Three-Dimensional Memory System

Overview

The workspace memory system adopts a three-dimensional architecture, storing and managing knowledge across three dimensions: account, enterprise, and session. Memory is deeply integrated with conversation — every conversation queries relevant memories to inject into the context, and new memories are automatically extracted after each interaction turn.

Technical foundation: mem0 REST API + Neo4j knowledge graph.

Three-Dimensional Architecture

Account-Level Memory

• Cross-organization sharing: Bound to userId, available across all of the user's organizations
• Write methods:
  • Manual: Add via commands in the memory management interface or in conversation
  • Automatic: Automatically detects information worth remembering during conversation
• Management interface: APP Settings → Memory
• Detailed documentation: Memory Management

Enterprise-Level Memory

• Organization isolation: Bound to orgId, visible only to members of the current organization
• Management permission: Only workspace administrators can maintain it
• Management interface: Workspace Management → Advanced Settings → Enterprise Memory
• Detailed documentation: Advanced Settings — Enterprise Memory

Session-Level Memory

• Single session: Bound to runId, valid only within the current session
• Per-turn extraction: Automatically analyzes and extracts memory after each conversation turn
• No management needed: The system handles it automatically

Memory Extraction Mechanism

Explicit Commands (Confidence 0.99)

The user directly asks the Agent to remember or forget information:

Implicit Detection (Confidence 0.5-0.93)

The system automatically detects persistent facts in conversation via regex patterns:

Confidence Thresholds

Automatic Exclusion Rules

The following content will not be extracted as memory:

• Pure questions (ending with a question mark, starting with an interrogative word)
• Small talk / greetings
• Content within code blocks
• Time-sensitive / time-bound information (dates, news, temporary states)
• Non-persistent topics (bug reports, error messages)

Knowledge Graph

Backend Storage (Neo4j)

Memory relationships are stored in the Neo4j graph database as triples:

                      
                      (source entity) --[relationship]--> (target entity)

                    

N-hop neighborhood query (1-4 hops):

                      
                      MATCH path = (n {name: $entity})-[*1..depth]-(m)
WHERE ALL(node IN nodes(path) WHERE node.user_id = $user_id)
UNWIND relationships(path) AS rel
RETURN source, relationship, target

                    

The Cypher query guarantees it never crosses user boundaries.

Frontend Visualization

Uses react-force-graph-2d to render a force-directed graph:

Graph Rendering Optimization (2026-04 Update) NEW

The new graph introduces two density gate thresholds to prevent label overlap caused by too many nodes:

Labels are shown only when both conditions are met. When they are not met, only dots are rendered.

Entity Type Color Stabilization:

• Computes entity type → color index via a djb2 hash
• Entities of the same type keep the same color across different views and different times
• The 10-color palette supports cyclic reuse for an unlimited number of types

Bug Fix: When mapping graph relationships, the sourceTypes / targetTypes fields were previously dropped, causing all nodes to fall back to the gray fallback color. This bug has been fixed (see sidecar/src/mem0Service.ts).

Multi-Level Drill-Down

Clicking a node expands its associated entities, drilling deeper into the knowledge network layer by layer. The graph engine supports:

• neighborhood() — N-hop neighborhood query
• shortestPath() — Shortest path between two entities
• extractEntitiesFromText() — Extract entity names from text

Synergy Between Memory and Conversation

Memory Query Timing

Memory Injection Location

Two memory profiles are injected into the system prompt:

                      
                      你是 GPTBots AI 助手...

## 用户概况
- 用户是前端开发工程师
- 偏好使用 TypeScript
- ...

## 企业概况
- 公司使用 React 技术栈
- 项目代号 Project Alpha
- ...

                    

Memory Tools

The Agent can proactively query and manage memory through the following tools:

Cross-Account Gateway Isolation NEW

When a node is invoked cross-account with enterprise scope:

• Account-level memory (userId-bound): ❌ Not accessible
• Enterprise-level memory (orgId-bound): ✅ Accessible
• Session-level memory (within this session): ✅ Accessible

Filtering is enforced during the mem0Service query phase via the isRemoteSession flag — a cross-account invocation cannot read the personal memory of the target node's owner.

Design purpose: Protect personal privacy while not affecting the collaborative sharing of organization-level knowledge. See Multi-Node Architecture for details.

mem0 Capabilities

The memory backend, based on mem0, provides the following automated capabilities:

Timeout Protection

What This Means for You

The memory system enables the Agent to "know you". Without memory, the Agent is like a stranger meeting you for the first time in every conversation; with memory, the Agent knows your preferences, project background, and work habits.

How the three dimensions manifest in practice:

• Account memory: "You said you like TypeScript" → the Agent prioritizes TypeScript in conversations across all organizations
• Enterprise memory: An administrator added "The company uses a PostgreSQL database" → when any member of the organization discusses databases with the Agent, the Agent recommends the PostgreSQL solution by default
• Session memory: You said "The current project is called Project Alpha" in this conversation → the Agent remembers it in this conversation, but not necessarily in a new one

You can manage memory in the following ways:

• For personal preferences: Say "记住 I prefer the dark theme" in conversation, or add it manually in APP Settings → Memory
• For enterprise knowledge: Contact an administrator to add it in Workspace Management → Advanced Settings → Enterprise Memory
• If you find a memory is incorrect: Say "忘掉 the previous preference about Python" in conversation, or delete it directly in memory management
• To view existing memories: APP Settings → Memory, where you can browse the list and the graph

Related Documentation

• Memory Management — Memory management interface on the APP side
• Advanced Settings — Enterprise Memory — Enterprise-level memory management
• Tool Management — Memory-related tools (memory_query, memory_manage)
• Multi-Node Architecture — Cross-account Gateway isolation mechanism