chat Session ID) and managed entirely by FlowX. You don't need to build any memory retrieval logic — just toggle it on per node.
Latest 3 message pairs + summary of earlier exchanges retrieved on each message
Older conversation history is automatically summarized to fit within context limits
Turn on memory per Custom Agent or Intent Classification node with the **Use Memory** toggle
View retrieved memory (raw turns + summary) in the workflow console log
***
## How memory works
The Chat component sends the message + `chatSessionId` to the Chat Driven workflow.
The Start node retrieves conversation history for the session: the latest 3 user/agent message pairs in full, plus a summary of all earlier exchanges.
For each node with **Use Memory** enabled, the retrieved history is appended to the LLM system prompt, giving the AI agent context from prior turns.
After the AI generates a response, the system stores both the user message and the agent response, updating the session memory for the next turn.
When the conversation exceeds 3 turns, older messages are automatically summarized. The summary is regenerated on each new message beyond the threshold.
***
## Memory structure
The memory injected into the LLM prompt follows this structure:
```
user: [message content]
agent: [response content]
user: [message content]
agent: [response content]
user: [message content]
agent: [response content]
Summary of previous replies: [generated summary]
```
* **Recent turns** — The latest 3 user/agent message pairs are included in full
* **Summary** — A compressed summary of all conversation history before the last 3 turns
Only user messages and agent responses are included in memory. Internal workflow data, routing decisions, and intermediate node outputs are not part of the memory.
***
## Enabling memory
Memory is controlled per node using the **Use Memory** toggle. It's available on:
* **Custom Agent** nodes — for context-aware AI responses
* **Intent Classification** nodes — for more accurate classification using conversation history
**Default:** OFF
When enabled, the node sends the `chatSessionId` to the AI platform, which retrieves and attaches the conversation history to the LLM call.
Memory is only available in **Chat Driven** workflows. The toggle does not appear on nodes in Output Focused workflows.
***
## Summarization
When a conversation exceeds 3 message turns, FlowX automatically summarizes the older messages using an LLM call. The summarization process:
1. Collects all messages older than the most recent 3 turns
2. Sends them to the LLM with a context extraction prompt
3. Stores the resulting summary for future retrieval
4. The summary is regenerated on each new message to incorporate the latest context
The summarization prompt instructs the LLM to extract the most important and relevant context from the conversation history, focusing on information needed to maintain coherent dialogue.
Summarization happens automatically — there is no configuration needed. The summary prefix `## Previous conversation summary:` is prepended to the summary text in the prompt.
***
## Debugging memory in the console
The workflow console log includes a **Memory** tab that shows exactly what memory was sent to the LLM for each workflow run.
| Level | Tab | Content |
| --------------------- | ---------- | ----------------------------------------------------- |
| **Workflow** | Memory tab | Full conversation history and summary sent to the LLM |
| **Workflow** | Input tab | User Message and Chat Session ID in JSON format |
| **Workflow** | Output tab | Chat response as text |
| **Custom Agent node** | Output tab | Node response as text |
### Memory use summary
When an AI node with **Use Memory** enabled completes, the console displays the `memoryUseSummary` — a snapshot of the conversation context that was available to the LLM. This includes:
* **summary** — the auto-generated summary of earlier conversation history
* **chatMessages** — the most recent message pairs (up to 3, excluding the current message) with their `messageId`, `userMessage`, and `agentMessage`
```json theme={"system"}
{
"summary": "User asked about account setup and verification steps.",
"chatMessages": [
{
"messageId": "msg-003",
"userMessage": "What documents do I need?",
"agentMessage": "You need a valid ID and proof of address."
},
{
"messageId": "msg-002",
"userMessage": "How do I verify my account?",
"agentMessage": "Navigate to Settings and select Verification."
}
]
}
```
The Memory tab only appears when `memoryUseSummary` data exists for the workflow instance. If you don't see it, verify that **Use Memory** is enabled on at least one node.
***
## Storage
| Data | Storage | Details |
| -------------------- | -------------------------------- | ------------------------------------------------------ |
| Session ID | Browser storage + FlowX Database | Links the client chat instance to the server session |
| Message history | FlowX Database | Complete record of user and agent messages per session |
| Conversation summary | FlowX Database | Auto-generated summary of older conversation turns |
| Session metadata | FlowX Database | Timestamps, workflow reference, user info |
Session memory is tied to the `chatSessionId` — the same session ID retrieves the same memory across workflow runs. The Chat component manages session IDs automatically.
***
## Limitations
* Memory is **session-scoped** — there is no cross-session or cross-user memory
* The latest **3 turns** are always included in full; older turns are summarized
* Summarization uses an LLM call, which adds latency on the first message after the 3-turn threshold
* Memory cannot be manually edited or cleared from the Designer UI
* Only **user messages** and **agent responses** are stored — internal workflow data is excluded
***
## Related resources
Full guide to building Chat Driven workflows with memory and intent routing
Runtime behavior, session management, and display modes
Route conversations based on detected user intent with optional memory
Configure AI nodes with memory, chat reply, and response settings