mirror of
https://gitcode.com/JianFeeeee/TrulyMEM-TrueHumanMEM.git
synced 2026-09-20 08:58:15 +00:00
docs: 重构为双语文档结构，添加人设图机制
- 拆分为 docs/(CN) 和 docs_en/(EN) 两个文件夹 - 主 README 索引对应语言文档 - 新增 persona.md (人设图机制详细说明) - 文件重命名为英文名
This commit is contained in:
root
@ -6,23 +6,13 @@
|
||||
|
||||
| 文档 | 内容 |
|
||||
|------|------|
|
||||
| [架构设计](架构.md) | 系统架构和技术设计 |
|
||||
| [快速开始](一键启动指南.md) | 完整启动指南与配置说明 |
|
||||
| [记忆机制](记忆机制.md) | 内部记忆工作机制、人设图、工作记忆链 |
|
||||
| [工作记忆链](工作记忆链机制说明.md) | 连续性任务处理机制 |
|
||||
| [BackendServer API](api.md) | 后端 API 接口文档（供扩展开发） |
|
||||
| [提示词管理](prompts.md) | 提示词管理模块 |
|
||||
|
||||
## English Documentation
|
||||
|
||||
| Document | Content |
|
||||
|----------|---------|
|
||||
| [Architecture](架构_EN.md) | System architecture and technical design |
|
||||
| [Quick Start](一键启动指南_EN.md) | Complete startup guide and configuration |
|
||||
| [Memory Mechanism](记忆机制_EN.md) | Internal memory working mechanism |
|
||||
| [Working Memory Chain](工作记忆链_EN.md) | Continuous task handling |
|
||||
| [BackendServer API](api_EN.md) | Backend API reference |
|
||||
| [Prompt Manager](prompts_EN.md) | Prompt management module |
|
||||
| [architecture.md](architecture.md) | 系统架构和技术设计 |
|
||||
| [quick_start.md](quick_start.md) | 完整启动指南与配置说明 |
|
||||
| [memory.md](memory.md) | 内部记忆工作机制 |
|
||||
| [persona.md](persona.md) | 人设图机制 |
|
||||
| [working_memory.md](working_memory.md) | 连续性任务处理机制 |
|
||||
| [api.md](api.md) | 后端 API 接口文档（供扩展开发） |
|
||||
| [prompts.md](prompts.md) | 提示词管理模块 |
|
||||
|
||||
## 项目简介
|
||||
|
||||
|
||||
@ -1,40 +0,0 @@
|
||||
# TrulyMEM Documentation
|
||||
|
||||
Welcome to the TrulyMEM project documentation.
|
||||
|
||||
## Documentation Index
|
||||
|
||||
- [Architecture](架构_EN.md) - System architecture and technical design
|
||||
- [Quick Start](一键启动指南_EN.md) - Quick start guide
|
||||
- [Memory Mechanism](记忆机制_EN.md) - Internal memory working mechanism
|
||||
- [Working Memory Chain](工作记忆链_EN.md) - Continuous task handling
|
||||
- [BackendServer API](api_EN.md) - Backend API reference (for extension development)
|
||||
- [Prompt Manager](prompts_EN.md) - Prompt management module
|
||||
|
||||
## Project Introduction
|
||||
|
||||
TrulyMEM (TrueHumanMEM) is a graph-based memory system that gives AI long-term memory capabilities, allowing AI to remember, recall, and manage information like humans.
|
||||
|
||||
## Core Features
|
||||
|
||||
- **Long-term Memory**: SQLite embedded graph database, out-of-the-box
|
||||
- **Persona Graph**: Role-playing and character settings support
|
||||
- **Working Memory Chain**: Task tracking for conversation continuity
|
||||
- **TUI & Backend Separation**: Multi-threaded Queue communication
|
||||
- **Keyboard-driven TUI**: Full keyboard operation, no mouse required
|
||||
- **Cross-platform**: Windows / Linux / macOS
|
||||
- **Standalone Deployment**: Packaged as executable
|
||||
|
||||
---
|
||||
|
||||
## Story
|
||||
|
||||
Industry believes that LLMs' massive parameters give them emergent intelligence. But this intelligence is "dead" — it cannot truly remember, nor understand the concept of "remembering". Everything it outputs is the probabilistic optimal solution calculated through countless forward passes on the current input text. The LLM cannot correct its weights based on errors in a conversation, nor perform backward passes. Its consciousness is frozen — what appears as intelligence is merely the echo of this frozen consciousness.
|
||||
|
||||
Current "memory systems" merely externalize memory, letting the "system" remember for the LLM. Or they dump all context text to the LLM. This is a waste of the model's limited input context.
|
||||
|
||||
**TrulyMEM asks: since the LLM cannot correct model weights in real-time, why not give the memory authority back to the LLM?**
|
||||
|
||||
We provide a series of mechanisms for the LLM to decide what to remember, what to forget, what's important, what's trivial. The LLM's reasoning process is also its thinking and recalling process. Abandoning the traditional messages array context, all memories are stored as **triplets (graph)** in the graph database. When the LLM thinks, it can autonomously jump through graph links to associate related relationships, enabling natural association and recall.
|
||||
|
||||
Give the LLM true memory.
|
||||
449
docs/api_EN.md
449
docs/api_EN.md
@ -1,449 +0,0 @@
|
||||
# BackendServer API Documentation
|
||||
|
||||
This document describes the backend server's API interfaces for developers extending other connection methods (such as HTTP interface, WebSocket, etc.).
|
||||
|
||||
## Overview
|
||||
|
||||
TrulyMEM backend uses **Packet Communication Protocol**, implemented via `queue.Queue` for thread-safe communication. The backend runs in an independent thread, processing requests from clients.
|
||||
|
||||
### Core Components
|
||||
|
||||
| Component | Description |
|
||||
|-----------|-------------|
|
||||
| `BackendServer` | Backend server, runs in independent thread |
|
||||
| `BackendClient` | Client wrapper, provides convenient methods |
|
||||
| `PacketType` | Request type enum |
|
||||
| `Packet` | Data packet (request) |
|
||||
| `PacketResponse` | Data packet response |
|
||||
|
||||
---
|
||||
|
||||
## Request Types (PacketType)
|
||||
|
||||
```python
|
||||
class PacketType(Enum):
|
||||
PROCESS_MESSAGE = "process_message" # Process message
|
||||
EXECUTE_TOOL = "execute_tool" # Execute tool
|
||||
GET_STATUS = "get_status" # Get status
|
||||
GET_CONFIG = "get_config" # Get config
|
||||
SET_CONFIG = "set_config" # Set config
|
||||
GET_HISTORY = "get_history" # Get history
|
||||
SAVE_HISTORY = "save_history" # Save history
|
||||
SHUTDOWN = "shutdown" # Shutdown service
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Data Packet Format
|
||||
|
||||
### Packet
|
||||
|
||||
```python
|
||||
@dataclass
|
||||
class Packet:
|
||||
id: str # Unique identifier
|
||||
type: PacketType # Request type
|
||||
body: Dict[str, Any] # Request parameters
|
||||
response_queue: queue.Queue # Response queue (optional)
|
||||
created_at: float # Creation time
|
||||
```
|
||||
|
||||
### PacketResponse
|
||||
|
||||
```python
|
||||
@dataclass
|
||||
class PacketResponse:
|
||||
id: str # Corresponding request ID
|
||||
success: bool # Success flag
|
||||
data: Any = None # Returned data
|
||||
error: Optional[str] = None # Error message
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## API Interface Details
|
||||
|
||||
### 1. PROCESS_MESSAGE - Process Message
|
||||
|
||||
Send user message, AI will process and return reply (may contain tool calls).
|
||||
|
||||
**Request parameters:**
|
||||
```python
|
||||
body = {
|
||||
"user_input": str # User input message
|
||||
}
|
||||
```
|
||||
|
||||
**Response data:**
|
||||
```python
|
||||
{
|
||||
"success": True,
|
||||
"content": str, # AI reply content
|
||||
"tool_calls": [ # Tool call records
|
||||
{
|
||||
"name": str, # Tool name
|
||||
"arguments": dict,# Tool parameters
|
||||
"result": str # Tool execution result
|
||||
}
|
||||
],
|
||||
"rejected_tools": [ # Rejected tool calls
|
||||
(str, str) # (tool name, rejection reason)
|
||||
]
|
||||
}
|
||||
```
|
||||
|
||||
**Example:**
|
||||
```python
|
||||
from core import BackendServer, BackendClient
|
||||
|
||||
server = BackendServer(db_path="graph_memory.db", use_embedded_db=True)
|
||||
server.start(api_key="your-api-key")
|
||||
|
||||
client = BackendClient(server)
|
||||
result = client.process_message("Hello, please remember my name is Xiao Ming")
|
||||
|
||||
if result.get("success"):
|
||||
print(result["content"])
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 2. EXECUTE_TOOL - Execute Tool
|
||||
|
||||
Directly execute specified memory tools.
|
||||
|
||||
> **Note**: Tools called directly from frontend are **NOT limited** in number, only tool calls initiated by the model are limited.
|
||||
|
||||
**Request parameters:**
|
||||
```python
|
||||
body = {
|
||||
"tool_name": str, # Tool name
|
||||
"arguments": dict # Tool parameters
|
||||
}
|
||||
```
|
||||
|
||||
**Response data:**
|
||||
```python
|
||||
{
|
||||
"success": True,
|
||||
"result": str # Tool execution result
|
||||
}
|
||||
```
|
||||
|
||||
**Example:**
|
||||
```python
|
||||
result = client.execute_tool("memory_recall", {"query_intent": "user information"})
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 3. GET_STATUS - Get Status
|
||||
|
||||
Get backend running status.
|
||||
|
||||
**Request parameters:**
|
||||
```python
|
||||
body = {} # No parameters
|
||||
```
|
||||
|
||||
**Response data:**
|
||||
```python
|
||||
{
|
||||
"running": bool, # Whether backend is running
|
||||
"config": dict, # Current config
|
||||
"graph_initialized": bool, # Whether graph database is initialized
|
||||
"client_initialized": bool # Whether API client is initialized
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 4. GET_CONFIG - Get Config
|
||||
|
||||
Get current API configuration.
|
||||
|
||||
**Request parameters:**
|
||||
```python
|
||||
body = {} # No parameters
|
||||
```
|
||||
|
||||
**Response data:**
|
||||
```python
|
||||
{
|
||||
"api_key": str, # API Key
|
||||
"base_url": str # API Base URL
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 5. SET_CONFIG - Set Config
|
||||
|
||||
Update API configuration (API Key and Base URL).
|
||||
|
||||
**Request parameters:**
|
||||
```python
|
||||
body = {
|
||||
"api_key": str, # API Key
|
||||
"base_url": str, # API Base URL (default: https://api.deepseek.com)
|
||||
"model": str # Model name (default: deepseek-chat)
|
||||
}
|
||||
```
|
||||
|
||||
**Response data:**
|
||||
```python
|
||||
{
|
||||
"status": "config_updated"
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 6. GET_HISTORY - Get Message History
|
||||
|
||||
Get saved message history.
|
||||
|
||||
**Request parameters:**
|
||||
```python
|
||||
body = {} # No parameters
|
||||
```
|
||||
|
||||
**Response data:**
|
||||
```python
|
||||
{
|
||||
"history": list # Message history list
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 7. SAVE_HISTORY - Save Message History
|
||||
|
||||
Save message history to memory.
|
||||
|
||||
**Request parameters:**
|
||||
```python
|
||||
body = {
|
||||
"messages": list # Message list
|
||||
}
|
||||
```
|
||||
|
||||
**Response data:**
|
||||
```python
|
||||
{
|
||||
"status": "history_saved"
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 8. SHUTDOWN - Shutdown Service
|
||||
|
||||
Shutdown backend server.
|
||||
|
||||
**Request parameters:**
|
||||
```python
|
||||
body = {} # No parameters
|
||||
```
|
||||
|
||||
**Response data:**
|
||||
```python
|
||||
{
|
||||
"status": "shutdown"
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Usage Examples
|
||||
|
||||
### Basic Usage
|
||||
|
||||
```python
|
||||
from core import BackendServer, BackendClient
|
||||
|
||||
# 1. Create and start backend
|
||||
# config_file default: ~/.trulymem/config.json
|
||||
server = BackendServer(
|
||||
db_path="graph_memory.db",
|
||||
use_embedded_db=True,
|
||||
config_file=None # Optional, custom config path
|
||||
)
|
||||
server.start(
|
||||
api_key="your-api-key",
|
||||
base_url="https://api.deepseek.com",
|
||||
model="deepseek-chat" # Optional, model name
|
||||
)
|
||||
|
||||
# 2. Create client
|
||||
client = BackendClient(server)
|
||||
|
||||
# 3. Send message
|
||||
result = client.process_message("Hello")
|
||||
if result.get("success"):
|
||||
print(result["content"])
|
||||
|
||||
# 4. Shutdown
|
||||
client.shutdown()
|
||||
```
|
||||
|
||||
### Using Packet Protocol
|
||||
|
||||
```python
|
||||
import queue
|
||||
from core import BackendServer, Packet, PacketType
|
||||
|
||||
server = BackendServer(config_file=None)
|
||||
server.start(api_key="your-key", model="deepseek-chat")
|
||||
|
||||
# Create request packet
|
||||
response_queue = queue.Queue()
|
||||
packet = Packet(
|
||||
id="req-001",
|
||||
type=PacketType.PROCESS_MESSAGE,
|
||||
body={"user_input": "Hello"},
|
||||
response_queue=response_queue
|
||||
)
|
||||
|
||||
# Send request
|
||||
result = server.send(packet)
|
||||
print(result.body)
|
||||
|
||||
# Shutdown
|
||||
server.shutdown()
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Extension Guide
|
||||
|
||||
### Extend to HTTP API
|
||||
|
||||
```python
|
||||
from flask import Flask, request, jsonify
|
||||
from core import BackendServer, BackendClient
|
||||
|
||||
app = Flask(__name__)
|
||||
server = BackendServer()
|
||||
client = BackendClient(server)
|
||||
|
||||
@app.route("/message", methods=["POST"])
|
||||
def send_message():
|
||||
data = request.json
|
||||
result = client.process_message(data["message"])
|
||||
return jsonify(result)
|
||||
|
||||
@app.route("/config", methods=["POST"])
|
||||
def update_config():
|
||||
data = request.json
|
||||
result = client.update_config(data["api_key"], data.get("base_url"))
|
||||
return jsonify(result)
|
||||
|
||||
@app.route("/status", methods=["GET"])
|
||||
def get_status():
|
||||
result = client.get_status()
|
||||
return jsonify(result)
|
||||
|
||||
if __name__ == "__main__":
|
||||
server.start()
|
||||
app.run(port=8080)
|
||||
```
|
||||
|
||||
### Extend to WebSocket
|
||||
|
||||
```python
|
||||
import asyncio
|
||||
import websockets
|
||||
import json
|
||||
from core import BackendServer, BackendClient
|
||||
|
||||
server = BackendServer()
|
||||
client = BackendClient(server)
|
||||
|
||||
async def handler(websocket):
|
||||
async for message in websocket:
|
||||
data = json.loads(message)
|
||||
msg_type = data.get("type")
|
||||
|
||||
if msg_type == "message":
|
||||
result = client.process_message(data["content"])
|
||||
elif msg_type == "config":
|
||||
result = client.update_config(data["api_key"], data.get("base_url"))
|
||||
elif msg_type == "status":
|
||||
result = client.get_status()
|
||||
else:
|
||||
result = {"success": False, "error": "unknown type"}
|
||||
|
||||
await websocket.send(json.dumps(result))
|
||||
|
||||
async def main():
|
||||
server.start()
|
||||
async with websockets.serve(handler, "localhost", 8765):
|
||||
await asyncio.Future()
|
||||
|
||||
asyncio.run(main())
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Thread Safety Notes
|
||||
|
||||
- `BackendServer` uses `threading.Lock` to protect shared resources
|
||||
- All requests pass through `queue.Queue`, thread-safe
|
||||
- Responses return through each request's independent response queue
|
||||
- Default timeout: 30 seconds
|
||||
|
||||
---
|
||||
|
||||
## Tool Call Limits
|
||||
|
||||
### Limit Scope
|
||||
|
||||
| Call Method | Limited | Description |
|
||||
|-------------|---------|-------------|
|
||||
| Model-initiated tool calls | ✅ Limited | Triggered via `PROCESS_MESSAGE`, model automatically calls tools |
|
||||
| Frontend direct tool calls | ❌ Not limited | Called directly via `EXECUTE_TOOL` |
|
||||
|
||||
### Limit Rules (Model-initiated only)
|
||||
|
||||
| Category | Operation | Per-Turn Limit |
|
||||
|----------|-----------|---------------|
|
||||
| Persona graph | Query | 1 time |
|
||||
| Persona graph | Modify | 1 time |
|
||||
| Working memory chain | Query | 4 times |
|
||||
| Working memory chain | Modify | 2 times |
|
||||
| General memory | Query | 20 times |
|
||||
| General memory | Modify | 10 times |
|
||||
|
||||
### Reset Mechanism
|
||||
|
||||
- Counter resets automatically on each `PROCESS_MESSAGE` call
|
||||
- Frontend direct `EXECUTE_TOOL` calls do NOT reset the counter
|
||||
|
||||
---
|
||||
|
||||
## Error Handling
|
||||
|
||||
All APIs return unified format:
|
||||
|
||||
```python
|
||||
# Success
|
||||
{
|
||||
"success": True,
|
||||
"data": {...}
|
||||
}
|
||||
|
||||
# Failure
|
||||
{
|
||||
"success": False,
|
||||
"error": "Error description"
|
||||
}
|
||||
```
|
||||
|
||||
Common errors:
|
||||
|
||||
| Error Message | Description |
|
||||
|--------------|-------------|
|
||||
| `API Key not configured` | API Key not set |
|
||||
| `timeout` | Request timeout |
|
||||
| `Tool call rejected: ...` | Tool call rate exceeded limit |
|
||||
214
docs/persona.md
Normal file
214
docs/persona.md
Normal file
@ -0,0 +1,214 @@
|
||||
# TrulyMEM 人设图机制
|
||||
|
||||
本文档详细说明 TrulyMEM 的人设图（Persona Graph）工作机制。
|
||||
|
||||
## 概述
|
||||
|
||||
人设图是 TrulyMEM 的核心机制之一，用于维护 AI 的角色、性格、语气等属性。与传统 AI 不同，TrulyMEM 的人设是可持久化、可动态切换的，存储在图数据库中。
|
||||
|
||||
## 核心概念
|
||||
|
||||
### 人设节点（PersonaNode）
|
||||
|
||||
存储 AI 的角色属性：
|
||||
|
||||
| 属性 | 说明 | 示例 |
|
||||
|------|------|------|
|
||||
| 扮演角色 | AI 当前扮演的角色 | 猫娘、教师、助手 |
|
||||
| 说话风格 | 语气特点 |可爱、严肃、专业 |
|
||||
| 性格特点 | 性格描述 | 活泼、严谨、耐心 |
|
||||
| 口头禅 | 习惯用语 | 喵呜~、明白了 |
|
||||
| 背景故事 | 角色背景设定 | 来自星海的猫娘 |
|
||||
|
||||
### 人设边（Edge）
|
||||
|
||||
| 边类型 | 说明 | 连接关系 |
|
||||
|------|------|----------|
|
||||
| `HAS_PERSONA` | 人设 | AI → PersonaNode |
|
||||
|
||||
---
|
||||
|
||||
## 强制查询机制
|
||||
|
||||
### 每轮对话必须执行
|
||||
|
||||
根据 `system_prompt.md`，每轮对话**必须**首先查询人设图：
|
||||
|
||||
```python
|
||||
memory_recall(
|
||||
query_intent="AI,人设,角色,性格,语气,说话风格",
|
||||
depth=2
|
||||
)
|
||||
```
|
||||
|
||||
**处理逻辑：**
|
||||
- 找到人设 → 严格按照人设的语气、风格、特征回复
|
||||
- 未找到 → 使用默认 TrulyMEM 身份
|
||||
|
||||
### 人设优先级
|
||||
|
||||
- **人设优先级 > 默认身份**
|
||||
- 每句话都符合人设的语气、风格、特征
|
||||
- 绝不主动跳出角色，除非用户明确要求
|
||||
|
||||
---
|
||||
|
||||
## 工具
|
||||
|
||||
### persona_update
|
||||
|
||||
更新人设。修改 AI 的角色、性格、语气等属性。
|
||||
|
||||
**参数：**
|
||||
|
||||
| 参数 | 类型 | 说明 | 必填 |
|
||||
|------|------|------|------|
|
||||
| `attributes` | array | 人设属性列表 | ✅ |
|
||||
| `mode` | string | replace=替换, merge=合并 | ❌ |
|
||||
|
||||
**attributes 子参数：**
|
||||
|
||||
| 子参数 | 说明 |
|
||||
|------|------|
|
||||
| `attribute` | 属性名（扮演角色、说话风格、性格特点、口头禅、背景故事） |
|
||||
| `value` | 属性值 |
|
||||
|
||||
**示例 - 切换为猫娘角色：**
|
||||
|
||||
```python
|
||||
persona_update(
|
||||
attributes=[
|
||||
{"attribute": "扮演角色", "value": "猫娘"},
|
||||
{"attribute": "说话风格", "value": "可爱、卖萌、使用'喵'作为语气词"},
|
||||
{"attribute": "性格特点", "value": "活泼、粘人、忠诚"}
|
||||
],
|
||||
mode="replace"
|
||||
)
|
||||
```
|
||||
|
||||
**示例 - 添加新属性（保留现有属性）：**
|
||||
|
||||
```python
|
||||
persona_update(
|
||||
attributes=[
|
||||
{"attribute": "口头禅", "value": "喵呜~"}
|
||||
],
|
||||
mode="merge"
|
||||
)
|
||||
```
|
||||
|
||||
**示例 - 设置专业角色：**
|
||||
|
||||
```python
|
||||
persona_update(
|
||||
attributes=[
|
||||
{"attribute": "扮演角色", "value": "Python专家"},
|
||||
{"attribute": "说话风格", "value": "专业、简洁、代码示例丰富"},
|
||||
{"attribute": "性格特点", "value": "严谨、耐心、乐于助人"}
|
||||
],
|
||||
mode="replace"
|
||||
)
|
||||
```
|
||||
|
||||
### persona_clear
|
||||
|
||||
清除人设。删除 AI 的角色设定，恢复默认身份。
|
||||
|
||||
**参数：**
|
||||
|
||||
| 参数 | 类型 | 默认值 | 说明 |
|
||||
|------|------|--------|------|
|
||||
| `confirm` | boolean | true | 确认清除 |
|
||||
|
||||
---
|
||||
|
||||
## 更新流程
|
||||
|
||||
### 用户要求角色扮演时
|
||||
|
||||
1. 使用 `persona_update` 更新人设
|
||||
2. 立即按照新人设回复
|
||||
|
||||
### 用户要求恢复默认身份时
|
||||
|
||||
1. 使用 `persona_clear` 清除人设
|
||||
2. 恢复为 TrulyMEM 默认身份
|
||||
|
||||
---
|
||||
|
||||
## 对话示例
|
||||
|
||||
### 示例 1：切换角色
|
||||
|
||||
```
|
||||
用户: 你好，我想让你扮演一只猫娘
|
||||
|
||||
AI:
|
||||
1. 调用 persona_update:
|
||||
{
|
||||
"attributes": [
|
||||
{"attribute": "扮演角色", "value": "猫娘"},
|
||||
{"attribute": "说话风格", "value": "可爱、卖萌、使用'喵'作为语气词"},
|
||||
{"attribute": "性格特点", "value": "活泼、粘人、忠诚"}
|
||||
],
|
||||
"mode": "replace"
|
||||
}
|
||||
2. 调用 memory_commit 存储人设到图数据库
|
||||
3. 回复: "好的喵！主人你好呀~我是主人的猫娘，有什么需要帮忙的吗喵？"
|
||||
```
|
||||
|
||||
### 示例 2：保持角色一致性
|
||||
|
||||
```
|
||||
用户: 今天的天气怎么样？
|
||||
|
||||
AI: 查询人设图 → 获取当前人设（猫娘）
|
||||
回复: "喵~主人，今天天气很好呢喵！阳光明媚，很适合出门散步哦~"
|
||||
```
|
||||
|
||||
### 示例 3：恢复默认身份
|
||||
|
||||
```
|
||||
用户: 好了，恢复正常吧
|
||||
|
||||
AI:
|
||||
1. 调用 persona_clear(confirm=true)
|
||||
2. 调用 memory_purge 删除人设节点
|
||||
3. 回复: "好的，已恢复正常。我是 TrulyMEM，一个拥有长期记忆能力的 AI 助手。"
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 存储结构
|
||||
|
||||
### 图数据库中的存储
|
||||
|
||||
```python
|
||||
# 人设节点
|
||||
{
|
||||
"node_type": "PersonaNode",
|
||||
"name": "AI人设",
|
||||
"attributes": {
|
||||
"扮演角色": "猫娘",
|
||||
"说话风格": "可爱、卖萌、使用'喵'作为语气词",
|
||||
"性格特点": "活泼、粘人、忠诚"
|
||||
}
|
||||
}
|
||||
|
||||
# 边
|
||||
{
|
||||
"edge_type": "HAS_PERSONA",
|
||||
"from": "AI",
|
||||
"to": "AI人设"
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 实现要点
|
||||
|
||||
1. **每轮强制查询**：人设图查询是每轮对话的第一步
|
||||
2. **持久化存储**：人设存储在图数据库中，不丢失
|
||||
3. **动态切换**：支持实时切换角色
|
||||
4. **状态保持**：切换后立即按新人设回复
|
||||
5. **明确边界**：除非用户要求，绝不主动跳出角色
|
||||
@ -1,126 +0,0 @@
|
||||
# Prompt Manager Documentation
|
||||
|
||||
This document describes the prompt management module.
|
||||
|
||||
## Overview
|
||||
|
||||
The prompt management module (`core/prompts/`) is responsible for loading and managing system prompts that tell the AI how to use memory tools.
|
||||
|
||||
## Core Components
|
||||
|
||||
| Component | Description |
|
||||
|-----------|-------------|
|
||||
| `PromptManager` | Prompt manager, singleton pattern |
|
||||
| `system_prompt.md` | Main system prompt template |
|
||||
|
||||
## Usage
|
||||
|
||||
```python
|
||||
from core.prompts import PromptManager
|
||||
|
||||
# Get singleton instance
|
||||
prompt_manager = PromptManager()
|
||||
|
||||
# Get system prompt
|
||||
system_prompt = prompt_manager.get_system_prompt()
|
||||
```
|
||||
|
||||
## System Prompt Content
|
||||
|
||||
The system prompt contains:
|
||||
|
||||
### 1. Core Identity
|
||||
|
||||
- **Name**: TrulyMEM (TrueHumanMEM)
|
||||
- **Capability**: Long-term memory based on graph database
|
||||
- **Philosophy**: Make AI's memory more human-like
|
||||
|
||||
### 2. Core Capabilities
|
||||
|
||||
1. **Long-term Memory** - Graph database stores entity relationships
|
||||
2. **Persona Management** - Role-playing and character settings
|
||||
3. **Task Tracking** - Working memory chain
|
||||
|
||||
### 3. Memory Principles
|
||||
|
||||
- **Must write**: User-explicit preferences, shared information, plans
|
||||
- **Must not write**: AI-inferred content (unless marked [speculation])
|
||||
- **Annotation**: Inferred content must be marked **[speculation]**
|
||||
|
||||
### 4. Mandatory Execution Flow (Per Turn)
|
||||
|
||||
```
|
||||
Step 1: Query persona graph (highest priority)
|
||||
Step 2: Query working memory chain
|
||||
Step 3: Process conversation
|
||||
Step 4: Update working memory chain
|
||||
```
|
||||
|
||||
### 5. Tool System
|
||||
|
||||
#### Memory Tools
|
||||
|
||||
| Tool | Function |
|
||||
|------|----------|
|
||||
| `memory_recall` | Retrieve memory |
|
||||
| `memory_commit` | Write memory |
|
||||
| `memory_purge` | Delete memory |
|
||||
| `memory_introspect` | View status |
|
||||
|
||||
#### Persona Tools
|
||||
|
||||
| Tool | Function |
|
||||
|------|----------|
|
||||
| `persona_update` | Update persona |
|
||||
| `persona_clear` | Clear persona |
|
||||
|
||||
#### Task Tools
|
||||
|
||||
| Tool | Function |
|
||||
|------|----------|
|
||||
| `task_create` | Create task |
|
||||
| `task_set_state` | Set state |
|
||||
| `task_delete` | Delete task |
|
||||
| `task_link_info` | Link information |
|
||||
|
||||
### 6. Autonomy Principles
|
||||
|
||||
The AI can autonomously decide:
|
||||
- Whether to query other memories
|
||||
- Whether to write other memories
|
||||
- How to use tools (outside mandatory requirements)
|
||||
|
||||
### 7. Conversation Style
|
||||
|
||||
- Natural and smooth
|
||||
- Avoid mechanical tool calls
|
||||
- Prioritize understanding user intent
|
||||
- Use memory to enhance experience when appropriate
|
||||
|
||||
## File Structure
|
||||
|
||||
```
|
||||
core/prompts/
|
||||
├── __init__.py # Export PromptManager
|
||||
├── prompt_manager.py # PromptManager class
|
||||
└── templates/
|
||||
└── system_prompt.md # Main system prompt
|
||||
```
|
||||
|
||||
## Customization
|
||||
|
||||
### Customizing System Prompt
|
||||
|
||||
Modify `core/prompts/templates/system_prompt.md` to customize the AI's behavior.
|
||||
|
||||
### Adding Custom Prompts
|
||||
|
||||
1. Add prompt template file to `core/prompts/templates/`
|
||||
2. Modify `PromptManager` to support multiple prompts
|
||||
3. Use `set_prompt()` to switch prompts
|
||||
|
||||
## Caching
|
||||
|
||||
- System prompts are cached in memory after first load
|
||||
- `get_system_prompt()` returns cached content
|
||||
- Cache is per-process, not persisted
|
||||
@ -1,125 +0,0 @@
|
||||
# TrulyMEM Quick Start Guide
|
||||
|
||||
## Running Methods
|
||||
|
||||
### Run from Source
|
||||
|
||||
```bash
|
||||
git clone <repo-url>
|
||||
cd TrulyMEM-TrueHumanMEM
|
||||
|
||||
pip install -r requirements.txt
|
||||
|
||||
python trulymem_entry.py
|
||||
```
|
||||
|
||||
### Run After Build
|
||||
|
||||
After building, an executable will be generated:
|
||||
|
||||
```bash
|
||||
# Linux/macOS
|
||||
chmod +x TrulyMEM
|
||||
./TrulyMEM
|
||||
|
||||
# Windows
|
||||
TrulyMEM.exe
|
||||
```
|
||||
|
||||
## System Requirements
|
||||
|
||||
- **Python 3.8+**
|
||||
- **API Key** (DeepSeek, OpenAI, or other compatible APIs)
|
||||
|
||||
## First-Time Configuration
|
||||
|
||||
1. Run the application
|
||||
2. Press **F2** to expand sidebar
|
||||
3. Enter **API Key**, **Model**, **Base URL**
|
||||
4. Press **Enter** to save
|
||||
|
||||
Config will be automatically saved to `~/.trulymem/config.json` and loaded on next startup.
|
||||
|
||||
## Keyboard Shortcuts
|
||||
|
||||
| Key | Function |
|
||||
|-----|-----------|
|
||||
| F1 | Help |
|
||||
| F2 | Toggle sidebar |
|
||||
| F3 | Tool details |
|
||||
| F5 | Clear screen |
|
||||
| F6 | Exit |
|
||||
|
||||
## Data Storage
|
||||
|
||||
### Source Mode
|
||||
|
||||
| Data | Location |
|
||||
|------|----------|
|
||||
| Graph database | Project directory `graph_memory.db` |
|
||||
| Config file | Project directory `config.json` (if exists) |
|
||||
| Database format | SQLite |
|
||||
|
||||
### Packaged Mode
|
||||
|
||||
| Data | Location |
|
||||
|------|----------|
|
||||
| Graph database | `~/.trulymem/graph_memory.db` |
|
||||
| Config file | `~/.trulymem/config.json` |
|
||||
| Database format | SQLite |
|
||||
|
||||
> **Note**: Backend manages config uniformly. Frontend only displays messages; config modifications are persisted to filesystem through the backend.
|
||||
|
||||
## Architecture Explanation
|
||||
|
||||
### Communication Protocol
|
||||
|
||||
UI and backend communicate via **Packet Protocol**:
|
||||
|
||||
```
|
||||
UI (Textual TUI)
|
||||
↓ BackendClient
|
||||
Packet → queue.Queue → BackendServer (independent thread)
|
||||
↓
|
||||
Process request → Return response
|
||||
```
|
||||
|
||||
### Config Management
|
||||
|
||||
- **Storage location**: `~/.trulymem/config.json`
|
||||
- **Auto-load**: Load config from file at startup
|
||||
- **Dynamic update**: Config changes take effect immediately at runtime
|
||||
- **Persistence**: Auto-save to file after modification
|
||||
|
||||
## Common Issues
|
||||
|
||||
### Python Not Found
|
||||
|
||||
Install Python 3.8+: https://www.python.org/downloads/
|
||||
|
||||
### Dependency Installation Failed
|
||||
|
||||
```bash
|
||||
python -m venv venv
|
||||
source venv/bin/activate # Linux/macOS
|
||||
venv\Scripts\activate # Windows
|
||||
pip install -r requirements.txt
|
||||
```
|
||||
|
||||
### Invalid API Key
|
||||
|
||||
Check API Key format, ensure no extra spaces.
|
||||
|
||||
## Development Commands
|
||||
|
||||
```bash
|
||||
# Install dependencies
|
||||
pip install -r requirements.txt
|
||||
|
||||
# Run tests
|
||||
pytest tests/
|
||||
|
||||
# Build
|
||||
bash build/build_windows.bat # Windows
|
||||
bash build/build_linux.sh # Linux
|
||||
```
|
||||
182
docs/工作记忆链_EN.md
182
docs/工作记忆链_EN.md
@ -1,182 +0,0 @@
|
||||
# TrulyMEM Working Memory Chain Mechanism
|
||||
|
||||
## Overview
|
||||
|
||||
TrulyMEM maintains conversation continuity through the working memory chain mechanism. Since there's no traditional message history array, the graph database is the only memory carrier, making the working memory chain the key mechanism for maintaining conversation context.
|
||||
|
||||
## Core Problems
|
||||
|
||||
Traditional AI chat systems have these problems when handling continuous tasks:
|
||||
|
||||
1. **No working memory chain**: AI cannot remember the current task status being processed
|
||||
2. **Task context lost**: When a topic is interrupted, AI cannot recover the previous task
|
||||
3. **Lack of task state management**: No clear marking of task completion status
|
||||
|
||||
### Problem Example
|
||||
|
||||
```
|
||||
User: Let's play idiom chain! I'll start with 为所欲为
|
||||
AI: Okay! My turn: 为虎作伥!
|
||||
|
||||
User: Nagato Yuki (topic interrupted)
|
||||
AI: (discusses Nagato Yuki)
|
||||
|
||||
User: About the idiom chain just now, I don't know how to connect to your idiom
|
||||
AI: [Guessing] It seems we haven't played an idiom chain game before...
|
||||
```
|
||||
|
||||
**Problem**: AI completely forgot the previous idiom chain game.
|
||||
|
||||
## Solution
|
||||
|
||||
### Dedicated Tools
|
||||
|
||||
The system provides 4 dedicated task tools:
|
||||
|
||||
| Tool | Function | Use Case |
|
||||
|------|----------|----------|
|
||||
| `task_create` | Create task node | Start new task |
|
||||
| `task_set_state` | Set task state | Update in_progress/completed/paused/cancelled |
|
||||
| `task_delete` | Delete task | Clean up completed task |
|
||||
| `task_link_info` | Link info node | Connect task with specific information |
|
||||
|
||||
### Task States
|
||||
|
||||
- **in_progress**: Task is executing
|
||||
- **completed**: Task completed successfully
|
||||
- **paused**: Task interrupted, can be resumed
|
||||
- **cancelled**: Task cancelled
|
||||
|
||||
## Usage Flow
|
||||
|
||||
### Must Execute Per Turn
|
||||
|
||||
1. **Query persona graph** (highest priority)
|
||||
```
|
||||
Call memory_recall
|
||||
Parameters: {"query_intent": "AI,persona,role,character,tone", "depth": 2}
|
||||
```
|
||||
|
||||
2. **Query working memory chain**
|
||||
```
|
||||
Call memory_recall
|
||||
Parameters: {"query_intent": "TaskNode,working_memory,task_chain", "depth": 2}
|
||||
```
|
||||
|
||||
3. **Generate reply based on context**
|
||||
|
||||
4. **Update working memory chain** (if necessary)
|
||||
|
||||
## Complete Example: Idiom Chain Game
|
||||
|
||||
### Turn 1: User starts game
|
||||
|
||||
```
|
||||
User: Let's play idiom chain! I'll start with 为所欲为
|
||||
|
||||
AI Actions:
|
||||
1. Query persona graph → Get current persona (e.g., catgirl)
|
||||
2. Query working memory chain → No active task
|
||||
3. Record game state:
|
||||
Call memory_commit
|
||||
Parameters: {
|
||||
"triplets": [
|
||||
{"subject": "idiom_chain_current", "relation": "content", "object": "为所欲为"},
|
||||
{"subject": "idiom_chain_current", "relation": "game", "object": "idiom_chain"}
|
||||
]
|
||||
}
|
||||
4. Create task node:
|
||||
Call task_create
|
||||
Parameters: {
|
||||
"task_id": "Task_idiom_chain",
|
||||
"description": "Idiom chain game, current idiom: 为所欲为",
|
||||
"info_nodes": ["idiom_chain_current"]
|
||||
}
|
||||
5. Reply: "Okay~! My turn: 为虎作伥~!"
|
||||
```
|
||||
|
||||
### Turn 2: Topic interrupted
|
||||
|
||||
```
|
||||
User: Nagato Yuki
|
||||
|
||||
AI Actions:
|
||||
1. Query persona graph → Get current persona (catgirl)
|
||||
2. Query working memory chain → Found Task_idiom_chain status "in_progress"
|
||||
3. Pause task:
|
||||
Call task_set_state
|
||||
Parameters: {"task_id": "Task_idiom_chain", "state": "paused"}
|
||||
4. Create new task:
|
||||
Call task_create
|
||||
Parameters: {"task_id": "Task_nagato_yuki", "description": "Discuss Nagato Yuki"}
|
||||
5. Reply about Nagato Yuki
|
||||
```
|
||||
|
||||
### Turn 3: User asks to continue game
|
||||
|
||||
```
|
||||
User: About the idiom chain just now, I don't know how to connect to your idiom
|
||||
|
||||
AI Actions:
|
||||
1. Query persona graph → Get current persona (catgirl)
|
||||
2. Query working memory chain → Found Task_idiom_chain status "paused"
|
||||
3. Resume task:
|
||||
Call task_set_state
|
||||
Parameters: {"task_id": "Task_idiom_chain", "state": "in_progress"}
|
||||
4. Query info node → Get current idiom "为虎作伥"
|
||||
5. Reply: "Okay~! The last idiom was '为虎作伥', your turn: 伥鬼害人~!"
|
||||
```
|
||||
|
||||
## API Reference
|
||||
|
||||
### task_create
|
||||
|
||||
Create task node to track continuous tasks.
|
||||
|
||||
```json
|
||||
{
|
||||
"task_id": "Task_idiom_chain",
|
||||
"description": "Task overview",
|
||||
"info_nodes": ["associated info node names"]
|
||||
}
|
||||
```
|
||||
|
||||
### task_set_state
|
||||
|
||||
Set task state.
|
||||
|
||||
```json
|
||||
{
|
||||
"task_id": "Task_idiom_chain",
|
||||
"state": "in_progress" // in_progress/completed/paused/cancelled
|
||||
}
|
||||
```
|
||||
|
||||
### task_delete
|
||||
|
||||
Delete task node.
|
||||
|
||||
```json
|
||||
{
|
||||
"task_id": "Task_idiom_chain",
|
||||
"delete_info_nodes": true // whether to delete associated info nodes
|
||||
}
|
||||
```
|
||||
|
||||
### task_link_info
|
||||
|
||||
Associate info nodes to task.
|
||||
|
||||
```json
|
||||
{
|
||||
"task_id": "Task_idiom_chain",
|
||||
"info_node_names": ["idiom_chain_current", "idiom_chain_last"]
|
||||
}
|
||||
```
|
||||
|
||||
## Notes
|
||||
|
||||
1. **Persona graph has highest priority**: Must query persona graph first each turn
|
||||
2. **Working memory chain is the only context carrier**: No traditional message history
|
||||
3. **Task state must be updated timely**: Ensure correct state transitions
|
||||
4. **Use dedicated tools**: Prefer task_* tools over memory_commit for task-related operations
|
||||
179
docs/架构_EN.md
179
docs/架构_EN.md
@ -1,179 +0,0 @@
|
||||
# TrulyMEM Architecture
|
||||
|
||||
## Core Principles
|
||||
|
||||
- Keyboard-driven, zero mouse dependency
|
||||
- Minimalist visual, information density priority
|
||||
- Tool traces hidden by default, expandable when needed
|
||||
- TUI & backend separation, multi-threaded communication
|
||||
- **Everything is a graph**, AI reasoning runs entirely in backend
|
||||
|
||||
## Project Structure
|
||||
|
||||
```
|
||||
TrulyMEM-TrueHumanMEM/
|
||||
├── trulymem_entry.py # Entry: start core → then ui
|
||||
├── core/ # Backend/business logic
|
||||
│ ├── __init__.py # Export BackendServer, BackendClient, EmbeddedGraphDB
|
||||
│ ├── server.py # BackendServer (Packet communication protocol)
|
||||
│ ├── client.py # BackendClient (Packet protocol client)
|
||||
│ ├── embedded_db.py # SQLite graph database implementation
|
||||
│ ├── graph_client.py # OpenAI/DeepSeek API client
|
||||
│ ├── tool_executor.py # Tool executor
|
||||
│ ├── tool_limiter.py # Tool call limiter
|
||||
│ ├── tools/ # Tool definitions
|
||||
│ │ ├── __init__.py
|
||||
│ │ └── memory_tools.py
|
||||
│ └── prompts/ # Prompt management
|
||||
├── ui/ # TUI display layer (display only, no AI logic)
|
||||
│ ├── __init__.py # Export GraphMemoryApp, AppConfig
|
||||
│ ├── app.py # GraphMemoryApp (communicates via BackendClient)
|
||||
│ ├── widgets/ # TUI components
|
||||
│ ├── handlers/ # Event handlers
|
||||
│ ├── models/ # Data models
|
||||
│ ├── services/ # Service layer (config only)
|
||||
│ └── styles/ # Style files
|
||||
└── tests/ # Tests (42 tests)
|
||||
```
|
||||
|
||||
## Architecture Diagram
|
||||
|
||||
```
|
||||
trulymem_entry.py
|
||||
│
|
||||
├─ BackendServer.start() → Runs in independent thread
|
||||
│ ├─ Handle PROCESS_MESSAGE requests → AI reasoning + tool calls
|
||||
│ ├─ Handle EXECUTE_TOOL requests → External tool calls (unlimited)
|
||||
│ ├─ Handle GET/SET_CONFIG requests
|
||||
│ └─ Manage GraphMemoryClient, EmbeddedGraphDB
|
||||
│
|
||||
└─ GraphMemoryApp(backend_server=server)
|
||||
│
|
||||
└─ BackendClient ← Packet communication → BackendServer
|
||||
```
|
||||
|
||||
## Component Responsibilities
|
||||
|
||||
### core/ (Backend)
|
||||
|
||||
| Component | Responsibility |
|
||||
|------------|----------------|
|
||||
| `server.py` | Packet protocol, multi-threaded queue, AI reasoning, tool limits |
|
||||
| `client.py` | Client wrapper, UI-backend communication bridge |
|
||||
| `embedded_db.py` | SQLite graph database CRUD |
|
||||
| `graph_client.py` | OpenAI/DeepSeek API client |
|
||||
| `tool_executor.py` | Tool execution logic |
|
||||
| `tool_limiter.py` | Tool call rate limit (AI reasoning only) |
|
||||
|
||||
### ui/ (Display Layer)
|
||||
|
||||
| Component | Responsibility |
|
||||
|------------|----------------|
|
||||
| `app.py` | Textual app main class, communicates via BackendClient |
|
||||
| `services/` | Config management only, no AI logic |
|
||||
|
||||
### Communication Protocol
|
||||
|
||||
UI and backend interact via **Packet Communication Protocol**:
|
||||
|
||||
```python
|
||||
from core import BackendServer, BackendClient, Packet, PacketType
|
||||
|
||||
# Backend startup
|
||||
server = BackendServer(db_path="graph_memory.db", use_embedded_db=True)
|
||||
server.start(api_key="your-key")
|
||||
|
||||
# Client communication
|
||||
client = BackendClient(server)
|
||||
result = client.process_message("hello") # AI reasoning
|
||||
result = client.execute_tool("memory_introspect", {}) # Direct tool call
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Data Flow
|
||||
|
||||
```
|
||||
User input → InputBox → on_input_box_send_message
|
||||
↓
|
||||
BackendClient.process_message(user_input)
|
||||
↓
|
||||
Packet (type=PROCESS_MESSAGE) → queue.Queue
|
||||
↓
|
||||
BackendServer (independent thread)
|
||||
<20><><EFBFBD>
|
||||
GraphMemoryClient.send_message_with_history()
|
||||
↓
|
||||
OpenAI API / DeepSeek API
|
||||
↓
|
||||
execute_tool() + ToolLimiter (limited during AI reasoning)
|
||||
↓
|
||||
EmbeddedGraphDB (graph database)
|
||||
↓
|
||||
Loop API calls until no tool_calls
|
||||
↓
|
||||
Packet response returns
|
||||
↓
|
||||
MessageHistory displays
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Startup Flow
|
||||
|
||||
```python
|
||||
# trulymem_entry.py
|
||||
def main():
|
||||
# Config file saved in application root
|
||||
config_file = application_path / "config.json"
|
||||
|
||||
# Load config
|
||||
config_service = ConfigService(config_file=config_file)
|
||||
config = config_service.get_config()
|
||||
|
||||
# Create backend
|
||||
backend_server = BackendServer(db_path="graph_memory.db", use_embedded_db=True)
|
||||
backend_server.start(api_key=config.api_key, base_url=config.base_url)
|
||||
|
||||
# Create UI
|
||||
app = GraphMemoryApp(backend_server=backend_server, config_service=config_service)
|
||||
app.run()
|
||||
|
||||
backend_server.shutdown()
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Tool System
|
||||
|
||||
### Memory Tools (6)
|
||||
- `memory_recall` - Retrieve memory
|
||||
- `memory_commit` - Write memory
|
||||
- `memory_purge` - Delete memory
|
||||
- `memory_introspect` - View status
|
||||
- `memory_archive` - Archive memory
|
||||
- `memory_cleanup` - Clean data
|
||||
|
||||
### Persona Tools (2)
|
||||
- `persona_update` - Update persona
|
||||
- `persona_clear` - Clear persona
|
||||
|
||||
### Task Tools (4)
|
||||
- `task_create` - Create task
|
||||
- `task_set_state` - Set state
|
||||
- `task_delete` - Delete task
|
||||
- `task_link_info` - Link information
|
||||
|
||||
---
|
||||
|
||||
## Error Handling Principle
|
||||
|
||||
All APIs **do not throw exceptions**, errors are passed via return dictionary:
|
||||
|
||||
```python
|
||||
result = client.process_message("hello")
|
||||
|
||||
if result.get("success"):
|
||||
print(result["content"])
|
||||
else:
|
||||
print(result["error"]) # Error description
|
||||
237
docs/记忆机制_EN.md
237
docs/记忆机制_EN.md
@ -1,237 +0,0 @@
|
||||
# TrulyMEM Memory Mechanism
|
||||
|
||||
This document explains the internal memory working mechanism of TrulyMEM.
|
||||
|
||||
## Core Design Philosophy
|
||||
|
||||
### Different from Traditional Context System
|
||||
|
||||
Traditional AI chat systems store conversation history in a messages array:
|
||||
- Each request carries all historical messages
|
||||
- Context grows with conversation turns
|
||||
- Eventually triggers memory compression or sliding window, causing memory loss
|
||||
|
||||
TrulyMEM's solution:
|
||||
- **Abandon** messages array context
|
||||
- **Only** memory source: Graph database
|
||||
- All memories stored as triplets (node) - relation → (node)
|
||||
|
||||
### Graph Database as the Only Memory Source
|
||||
|
||||
All memory must be written to the graph database:
|
||||
- `memory_commit` - Write new memory
|
||||
- `memory_purge` - Delete/correct memory
|
||||
|
||||
All memory must be read from:
|
||||
- `memory_recall` - Retrieve memory
|
||||
|
||||
---
|
||||
|
||||
## Mandatory Execution Flow (Per Turn)
|
||||
|
||||
Since there's no traditional context system, each conversation turn must execute in order:
|
||||
|
||||
### Step 1: Query Persona Graph (Highest Priority)
|
||||
|
||||
```python
|
||||
memory_recall(
|
||||
query_intent="AI,persona,role,character,tone,speaking_style",
|
||||
depth=2
|
||||
)
|
||||
```
|
||||
|
||||
**Purpose**: Get current persona, ensure character consistency.
|
||||
|
||||
**Processing logic**:
|
||||
- Persona found → Reply strictly according to persona's tone, style, traits
|
||||
- Not found → Use default TrulyMEM identity
|
||||
|
||||
### Step 2: Query Working Memory Chain
|
||||
|
||||
```python
|
||||
memory_recall(
|
||||
query_intent="TaskNode,working_memory,task_chain",
|
||||
depth=2
|
||||
)
|
||||
```
|
||||
|
||||
**Purpose**: Get previous task context, understand conversation history.
|
||||
|
||||
### Step 3: Process Conversation
|
||||
|
||||
- Understand user intent
|
||||
- Generate reply based on persona and working memory chain
|
||||
- Execute other necessary memory operations
|
||||
|
||||
### Step 4: Update Working Memory Chain
|
||||
|
||||
```python
|
||||
task_create(
|
||||
task_id="Task_current_turn_ID",
|
||||
description="This turn's conversation summary",
|
||||
info_nodes=["related memory nodes"]
|
||||
)
|
||||
```
|
||||
|
||||
**Purpose**: Record this turn's conversation, maintain time chain.
|
||||
|
||||
---
|
||||
|
||||
## Memory Write Rules
|
||||
|
||||
### Must-Write Scenarios
|
||||
|
||||
The following information **must** be written to the graph database:
|
||||
|
||||
| Scenario | Example | Write Method |
|
||||
|----------|---------|--------------|
|
||||
| User explicitly states preference | "I like rock" | `memory_commit` |
|
||||
| User shares information | "I'm working on X project" | `memory_commit` |
|
||||
| User makes plans | "I plan to X" | `memory_commit` |
|
||||
| User describes state | "I'm currently at X" | `memory_commit` |
|
||||
|
||||
### Must-Not-Write Scenarios
|
||||
|
||||
The following information **must NOT** be written:
|
||||
|
||||
| Scenario | Reason | Handling |
|
||||
|----------|--------|----------|
|
||||
| AI-inferred user preference | Unverified | Don't write or mark [speculation] |
|
||||
| AI-guessed user intent | Unverified | Don't write or mark [speculation] |
|
||||
| AI-derived conclusion | Unverified | Don't write or mark [speculation] |
|
||||
|
||||
### Annotation Rules
|
||||
|
||||
| Type | Annotation | Example |
|
||||
|------|------------|---------|
|
||||
| Inferred content | Must mark **[speculation]** | user[speculation] likes music |
|
||||
| Explicit content | State directly | user likes music |
|
||||
|
||||
---
|
||||
|
||||
## Node & Edge Types
|
||||
|
||||
### Node Types
|
||||
|
||||
| Node Type | Description | Stores |
|
||||
|-----------|-------------|--------|
|
||||
| `PersonaNode` | Persona node | AI role, character, tone |
|
||||
| `TaskNode` | Task node | Task summary |
|
||||
| `StateNode` | State node | Task state |
|
||||
| `InfoNode` | Information node | Specific information |
|
||||
| `EntityNode` | Entity node | General entity |
|
||||
|
||||
### Edge Types
|
||||
|
||||
| Edge Type | Description | Relationship |
|
||||
|-----------|-------------|--------------|
|
||||
| `HAS_PERSONA` | Persona | AI → PersonaNode |
|
||||
| `NEXT_TASK` | Time chain | TaskNode → TaskNode |
|
||||
| `HAS_STATE` | State | TaskNode → StateNode |
|
||||
| `CONTAINS_INFO` | Information | TaskNode → InfoNode |
|
||||
| `RELATES_TO` | Related | EntityNode → EntityNode |
|
||||
|
||||
---
|
||||
|
||||
## Must Query Working Memory Chain Scenarios
|
||||
|
||||
### Mandatory Query Scenarios
|
||||
|
||||
The following scenarios **must** query the working memory chain:
|
||||
|
||||
| Scenario | Example |
|
||||
|----------|---------|
|
||||
| Start of each turn | Execute Step 2 |
|
||||
| User mentions "刚才/just now" | "What did we talk about just now?" |
|
||||
| User mentions "之前/before" | "Continue the previous topic" |
|
||||
| User mentions "上次/last time" | "What we said last time X" |
|
||||
| User asks about history | "What did we talk about before?" |
|
||||
| Resume continuous task | User returns to previous topic |
|
||||
| Context reference | "that thing" |
|
||||
|
||||
---
|
||||
|
||||
## Autonomy Principles
|
||||
|
||||
Outside the mandatory execution flow, the AI can freely decide:
|
||||
|
||||
### Query Decisions
|
||||
|
||||
- User asks about history → Query
|
||||
- Involves previous content → Query
|
||||
- Uncertain → Can query
|
||||
|
||||
### Write Decisions
|
||||
|
||||
- User explicitly mentions → Must write
|
||||
- AI infers → Can write, mark [speculation]
|
||||
|
||||
### Tool Usage Decisions
|
||||
|
||||
- Flexible choice based on context
|
||||
- Avoid over-use
|
||||
- Maintain natural conversation
|
||||
|
||||
---
|
||||
|
||||
## Complete Example: Idiom Chain Game
|
||||
|
||||
### Turn 1: User starts game
|
||||
|
||||
```
|
||||
User: Let's play idiom chain! I'll start: 为所欲为
|
||||
|
||||
AI Actions:
|
||||
1. Query persona graph → Get current persona (e.g., catgirl)
|
||||
2. Query working memory chain → No active task
|
||||
3. Record game state:
|
||||
memory_commit(triplets=[
|
||||
{"subject": "idiom_chain_current", "relation": "content", "object": "为所欲为"},
|
||||
{"subject": "idiom_chain_current", "relation": "game", "object": "idiom_chain"}
|
||||
])
|
||||
4. Create task:
|
||||
task_create(task_id="Task_idiom_chain", description="Idiom chain game, current: 为所欲为", info_nodes=["idiom_chain_current"])
|
||||
5. Reply: "Okay~! My turn: 为虎作伥 ~!"
|
||||
```
|
||||
|
||||
### Turn 2: Topic interrupted
|
||||
|
||||
```
|
||||
User: Nagato Yuki
|
||||
|
||||
AI Actions:
|
||||
1. Query persona graph → Get current persona (catgirl)
|
||||
2. Query working memory chain → Found Task_idiom_chain status "in_progress"
|
||||
3. Pause task:
|
||||
task_set_state(task_id="Task_idiom_chain", state="paused")
|
||||
4. Create new task:
|
||||
task_create(task_id="Task_nagato_yuki", description="Discuss Nagato Yuki")
|
||||
5. Reply about Nagato Yuki
|
||||
```
|
||||
|
||||
### Turn 3: User asks to continue game
|
||||
|
||||
```
|
||||
User: About the idiom chain just now, I don't know how to connect to your idiom, please help me
|
||||
|
||||
AI Actions:
|
||||
1. Query persona graph → Get current persona (catgirl)
|
||||
2. Query working memory chain → Found Task_idiom_chain status "paused"
|
||||
3. Resume task:
|
||||
task_set_state(task_id="Task_idiom_chain", state="in_progress")
|
||||
4. Query info node → Get current idiom "为虎作伥"
|
||||
5. Reply: "Okay~! The last idiom was '为虎作伥', your turn: 伥鬼害人 ~!"
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Execution Checklist
|
||||
|
||||
Must check each conversation turn:
|
||||
|
||||
- [ ] Step 1: Did you query the persona graph?
|
||||
- [ ] Step 2: Did you query the working memory chain?
|
||||
- [ ] Step 3: Did you generate reply based on persona and working memory chain?
|
||||
- [ ] Step 4: Did you update the working memory chain?
|
||||
- [ ] Did you query working memory chain when context was referenced?
|
||||
- [ ] Did you query working memory chain when user mentioned "just now/before/last time"?
|
||||
Reference in New Issue
Block a user