Memory Management
Overview
The Memory Management page is used to view and maintain account-level memory. Account-level memory is shared across organizations and bound to the individual user — when you log in to the same account in different organizations, your memory follows you.
When an Agent starts a new conversation, it automatically queries account memory and injects it into the system prompt as a "user profile." This lets the Agent truly "know you."
Two Views
| View | Description |
|---|---|
| List | Paginated browsing of all memory entries + search + edit / delete / clear |
| Graph | Force-directed graph visualization showing the network of associations between memories |
Switching views: there is a List / Graph tab at the top of the page.
List View
Memory Entries
Each memory is displayed as a card:
| Field | Description |
|---|---|
| Text Content | The core information of the memory (e.g., "User prefers TypeScript") |
| Confidence | The confidence score at extraction time (0 ~ 1) |
| Metadata | Additional information (source session ID, extraction time, etc.) |
Actions
| Action | Description |
|---|---|
| Search | Filter the memory list by keyword |
| Manual Add | Enter a new memory entry |
| Edit | Modify existing memory content |
| Delete | Remove a single memory |
| Clear All | Clear all account-level memory in one click (⚠️ irreversible) |
Pagination
- Default 50 entries per page
- Initial load fetches 50 entries + graph data
- The
Load Morebutton at the bottom appends the next page
Graph View
Switch to the graph view to visualize the relationship network between memories:
Node Types
| Type | Color | Description |
|---|---|---|
| Hub | Purple #6d28d9 |
User center node |
| Fact | Blue #2563eb |
A single memory entry |
| Entity | Amber #f59e0b |
An entity extracted from memory |
Edge Types
| Type | Style | Description |
|---|---|---|
| Hierarchical relationship | Gray edge | Fact → Hub, Entity → Hub |
| Association relationship | Amber edge | Entity ↔ Entity cross-association |
Interactions
- Click a node: view details, with support for edit / delete
- Hover a node: show the full text
- Drag: adjust node position
- Zoom: scroll to zoom the view
- Collapsed mode: when there are more than 100 memories, the center node displays a count label (e.g., "128 memories") instead of all fact nodes
Graph Rendering Optimization (2026-04 Update) NEW
The new graph introduces a dual density gate mechanism to avoid unreadable label overlap when there are too many nodes:
| Threshold Constant | Value | Meaning |
|---|---|---|
PILL_READABILITY_MIN_SCALE |
1.5 | Text labels are not rendered when the zoom scale is below 1.5x |
PILL_MIN_SCREEN_AREA_PER_NODE |
3000 | Text labels are not rendered when the average on-screen area per node is less than 3000 pixels |
Labels are only rendered when both conditions are met simultaneously. When they are not met, only dots are shown to keep the view visually clear.
Entity type color stabilization: The color of all Entity nodes is computed via a djb2 hash — the same entity type keeps the same color across different views and at different times, making visual recognition easier. The 10-color palette supports unlimited type cycling.
Entity type field fix: The sourceTypes and targetTypes fields are now correctly passed to the frontend graph, fixing a bug where all nodes previously fell back to gray because these fields were lost.
How Memory Is Written
Manual Add
In the list view, click the "Add" button, enter the memory text, and save.
Automatic Extraction from Conversation
The Agent automatically detects information worth remembering during a conversation and stores it:
| Trigger | Example | Confidence |
|---|---|---|
| Explicit command | "Remember I like to use TypeScript" | 0.99 |
| Implicit detection | "My name is Zhang San, I'm a frontend developer" | 0.93 |
For the detailed extraction mechanism, see Three-Dimensional Memory System.
The Role of Memory in Conversations
- New conversation: the system queries the account-level memory profile and injects it into the system prompt as a "user profile"
- Asking questions during a conversation: the Agent queries relevant memories based on the semantics of the user's question (the memory_query tool)
- Memory updates: new conversation information may trigger updates or merges of existing memories
Cross-Account Gateway Isolation NEW
To protect privacy, sessions dispatched via a cross-account Gateway cannot access your account memory:
- When your node is set to enterprise and called by another account in the same organization, the execution process cannot read your account memory
- Only enterprise memory (organization-level) can be accessed by cross-account sessions
- Your own local sessions are not affected
For details, see Multi-Node Architecture — Cross-Account Isolation.
Operation Guide
How to Manually Add Memory
- Open APP → Settings → Memory
- In the list view, click the Add button
- Enter the information you want the Agent to remember, for example: "I prefer to write code in TypeScript"
- Save
Effect: the next time you start a conversation, the Agent will know you prefer TypeScript and will automatically use TypeScript in code-related replies.
How to View the Memory Graph
- On the Memory Management page, switch to the Graph View
- The large purple circle in the center is your user center node
- The blue nodes are memory entries (e.g., "likes TypeScript"), connected to the center via gray edges
- The amber nodes are extracted entities (e.g., "TypeScript", "React"), and entities are connected to each other with amber edges to indicate associations
- Actions:
- Click any node to view details, edit, or delete
- Drag a node to adjust its position
- Scroll to zoom the view
- Hover a node to view the full text
How to Clear Memory
- On the Memory Management page, click the Clear All button
- Confirm the clear operation
Note: the clear operation is irreversible. After clearing, the Agent will no longer know any of your preferences, background, or historical information in future conversations. The Agent will interact with you as if meeting for the first time. Only use this feature when you truly need to "start over."
The Impact of Memory on Conversation Quality
| Memory State | Conversation Experience |
|---|---|
| Rich and accurate memory | The Agent understands your identity, preferences, and project background, delivering personalized and precise replies |
| Empty memory | The Agent acts as if meeting you for the first time, requiring you to re-explain your background and preferences each time |
| Outdated or incorrect memory | The Agent may give inappropriate suggestions based on incorrect information |
| Too much memory | The memory profile injected into the conversation grows longer and may take up context space |
Recommendation: Periodically (e.g., once a month) browse the memory list and delete outdated or no-longer-accurate entries. Keeping your memory concise and accurate is the key to the best conversation experience.
Automatic Extraction vs. Manual Add
| Comparison | Automatic Extraction | Manual Add |
|---|---|---|
| Trigger | The Agent automatically detects during a conversation | You manually enter it on the Memory Management page |
| Accuracy | Depends on the confidence threshold; may miss or mis-extract | Fully under your control, 100% accurate |
| Use case | Naturally accumulating during everyday conversations | When you want the Agent to remember specific information (e.g., project conventions) |
Tip: The two approaches can be combined. Let automatic extraction handle everyday information, and for important project conventions and work preferences, manually add them to ensure accuracy.
FAQ
Why didn't the Agent remember what I said?
Possible reasons:
- Insufficient confidence: the information didn't seem "like a fact worth remembering." Try to express it more explicitly: "Please remember, I like using a dark theme."
- Filtered by exclusion rules: pure questions, small talk, content inside code blocks, and time-sensitive information (e.g., "today's weather") are not extracted.
- Extraction delay: memory is only extracted after each conversation round ends; if the conversation is interrupted, extraction may not complete.
Why did the Agent answer based on incorrect memory?
You may have an outdated memory entry. How to fix it:
- Open Memory Management and search for the relevant memory
- Find the incorrect entry and click Edit to correct it, or Delete it
- You can also say in the conversation: "Forget that I said I liked Python before, I use TypeScript now."
How much context space does memory take up?
When starting a new conversation, the system injects memory into the system prompt as a profile, typically taking up 200-1000 tokens (depending on the number of memories). Relative to a 128K-200K context window, this is a very small proportion and does not significantly affect available space.
Related Documentation
- Three-Dimensional Memory System — technical deep dive into the memory system (including enterprise-level and session-level)
- Advanced Settings — Enterprise Memory — enterprise-level memory management (admin)
- Multi-Node Architecture — cross-account Gateway isolation